Background and significance

Brain-Computer Interface (BCI) technology is of great significance in the field of neural engineering and brain-inspired intelligence by connecting the brain with external devices to realize the reading of brain information and the interaction of external instructions. However, the existing non-invasive BCI is limited by signal resolution and reliability, and it is difficult to decode the fine-grained information of the deep structure of the brain, resulting in a high error rate. At the same time, traditional BCI mostly focuses on pattern recognition at the signal level, lacks understanding of the semantic level of brain activity, and is difficult to effectively intervene in complex cognitive states (such as emotion and purpose). This is particularly prominent in the intervention of neuropsychiatric diseases: anxiety disorders, [post-traumatic stress disorder]{.underline} (PTSD), etc., are characterized by abnormal emotional circuits in the brain, and existing drugs and superficial brain stimulation therapy often have limited effect, and a large number of patients do not respond to first-line treatment. There is an urgent need for a more accurate, controllable and intelligent brain-computer interaction method to decode and regulate specific states of the brain and provide innovative therapies for related diseases.

The rise of high spatiotemporal resolution ultrasound brain-computer interface technology provides new ideas for the above problems. Ultrasound is non-invasive, deep-focused, and can directly regulate the activity of deep subcortical nuclei such as [the amygdala and]{.underline} other emotion-related areas through low-intensity focused ultrasound (LIFU). Compared with transcranial magnetic stimulation and other methods that mainly affect the cortex, ultrasound can precisely act on deep neural circuits, avoid roundabout adjustment, and improve the potential for efficacy. At the same time, functional [ultrasound imaging]{.underline} (fUS) can detect local blood flow changes in the brain with extremely high sensitivity and spatial resolution, which is highly correlated with neuronal discharge, providing a new means to obtain brain activity between EEG and fMRI. Studies have shown that fUS can achieve a spatial resolution of about 100 microns and a temporal resolution of less than 1 second, and can cover centimeter-level brain regions in a single image. This large-scale, high-precision brain imaging capability enables non-invasive BCI to combine the accuracy of previous invasive electrodes with the breadth of whole-brain coverage. Recently, some studies have tried to introduce tFUS neuromodulation into the BCI circuit, in order to reduce the error rate of BCI and enhance the interaction effect. For example, targeting transcranial ultrasound to the V5 region of the visual cortex can significantly reduce the error rate of BCI spelling tasks while enhancing the rhythmic activity of related brain regions. These advances indicate the dual role of ultrasound in BCI: it can be used as a high-precision information acquisition channel and an execution channel to intervene in brain state, laying the foundation for the construction of a closed-loop brain-computer interface system.

Although ultrasound neuromodulation has shown preliminary results in improving BCI performance and treating psychiatric disorders, most of the existing studies use open-loop or pre-set stimulation strategies, and lack intelligent closed-loop regulation. For example, ultrasound intervention trials in patients with anxiety/PTSD have shown a significant reduction in anxiety and traumatic stress scores with no serious adverse effects after several weeks of daily ultrasound stimulation; Another study using ultrasound stimulation of the auricular vagus nerve similarly observed significant relief in anxiety and depression symptoms. However, most of these treatments are based on fixed-parameter stimulation and have not been adaptively adjusted to changes in individual brain status. In order to further improve the efficacy and reduce side effects, it is urgent to introduce closed-loop control: that is, to decode the patient's brain state in real time, dynamically adjust the stimulation strategy according to the treatment goal, and realize the automatic cycle of "brain reading-recognition-decision-treatment". This places higher demands on system intelligence: traditional algorithms struggle to understand high-level brain states or purposes, and are incapable of making such complex decisions. Therefore, we need to introduce a new type of artificial intelligence theory, the artificial consciousness model, so that the machine can have human-like semantic cognition and autonomous decision-making ability, so as to truly realize the closed-loop brain-computer interface centered on the patient's purpose and semantic state.

This project is guided by the "Data-Information-Knowledge-Wisdom-Purpose" artificial consciousness theory proposed by Professor Yucong Duan, focusing on **[high-spatiotemporal resolution ultrasound brain-computer interface]{.underline}**Application in the intervention of neuropsychiatric diseases. The DIKWP model provides a new cognitive architecture with a Purpose hierarchy, emphasizing the whole chain of semantic processing from raw data to high-level Wisdom decision-making to Purpose-oriented. Combined with the precise perception and regulation of the brain by ultrasound, we are expected to build a self-explanatory, purpose-driven brain-computer interaction system, so that the machine can "read" the semantic state of the brain and impose purposeful interventions, so as to more effectively alleviate the symptoms of anxiety, PTSD and other diseases. This will not only innovate the theoretical and technical connotation of brain-computer interface, but also open up a new path for the field of neuromodulation, which has important academic value and clinical significance.

In summary, aiming at the current bottleneck of BCI and brain disease intervention, this project proposes to deeply integrate the DIKWP artificial consciousness model with high-resolution ultrasound brain-computer interface, aiming to achieve: brain state decoding of high-level semantic information, purpose-driven precise ultrasound intervention, adaptive closed-loop adjustment and self-explanatory decision-making. The expected results will make a breakthrough in the field of brain-computer interaction and artificial consciousness at the semantic layer in China, and lay a new foundation for intelligent medical care and brain-like AI in the future.

Theoretical basis: DIKWP network artificial consciousness model

DIKWP Model Concept: DIKWP is an artificial consciousness framework composed of five hierarchical elements: Data, Information, Knowledge, Wisdom and Purpose. The model is extended by Professor Yucong Duan's team on the basis of the classic DIKW (pyramid) model, by adding a "Purpose" layer on top of knowledge-Wisdom, and breaking the traditional bottom-up one-way hierarchical structure, and adopting a network interaction structure. This means that the five elements in the DIKWP model can not only be transferred to each other, but can interact in both directions, so that there are 25 potential interaction paths in the system, 5×5. This network dynamic architecture is closer to the information flow pattern in the real brain: the cognitive processes in reality are not simple straight lines, but are full of feedback loops and parallel interactions, high-level purpose affects low-level perception, and memory emotions reshape the understanding of information. DIKWP is proposed to model this complexity, which scientifically incorporates the "purpose" into the cognitive process, emphasizing the guiding role of subjective goals on objective cognitive activities, so that artificial intelligence systems have human-like motivation-driven characteristics.

Mesh interaction and semantic dynamic transformation: In the DIKWP model, elements at each level can be input and output from each other, realize perceptual accumulation from the bottom up, realize purpose regulation from the top down, and carry out parallel information exchange at the same or across layers. For example, the "Data → Information" channel represents the meaningful pattern extracted from the original sensory data, the "Knowledge ↔ Information" represents the two-way impact of empirical knowledge on the current information processing, and the "Purpose→ Wisdom" reflects the guiding role of high-level goals in decision evaluation. These 25 basic transformation modules provide a complete semantic dynamic transformation mechanism for understanding and designing cognitive systems. In the neurophysiological research of Prof. Yucong Duan's team, the five elements of DIKWP have been preliminarily mapped to the main functional anatomical regions of the brain: data D corresponds to peripheral perceptual pathways such as the sensory cortex, information I corresponds to communication pathways such as the limbic system and primary prefrontal lobe, and knowledge K corresponds to long-term memory integration areas such as hippocampus and default mode network (DMN). Wisdom W corresponds to multimodal decision-making areas such as the parietal lobe symphysis and higher prefrontal lobe, and Purpose P corresponds to the motivational control areas such as the medial prefrontal lobe, orbitofrontal cortex, and anterior cingulate gyrus. This brain-semantic mapping shows that each semantic level of DIKWP may find a corresponding physiological basis in the brain, which in turn supports us to implement the mutual translation of "semantic space" and "physiological space" in brain-computer interfaces. This project will further use the mapping model to develop [a]{.underline} decoding method from brain signals to semantic state, as well as a coding method from reverse semantic target to brain stimulation strategy (see the following chapters for details) to realize the semantic layer alignment and transformation of information between brain and computer.

Artificial Consciousness and [ACPU]{.underline} Framework: The DIKWP model provides an interpretable and controllable cognitive framework for artificial intelligence, which is regarded as an important foundation for moving towards artificial consciousness (AC). Under this framework, we introduce [the concept of the Artificial Consciousness Processing Unit]{.underline} (ACPU) as a software/hardware carrier and central controller to implement the DIKWP model. ACPU is essentially a heterogeneous agent architecture that integrates subliminal computing and conscious decision-making: it contains both GPU-like massively parallel processing units for low-level data and information processing, and CPU-like sequential logic units for high-level processing Wisdom and Purpose of inference decision-making, and establish a special semantic store and knowledge graph internally for the representation of knowledge levels. According to the existing design, the ACPU can be divided into multiple functional modules, such as the data processing unit (DPU), which is responsible for sensor data preprocessing and pattern extraction, and converts the original signal into a standardized conceptual representation; The Information Processing Unit (IPU) performs complex feature operations and content classification to refine higher-level structured information; Knowledge Processing Units (KPUs) build internal models and long-term experiences of the environment through learning and induction; Wisdom Processing Unit (WPU) integrates knowledge and value systems for judgment and reasoning, and introduces contextual, ethical and other considerations to form decision-making schemes; The Purpose Processing Unit (PPU) dynamically reconstructs the execution logic according to the top-level target to generate specific action instructions or control flows. Together, these modules realize the semantic processing and transformation of all layers of DIKWP, and architecturally ensure that the AI system "knows why" - each step has a clear semantic meaning and a traceable basis. It is worth mentioning that the relevant patent proposes a "dual loop" DIKWP ×DIKWP architecture: a set of metacognitive loops is added to the basic cognitive process, and the former is self-monitored and regulated. In other words, an ACPU can contain two layers of DIKWP processing: one layer for cognitive activities oriented to external tasks (object layer), and the other layer as a supervision layer to evaluate, reflect, and optimize the state of the former, thus giving the system a preliminary self-awareness ability. This project will draw on this "dual circulation" idea when designing ACPU, so that the system can self-monitor its decoding and decision-making process, detect deviations in time and adjust strategies, and ensure the safety and robustness of closed-loop control.

Model Advantages and Innovation: Compared with traditional black-box machine learning algorithms, the DIKWP artificial awareness model has many advantages: (1) Interpretability: The model decomposes complex decision-making processes into data→ information → knowledge→ and the evolution of Wisdom → Purpose, each step is clearly defined, and can be monitored and audited. This helps to solve the "black box" problem in the current large model, so that every behavior of AI is "well documented", and greatly improves the transparency and credibility of the system. (2) Purpose-driven: By explicitly incorporating "Purpose" into the model, system decisions are no longer driven solely based on data relevance, but are always optimized around preset goals. This ensures that AI behavior is aligned with human expectations for value, reducing the risk of irrelevant or even harmful outputs. (3) Human-machine semantic alignment: The DIKWP model provides a cognitive semantic space shared by humans and machines, allowing machines to represent and interpret information in a way that is close to human thinking. Professor Yucong Duan pointed out that the model builds a common cognitive language between humans and machines, and humans can understand the semantic meaning behind each step of AI reasoning. This is especially critical for brain-computer interfaces – it means that machines can "understand" the brain's language and respond in a way that humans can understand. (4) Adaptive evolution: The network structure and dual-loop architecture give the system the ability to self-regulate, and can continuously correct the internal state according to feedback, so as to achieve continuous learning and optimization. In summary, the DIKWP artificial consciousness model has laid a solid theoretical foundation for us to build a new generation of intelligent closed-loop BCI system, which is expected to break through the bottlenecks of semantic understanding, purpose decision-making and self-evolution of previous systems, and truly realize strong semantic perception and purpose-driven regulation.

Project innovation

This project combines the above theoretical innovations with ultrasound brain-computer interface, and intends to break through the bottleneck of current brain-computer interaction, and the main innovations are reflected in the following aspects:

• **1. Ternary semantic conversion mechanism of brain-ultrasound system-artificial consciousness: Establish a dynamic transformation model between brain signals, ultrasound media, and DIKWP artificial consciousness semantic space. This paper clarifies the representation of brain activity at the semantic level of "data-information-knowledge-wisdom-purpose", and the mapping and reduction mechanism in the machine semantic space through ultrasound interface, so as to realize the unified framework of decoding brain signals to semantic information and encoding from semantic purpose to brain stimulation.

• 2. ACPU hub and ultrasound linkage control for artificial consciousness: The artificial consciousness processing unit (ACPU) is introduced as the semantic center of the system, and the artificial consciousness model is integrated for closed-loop control decision-making. The ACPU will be tightly coupled with high spatiotemporal resolution ultrasound hardware to achieve purpose-driven stimulation strategy generation. Through ACPU's semantic understanding of brain states and prediction of future evolution paths, intelligent decision-making is made on when and where to apply ultrasound stimulation to optimize the evolution of brain states and achieve the purpose of autonomous regulation and precise intervention.

• **3. Bidirectional mapping model of semantic space and physiological space: Constructing "semantic space (DIKWP)" andA mapping model between physiological spaces (cerebral blood flow, neural activity patterns). In terms of decoding, a mapping from physiological signal features to semantic states (e.g., emotion, purpose) is established. In terms of regulation, the rules for the transformation from the target semantic effect to the specific stimulus parameter are defined. The mapping model will act as a bridge to ensure that the decoding and stimuli follow a common semantic coordinate system, so that the closed-loop system can use the "understood" information directly for the "intervention". For example, the model can be used to interpret the pattern of brain activity obtained from functional ultrasound imaging as a cognitive/emotional state, and then in turn to generate an ultrasound stimulation program based on the desired state change.

• 4. Unified protocol standard for brain semantic decoding and stimulus coding: A unified brain-computer semantic interaction protocol is proposed to standardize the semantic representation of brain signals and the semantic coding of stimulus commands. The protocol will cover the description format of semantic content at all levels of DIKWP, the machine-understandable semantic tag system, and the interface specification of decoding/encoding conversion to ensure the consistency of semantic information transmission between different modules. The protocol design fully considers the implementation of software and hardware, so as to efficiently deploy on domestic brain-computer interface chips or special processors, and promote the formation of an independent and controllable standard system for brain-semantic interaction.

• **5. DIKWP-AI Adaptive Intervention Pathway for Anxiety/PTSD and Other Diseases: **Focus on exploring the application prospect of ultrasound neuromodulation in emotion-related diseases such as anxiety disorders and post-traumatic stress disorder. Based on DIKWP artificial intelligence, the central system will evaluate the patient's brain semantic state (such as stress level, fear memory reproduction, etc.) in real time, and adaptively adjust the ultrasound stimulation parameters to achieve an individualized closed-loop intervention path. For example, when a semantic pattern of excessive anxiety is detected in the patient's brain, the system automatically triggers ultrasound stimulation for the amygdala/limbic system to alleviate symptoms. Through a large number of preclinical experiments and preliminary clinical trials, the efficacy and safety of this DIKWP semantic-driven AI intervention in alleviating pathological anxiety have been verified, and a new intelligent and adaptive therapy for the treatment of neuropsychiatric diseases has been verified.

The above-mentioned innovations cover the complete chain from basic theoretical models to engineering implementation to clinical applications, reflecting the depth of exploration in the multidisciplinary aspects of this project. Through these innovations, we will create an unprecedented self-interpreting closed-loop ultrasound brain-computer interface system, leading a new paradigm of brain-computer interaction from signal-driven to semantic-driven.

Core system architecture design

In this architecture, the high spatiotemporal resolution ultrasound system plays a dual role. On the one hand, it utilizes functional ultrasound imaging techniques to obtain fine-grained images or signals of brain activity. For example, when the system is used for anxiety monitoring, ultrasound can detect hemodynamic changes in the limbic system (amygdala, hippocampus, etc.) and capture the occurrence of stress responses with millisecond temporal accuracy. These raw data are equivalent to the "data (D)" layer input in the DIKWP model, which needs to be transmitted to the upper-level ACPU for further semantic extraction due to unprocessed and lack of clear meaning. On the other hand, the ultrasound system also acts as an actuator: with the help of focused ultrasound stimulation technology, ultrasound pulses of specific parameters (frequency, intensity, pulse pattern, etc.) are applied to selected brain areas, directly altering the neural activity in that area. For example, ultrasound stimulation may relieve anxiety in an overactive amygdala by inhibiting nerve firing or modulating synaptic plasticity. Ultrasound stimulation corresponds to the "action execution" and can be seen as the realization in physical space of the external behavior derived by the "Purpose (P)" layer. It is important to note that imaging and stimulation are usually not performed at the same time, but rather in the working mode of the multiplex: the system can alternate ultrasound imaging (reading out brain activity) and ultrasound stimulation (writing regulation) on millisecond time slices, thus guaranteeing both continuous monitoring of the brain state and the ability to exert feedback control at a sufficiently fast frequency.

The Artificial Consciousness Processing Unit (ACPU) is the brain of the whole system, which runs the DIKWP artificial consciousness model inside it to realize the semantic level analysis of signals from the brain and intelligent decision-making on regulatory strategies. Specifically, when the data stream from the ultrasound system enters the ACPU, it first undergoes data → information conversion: the underlying data processing module denoises and extracts the original signal, and organizes it into a mode with preliminary significance (corresponding to the I layer of DIKWP). For example, a series of amygdala blood flow signals can be processed to extract patterns such as "tension/relaxation". Subsequently, the information → knowledge conversion module combines context and memory to comprehensively analyze these patterns and determine their higher-level meanings (corresponding to K-layer). In the above example, the system may combine other physiological signals and historical data to determine that the user is currently in a state of knowledge such as "triggering traumatic memory" and "being in the early stage of stress response". Next, the ACPU enters the Wisdom (W) layer for decision reasoning: this layer synthesizes the expert knowledge built into the system, the ethical constraints, and the requirements of the current Purpose (P) layer to evaluate possible interventions. In our system, the Purpose (P) layer is usually composed of a combination of pre-set treatment goals and a real-time updated user Purpose. For example, for patients with PTSD, the goal layer includes goals such as "relieving panic attacks" and "avoiding excessive suppression of normal emotions". Based on this, the Wisdom layer weighs the pros and cons of different intervention options, such as "immediate stimulation of the amygdala to reduce its activity" versus "diversion of the patient's attention", and gives the best strategy. The Purpose (P) layer will finally generate specific ultrasound stimulation instructions based on the recommendations of the Wisdom layer, which will be executed by the ultrasound system. It should be emphasized that the ACPU is not a simple feedforward chain in the working process, but is equipped with a metacognitive monitoring mechanism: it continuously monitors the semantic correctness and consistency of the output of each layer (for example, whether the inference of the knowledge layer is consistent with the original evidence of the information layer, whether the decision of the Wisdom layer is in line with the ultimate goal), and once a deviation is detected, it will adjust the processing parameters of the previous layer through feedback signals to achieve self-calibration. This design corresponds to the above-mentioned "dual circulation" architecture, that is, the second set of DIKWP loops inside the ACPU supervises and regulates the first set to ensure that the system decision-making is robust, reliable, transparent and controllable.

Closed-loop control process: Combining the above components, the closed-loop interaction process of the system is as follows: (1) Data collection: The neural activity generated by the brain is first obtained by ultrasound imaging to obtain real-time data, and after preprocessing, a structured information flow is formed and sent to the ACPU. (2) Semantic decoding: ACPU processes the information stream layer by layer, transforming it into a high-level semantic description of the current brain state within the DIKWP semantic space (e.g., identifying "the patient is panicking"). (3) Strategic decision-making: Based on the current semantic state and treatment goals, ACPU deduces the future state change path with the support of the internal knowledge base and artificial consciousness model, and selects the best intervention strategy. (4) Instruction encoding: The selected strategy is encoded as a specific sequence of ultrasound stimulation parameters (target, sound intensity, duration, etc.) as ACPU output. (5) Execution and feedback: The ultrasound system receives instructions and exerts stimulation to the corresponding brain area, causing changes in brain activity; At the same time, brain signals continue to be monitored, and changes are fed back to the ACPU. (6) Cyclic iteration: ACPU compares the brain semantic state before and after stimulation, evaluates the intervention effect, and adjusts the strategy to enter the next cycle if the goal is not reached, until the termination condition is met (symptom relief or end of the preset course of treatment). Through such closed-loop iteration, the system can "read" the patient's brain like a skilled therapist at all times, and "prescribe the right medicine" to achieve highly intelligent brain state regulation.

In short, this architecture integrates the hardware tool of high-resolution ultrasound with the software Wisdom of DIKWP artificial consciousness to form a complete closed loop of "perception-cognition-action". It breaks through the limitations of traditional brain-computer interfaces that only use one-way information reading or preset stimuli, and instead creates a new type of agent that can understand brain semantics, actively make decisions and govern, and self-supervise evolution. Below we will further break down the stage tasks, technical roadmap and key module design of the project implementation.

Study content and task decomposition in stages

In order to promote the research of this project in an orderly manner, we divide the task into three stages according to the difficulties and logical relationships, and each stage focuses on overcoming a number of key problems to achieve the project goals step by step.

1. Phase 1 (Theoretical Modeling and Key Technology Research, Estimated 1-2 Years):

• DIKWP Semantic Modeling: Improve the theoretical model of brain-ultrasound-artificial consciousness interaction. Based on the existing DIKWP semantic mathematical framework, a semantic representation method of brain signals was established to define the correspondence and transformation functions between each DIKWP level and typical neural activity patterns. The mathematical representation of the brain semantic state space is formed, which provides a theoretical basis for the subsequent algorithm development.

• Development of semantic decoding algorithms: Development of multi-level brain signal decoding algorithms for high spatiotemporal resolution ultrasound data. Including: limbic system emotional signal extraction, whole brain functional network dynamic analysis, semantic feature extraction, etc., to extract useful information from each layer of DIKWP from massive time series. Try to use the combination of machine learning/deep learning and knowledge graph to integrate data-driven and prior knowledge to improve the accuracy of semantic decoding.

• Stimulation Strategy Optimization Model: Study the purpose-driven decision-making mechanism and establish a mapping model from semantic goals to stimulus parameters. Reinforcement learning or optimization algorithms are used to simulate the influence of different stimulation schemes on the evolution of brain states, and a stimulus strategy optimization framework is formed. In particular, consider incorporating security constraints and ethical rules into decision-making (the Wisdom layer) to ensure that the resulting policies are safe and feasible.

 Phase 2 (Prototype System Development and Simulation Validation, Expected Year 2-3):

• Closed-loop system prototype integration: Develop hardware and software prototypes to integrate the ACPU software platform with ultrasound imaging/stimulation devices. Write interface programs to realize data collection, instruction sending, and real-time communication, and initially build a closed-loop operating environment. The human-computer interface is designed to monitor the semantic state and decision-making process within the system, and to facilitate debugging and demonstration of its interpretability.

• Simulation Platform Testing: Simulation validation of prototype systems in vitro/offline environments. Firstly, a virtual brain environment (such as a digital twin based on computational models or offline physiological data) was established to simulate various controllable brain state change scenarios (such as progressively enhanced anxiety and stress responses), and the accuracy and response speed of ACPU decoding these states were verified. Then, software in-the-loop simulation was introduced to simulate the impact of ultrasound stimulation on the virtual brain model, and the convergence and stability of closed-loop decision-making were tested. If possible, further ex vivo tissue experiments (e.g., brain slices or cell models) are carried out to validate the ultrasound stimulus-neural response relationship to lay the groundwork for application in overall biological systems.

• Key Indicator Evaluation: Evaluate the performance indicators of the prototype system in the simulation environment, including: decoding accuracy (the accuracy of semantic state recognition), decision delay (the time taken by a loop in a loop), stimulus control accuracy (the deviation of the degree of brain state regulation relative to the target), etc. Algorithms and parameters are continuously improved in response to identified issues. If the decoding accuracy is not up to standard, the feature extraction method will be optimized, and if the decision delay is too long, the ACPU calculation process will be optimized or hardware acceleration will be used to ensure that the prototype has the performance level to enter the next animal experiment.

 Phase 3 (Animal Experiments and Initial Clinical Validation, Expected Year 4-5):

• Animal model experiments: Select appropriate animal models (such as rodent mice or rats) to carry out in vivo closed-loop regulation experiments to verify the effectiveness and safety of the system in real biological systems. Animal models of stress disorder (e.g., mice that are dosed to induce anxiety-like behaviors) can be implanted or fixed with an ultrasound probe to monitor the ultrasound signal in anxiety-related regions of the brain (amygdala, hypothalamus, etc.) in real time. When an abnormal anxiety state is detected in the animal, ultrasound stimulation is triggered by the ACPU decision to intervene. To observe the effects of the intervention on behavioral and physiological indicators (e.g., heart rate, stress hormone levels) and to verify changes in brain activity by neuroimaging methods. Key evaluations: (1) the accuracy of the system in detecting abnormal brain states, (2) the effectiveness of closed-loop intervention in alleviating abnormal behaviors, and (3) the safety of ultrasound stimulation on tissues under long-term repeated action (with microbleeding, inflammation, etc.). According to the experimental results, the system parameters such as stimulation dose, monitoring threshold, ACPU decision rules, etc. were adjusted.

• Preparation for preliminary clinical trials: On the premise of ensuring the safety and effectiveness of animal experiments, we will cooperate with medical institutions to carry out trial or preclinical testing of small sample volunteers. A small number of patients with anxiety/PTSD who are refractory to medication were selected to participate in the trial with informed consent. For the sake of safety, an intervention model that is acceptable to the population, such as the use of ultrasound closed-loop adjustment during the patient's psychological exposure therapy, can be used to alleviate the excessive stress response in real time. The effect of the system on humans was evaluated step by step, including changes in symptom scores, subjective feelings of the patient, and any discomfort. Collect feedback from doctors and patients, improve the human-computer interaction interface and use process, and formulate a perfect plan for formal clinical trials.

• Clinical Translational Analysis: Evaluate the advantages and potential benefits of the system over traditional treatments based on trial data, such as improvement in anxiety scores, onset time, patient compliance, etc. At the same time, the challenges that may be faced in clinical translation (such as model generalization problems caused by individual differences, device portability and cost issues, regulatory and ethical requirements) are identified, and corresponding solutions and improvement routes are formulated. For example, consider using a customized head-mounted ultrasound device to improve convenience, or introduce more patient data to retrain the ACPU model to improve robustness. Make technical and strategic reserves for subsequent large-scale clinical applications.

Through the implementation of the above three phases, we will gradually turn the proof of concept into a practical system: from theoretical models and algorithms, to closed-loop prototypes, to biological experiments and clinical exploration, and finally to achieve the project goals. At the end of each phase, we conduct a milestone assessment to ensure that the progress and quality of the study is as expected, and to provide clear guidance for the next phase.

Technical route and scheme design

In order to achieve the goal of the project, we comprehensively use multidisciplinary methods such as brain science, artificial intelligence, and ultrasound engineering, and design the following technical routes to ensure the systematization and efficiency of the research:

• (1) Brain semantic feature extraction: Firstly, multi-level feature extraction and representation are performed on brain signals from ultrasound imaging. The combination of signal processing and deep learning is used to extract the feature set that can characterize the semantics of each layer of DIKWP. For example, low-level data features are obtained by frequency domain analysis and wavelet transform on the original time series. Use graph algorithms or network analysis to extract network features of the knowledge layer from the functional connection matrix; A pre-trained neural network is used to extract high-level features related to emotion/cognitive state from brain images, corresponding to the Wisdom/Purpose layer, etc. This step focuses on maximizing the retention of valid information in the brain signals and reorganizing them in a meaningful way to lay the foundation for subsequent semantic mapping.

• (2) DIKWP semantic mapping modeling: Establish a formal semantic mapping model to map the above multi-level features to the corresponding positions in the DIKWP semantic space. Specifically, the mapping functions $f_D, f_I, f_K, f_W, and f_P$ are designed, which correspond to the process of converting the observed feature vectors into semantic variables at each layer, respectively. For example, $f_I$ maps several low-level data features to an informational semantic label (e.g., "fear" signal appears), $f_K$ further maps the set of information labels to knowledge states (e.g., "recalls a specific traumatic situation"), and so on until $f_P$ gives an estimate of the purpose of the current subject (e.g., "fleeing a potential threat"). These mappings can be learned by machine learning models or constructed in part based on expert knowledge. During the training process, the parameters of these mappings are optimized to accurately reflect the true semantics through the annotated data of known semantic states (such as experiment-induced emotional states).

• (3) Artificial consciousness decision reasoning: After obtaining the DIKWP semantic representation of the current brain state, the ACPU initiates the artificial consciousness model for decision reasoning. This step combines two paradigms: symbolic inference and data-driven prediction: symbolic inference uses the rules embedded in the DIKWP model (such as adjusting the judgment criteria of the Wisdom layer according to the purpose layer) to logically evaluate the current state; The data-driven part uses the trained strategy network or value network to score and evaluate different candidate stimulus actions. In terms of implementation, the actor-critic architecture in reinforcement learning can be adopted, in which the actor network outputs an action (stimulus scheme) based on the current semantic state, and the critic network evaluates whether the action is helpful to achieve the purpose (based on the knowledge within the model and the external reward signal). In addition, incorporating safety rules and ethical constraints (such as the upper limit of stimulus amplitude, not touching non-targeted functional areas, etc.) in the inference process is equivalent to setting boundary conditions for decision-making at the Wisdom layer to ensure that the resulting scheme is feasible and safe.

• (4) Ultrasound stimulation coding and execution: Once the optimal stimulation protocol is determined (e.g., "pulse stimulation to reduce amygdala activity for 3 seconds"), the system enters the action implementation phase. According to the requirements of the protocol, it is coded into specific groups of ultrasound stimulation parameters: including the location of action (controlled by the ultrasonic phased array focus), frequency intensity (selected according to physiological effects, such as low-frequency bias neural modulation, high-frequency bias memory intervention), and modulation mode (continuous/pulsed wave, pulse duty cycle, etc.). These parameters are delivered through the drive interface of the ultrasonic device and converted into the electrical signal output of the ultrasonic probe, which in turn generates the desired ultrasonic field at the physical level. In this step, it is necessary to solve the parameter inversion problem, that is, calculate the corresponding ultrasound technical parameters according to the desired biological effect. This can be done through simulation and experimental calibration: a pre-established response curve or database of ultrasound parameters to neural effects (e.g., the intensity of neural excitation/inhibition effects at different frequencies) for decision-making table lookup or interpolation. At the same time, ensure that the stimulus sequence is within the range of hardware real-time, and if necessary, the signal is optimized to avoid hardware bottlenecks (such as limiting the switching frequency, loading waveform buffers in advance, etc.).

• (5) Closed-loop evaluation and adjustment: After the stimulation is performed, the system re-enters the monitoring mode to evaluate the changes in brain state. If the goal has not been reached, move on to the next closed loop. The key in this process is the feedback evaluation algorithm: by comparing the differences in DIKWP semantic vectors before and after stimulation, the intervention effect is quantitatively evaluated. We'll define a set of evaluation metrics, such as "semantic error" $E = | S_{current} - S_{target}|$ indicates the distance between the current semantic state and the target state. If $E$ is insufficient, the algorithm considers whether to adjust the strategy: for example, to increase the intensity of the stimulus, to change the stimulus site, or to infer that the current hypothesis (knowledge-level judgment) may be wrong if multiple cycles are ineffective, and the interpretation of the brain state needs to be reverted and re-amended (triggering metacognitive feedback of the ACPU). Through this closed-loop adjustment mechanism, the system can continuously modify its own behavior and approach the optimal treatment pathway.

• (6) Human-machine interface and manual intervention: Although the system is highly automated, we still reserve the interface for manual monitoring and intervention. At each step of the technical route, the internal status can be projected to the operator. For example, visualizing the semantic labels currently decoded, ACPU decision justification (why the stimulus was chosen), etc., to achieve truly explainable AI-assisted medicine. Doctors can intervene as needed: for example, when the system misjudges the purpose, the doctor can manually correct the semantic tag; or adjust system goals based on patient subjective feedback. This human-robot collaborative interface design ensures medical safety and ethical requirements, and also provides a valuable basis for further optimization of algorithms.

The above-mentioned technical routes form a closed loop from information acquisition, semantic modeling, intelligent decision-making to executive feedback. Each step is relatively independent and closely connected, forming an overall solution to the problem. In the implementation process, we will concretize each link according to the phase focus, such as the first phase is mainly completed (1) (2), the second stage is improved (3) (4), and the third stage is tested (5) and (6) and integrated into human supervision. Through multiple rounds of iterative development and verification, the performance and adaptability of each link are gradually optimized, and a stable and efficient closed-loop system is finally realized.

Design of simulation system and experimental platform

Building a high-fidelity simulation system is critical to validating and refining our protocol before moving into real-world biological experiments. This project will develop a hardware-in-the-loop simulation platform to simulate the closed-loop process of real brain-computer interfaces in a virtual environment, so as to find and solve problems at a low cost and high efficiency. The design of the simulation platform considers the following aspects:

1. Virtual brain and environmental modeling: Based on the neurotic point model or neural network simulation, we will build a virtual brain model, focusing on the brain regions and pathways related to anxiety and stress. For example, the Wilson-Cowan model can be used to simulate the dynamics of the amygdala-prefrontal cortex circuit, and its parameters can be adjusted to show a high activity state similar to anxiety attacks. Or leverage an open-source brain simulation platform such as TheVirtualBrain to build neural network models that incorporate limbic systems. The virtual brain model will act as a controlled object for a closed-loop system, capable of receiving simulated "ultrasound stimulus" inputs (e.g., changing the weights of certain connections or neuronal excitability) and generating state changes, while outputting signals that can be acquired by "ultrasound imaging" (e.g., "blood flow" time series in various regions). In addition, we will build a simple virtual body and environment model, so that the virtual brain state can be mapped to behavioral performance (such as anxiety levels affecting imaginary heart rate, exercise behavior, etc.), creating an observable and interactive simulation environment.

2. Ultrasound device model: In the simulation, we need a device model that approximates the behavior of a real ultrasound system. This includes: a beam characteristic model of an ultrasound probe to calculate the energy distribution acting on a specific brain region under a given parameter; The sampling and noise model of ultrasound imaging is used to add noise and resolution limitations to the virtual brain after the "ground truth" signal is generated to obtain a simulated observation signal. We will refer to the literature and experimental data to establish formulas or look-up tables for the propagation characteristics of ultrasound of different frequencies and intensities in brain tissue, and consider the effects of skull attenuation. For the imaging part, the sampling process of spatial resolution (about 100 microns) and temporal resolution (tens of milliseconds) is simulated to ensure that the simulated data and future real data have similar statistical characteristics.

3. ACPU software simulation: We can choose two implementation methods for ACPU running in the simulation environment: one is to directly run the real ACPU software we developed, just connect the input and output to the virtual model to realize the loop-in-the-loop test of the whole system； Second, in order to test more quickly, an accelerated simulation version of ACPU can also be developed, which directly represents the behavior of each module with high-level mathematical functions. For example, a function is used to approximate the mapping of virtual brain signals to the semantic state (skipping the details of the deep learning model) and outputs stimulus decisions based on pre-set policy rules. This fast simulation mode can be used for a wide range of parametric searches and proof-of-concept scenarios. In the early stages of the project, we will mainly use the second way to iterate quickly; After the real ACPU software is formed, switch to the first method for full-link verification.

**4. Data Logging & Visualization: The simulation platform will have sophisticated data logging and visualization tools. In each simulation run, we record all the information about the "true state" of the virtual brain, ultrasound observations, the semantic state of ACPU resolution, the details of the decision-making process, and the final output stimulus. This data will be used for offline analysis to verify that the modules are working as intended. For example, it is possible to verify whether the semantics parsed by the ACPU correspond to the real preset state of the virtual brain (decoding accuracy), whether the stimulus of the ACPU decision actually makes the virtual brain evolve towards the target state (control effectiveness), and so on. In terms of visualization, we will develop interactive dashboards that visualize the changes in each key quantity in the closed-loop process: the activity heat map of the virtual brain model, the curve of semantic variables over time, the option evaluation score for each round of decision-making, etc. These interfaces allow researchers to gain insight into system behavior and facilitate the tuning of algorithms and parameters.

5. Scenario and Case Library: We plan to design a series of representative simulation scenarios (use cases) to form a case library to cover as many situations as possible. For example, these include "escalating anxiety situations", "sudden panic situations", "abnormal situations with incorrect decoding", "multi-objective conflict situations", etc. Each scenario is preset with a different initial state of the virtual brain and external stimuli, and the system is allowed to run in it, testing its robustness and adaptability. By running a large number of simulation experiments, we can count the performance indicators of the system in various scenarios and find possible problems under extreme conditions. For example, if the system does not respond in a "sudden panic" scenario, we need to improve the real-time performance of the decision-making algorithm. If the error correction cannot be made in time in the "error decoding" scenario, the sensitivity of metacognitive monitoring needs to be strengthened.

Taken together, the simulation platform will continue throughout the project, serving as a "virtual proving ground" to continuously test and optimize our designs. Before key algorithms are put into animal and human experiments, try to use simulation to find and solve problems to reduce R&D risks and costs. Once the simulation results are satisfactory, we are more confident of success in the real world. At the same time, the empirical output of the simulation itself (such as simulation models and datasets) is also a by-product of the project, which can provide a reference for further research in the academic community.

Interpretability of the model mechanism

Explainability is a major feature and necessary requirement of this project. In a complex system that integrates artificial consciousness and closed-loop brain-computer interfaces, ensuring that every step of the decision is transparent and solvable for humans is not only helpful for R&D and debugging, but also a prerequisite for clinical application. We'll start with both model design and tool support to create a self-explanatory system.

DIKWP Semantic Transparent Design: The DIKWP model naturally has a hierarchical semantic structure, which allows the AI decision-making process to be disassembled into well-known cognitive steps. We will strictly follow the DIKWP architecture inside the ACPU, with clear semantic definitions and mathematical descriptions for each layer of processing. For example: (a) the input and output of the data layer are the raw signal and its simple statistical characteristics; (b) The information layer outputs events or patterns with semantic tags, e.g. "increased amygdala activation"; (c) Knowledge level output of awareness of the current situation, e.g. "stress response may occur due to recall of a traumatic situation"; (d) Wisdom layer outputs the basis for decision-making, e.g., "given that the patient is in the early stages of panic, immediate intervention should be made to prevent symptoms from worsening"; (e) The Purpose layer outputs the final action plan, e.g., "stimulation of brain region Y with X parameters for 3 seconds". Each of these intermediate results has a specific meaning, for which we design a formal representation (e.g. with readable labels or graphical representations) for presentation and review in the HMI. Through this layer of traces, every step of the system's reasoning can be traced back to the source and the reason can be understood. When there is an unexpected outcome, we are able to pinpoint where the problem lies (perceptual error?). Incomplete knowledge base? Improper decision-making rules? ) for targeted optimization.

"White Box" Evaluation Framework: We draw on the white box assessment method proposed by Yucong Duan's team to develop a set of interpretable evaluation indicators output by the artificial consciousness model. For example, for each closed-loop decision, the following assessments: (1) consistency: whether the outputs of each layer support each other and are not contradictory (e.g., the information layer detects "fear" but the knowledge layer judges "safety" inconsistently); (2) adequacy: whether the decision makes full use of known knowledge and data, or whether important factors are omitted; (3) comprehensibility: whether the explanations given by the system are expressed in concepts that humans can understand; (4) Necessity: whether each step of reasoning contributes to the final decision, and whether it can be deleted. Through the analysis of a large number of cases, we can quantify the interpretability of the model and guide improvement. For example, if we find that a layer often outputs internal codes that are difficult for humans to understand, we adjust the representation of that layer to more intuitive semantic notation.

Real-time visualization and user interaction: We will develop an interpretive interface for researchers and clinicians to see the internal state of the system in real time. The interface, or "semantic dashboard", includes: the raw waveform of the current brain signal and ultrasound images; Textual/graphical descriptions of the output of each layer of the ACPU, such as the currently recognized emotional state, the reason inferred by the system, the stimulus chosen and its intended effect; Confidence indicators, such as decoding confidence, decision confidence, etc. In addition, the interface is interoperable: users can question or adjust assumptions about the output of a layer, such as modifying a hypothesis of the knowledge layer to see how it affects the decision. This human-machine interaction helps to validate the internal mechanisms of the model. For example, doctors can adjust the Purpose goal to see how the system stimulates the regimen and how it is appropriate to determine the sensitivity and rationality of the treatment goal. If the explanation displayed on the interface is contrary to the doctor's professional knowledge, the doctor can intervene in time and pause the automatic mode to avoid potential risks. This transparent presentation and manual correction mechanism greatly enhances the safety and reliability of the system in clinical application.

Editability of knowledge base and rule base: ACPU depends on certain knowledge rules (especially in the Wisdom and Destination layers). To prevent these rules from becoming the new "black box", we will design the knowledge base to be editable modules. Specifically, the knowledge base will store medical and cognitive knowledge in a structure similar to a knowledge graph or ontology, such as "hyperventilation can exacerbate panic" and "high-frequency stimulation of the amygdala can inhibit its activity". This knowledge is entered and verified by a team of experts before the system goes live, and can be updated as new clinical discoveries are made. The rule base will also contain reasoning and decision-making rules, such as priority in case of multi-objective conflicts, rules for handling exceptions, etc., which will also be recorded. Our system will provide tools for experts to review and modify this knowledge/rules to suit different patients or applications. Through the intervention of artificial knowledge, combined with the patterns that machines learn from data, we strive to achieve a "controllable and evolving" artificial consciousness: one that does not deviate from the common sense of human beings, and can constantly improve itself.

Continuous model monitoring and learning: During the actual operation of the system, we continuously monitor the quality of its interpretation output. On the one hand, the feedback of doctors and patients on the behavior of the system is collected as an important basis for model adjustment. On the other hand, the correctness and usefulness of the explanations were evaluated by comparing the systematic explanations with the objective results (e.g., the system predicted that the patient's panic was about to occur and whether it actually occurred). If a certain type of explanation is found to be often unreliable, the corresponding part of the model will be analyzed and improved in depth. In addition, we will implement online learning capabilities that allow the model to learn better interpretations from the new data acquired in the run. For example, after multiple patient interactions, the system may learn new patterns of triggers, which can be incorporated into the knowledge base to make subsequent explanations more accurate and comprehensive.

In summary, this project implements the concept of "transparency, comprehensibility and controllability" in the model mechanism. Ensure that every step of the AI is meaningful from the design source, and ensure that humans can see and influence AI decisions in real time. This highly explainable AI, combined with brain-computer interfaces, not only facilitates scientific debugging, but more importantly, provides clinicians and patients with a foundation of trust that they can know what machines are doing and why, so they can confidently hand themselves over to system-assisted treatment. This will greatly improve the acceptance and success rate of our system in real-world medical scenarios.

Interface specification and hardware scalability

To realize the semantic interaction of closed-loop ultrasound brain-computer interface, in addition to the core algorithm model, this project will also formulate a unified interface specification and pay attention to hardware adaptability to ensure that the system has good scalability and engineering landing potential, especially to adapt to domestic chips and devices.

Brain-Computer Semantic Protocol Standard: We plan to propose the "Brain Semantic Communication Protocol (BSCP)" to unify the definition of the semantic expression of brain signals and the coding format of stimulus instructions. The core of this protocol is to introduce the concepts of semantic tags and semantic metadata into brain-computer interactions. For example, define standard codes for different brain states and cognitive semantics (similar to the event code of EEG atlas, but rise to the semantic level), such as using <EMOTION:FEAR> to represent fear emotion semantics, and <INTENT:ESCAPE> to represent the purpose semantics, etc.; At the same time, it is stipulated that the stimulus instruction is also accompanied by the target semantic identifier and the effect description, such as STIM[Target=Amygdala, Goal=CALM] indicates that the amygdala is administered to the amygdala for the purpose of calming emotions. The protocol will include a message format (e.g., JSON/XML structure or binary frame format) covering the following fields: timestamp, semantic label set (each DIKWP layer can have separate fields), confidence, priority, data validation, etc. In the decoding direction, the ultrasonic signal is parsed by ACPU and then packaged as a BSCP packet and transmitted to the decision-making module. After the coding direction and decision is made to generate a stimulus protocol, it is also sent to the ultrasound device driver in the form of a BSCP message. The advantage of this design is the decoupling of modules and the unification of standards: brain signals from different sources (not only limited to ultrasound, but also EEG, fNIRS, etc. can also be accessed in the future) can be abstracted into BSCP semantic output; Different types of stimulation devices (ultrasound, electrical stimulation, etc.) can also perform corresponding functions according to the BSCP command field. We will compile detailed protocol specification documents and sample libraries, open them to the community for use and improvement, and promote the evolution of brain-computer interfaces from "signaling layer protocol" to "semantic layer protocol".

Software and hardware interfaces and middleware: In order to make the BSCP protocol truly applied to the system, we will develop a set of middleware or APIs to facilitate the call of each module. It includes: data interface library (responsible for converting the drive data format of ultrasound equipment to BSCP format and reverse conversion), network communication module (supporting local and remote BSCP message transmission, and efficient communication frameworks such as ZeroMQ can be selected), and parsing library(The ACPU can quickly parse BSCP packets into internal objects or serialize internal objects into packets.) The middleware will be implemented in a cross-platform language (e.g. C++/Python hybrid) to ensure that it can be used in different runtime environments. For real-time requirements, we will optimize the middleware, such as using shared memory, RingBuffer and other mechanisms to reduce latency and ensure the high speed of closed-loop control.

Consideration of domestic chip adaptation: China is vigorously developing independent and controllable artificial intelligence chips and brain-computer interface hardware, such as Cambrian NPU, various RISC-V architecture processors, etc. This project attaches great importance to the feasibility of combining technical routes with domestic hardware. When designing the ACPU algorithm, we will consider its computing power requirements and parallelization characteristics, and try to make use of the characteristics of domestic AI chips. For example, for deep learning inference, model optimization (cropping and quantization) and inference framework adaptation can be performed for Cambrian MLU and Huawei's Ascend NPU. For symbolic inference, the RISC-V architecture's open instruction set expansion capability can be used to design customized instructions to improve the efficiency of logical inference. For the ultrasonic signal processing part, it can be combined with domestic DSP or FPGA to achieve high-speed filtering, feature extraction, etc. In addition, we will pay attention to the implementation of the DIKWP theory at the hardware level. The research of Yucong Duan's team has proposed a DIKWP-oriented computer architecture design, in which each semantic layer is mapped to dedicated processing units, such as DPU, IPU, KPU, etc. This implies that we can implement a modular architecture on the chip that corresponds to the DIKWP model to accelerate the artificial consciousness algorithm from the hardware layer. For example, in the future, DIKWP chips can be developed, which integrate data stream processing accelerators, knowledge storage units, Purpose decision logic, etc., to support semantic layer operations. This project will maintain communication with relevant chip R&D teams to explore the possibility of building ACPU prototypes on existing FPGAs or SoCs. If conditions permit, we plan to transplant the core algorithm of ACPU to a domestic embedded board for experimental demonstration in the later stage of the project to verify the performance of the system on low-power and high-integration hardware. This will lay the foundation for subsequent commercialization and clinical application.

Security and compatibility: Security and standards compatibility should also be considered in the design of interface specifications. For example, we incorporate permissions and encryption mechanisms to prevent unauthorized instruction injection or data eavesdropping. BSCP messages can be transmitted in conjunction with existing security protocols (e.g., TLS) and authenticated to ensure that medical data and control instructions are not tampered with. In terms of compatibility, we should try to integrate with the existing brain-computer interface communication protocol standards, such as the IEEE's draft brain-computer interface standard, the general interface of the BMI system, etc., and expand the semantic layer content on the basis of it, rather than completely starting from scratch. This will make the results of this project more acceptable and integrated by the industry. At the same time, we will provide different complexity versions of the protocol: for example, a lite version for compute and bandwidth-constrained chips, and an extended version for feature-rich servers. This hierarchical design allows the protocol to be backward compatible according to the application.

By formulating a unified interface protocol and focusing on the adaptation and optimization of domestic hardware, we hope to ensure that the results of this project are not only a laboratory prototype, but also standardized, modular, and easy to promote. In this way, after the completion of the project, our technology can be integrated into the national new generation of brain-computer interface industry chain more quickly, and combined with domestic equipment to form a complete independent and controllable solution. This is of great significance to seize the commanding heights of the formulation of international brain-computer interface standards and ensure the safety and controllability of relevant technologies in China.

Risk assessment and risk response

In view of the cutting-edge and complex nature of this project, we have identified the following key risk points and developed corresponding prevention and response measures to ensure the smooth achievement of the project objectives:

• Risk 1: Ultrasound Signal Quality and Safety Risks – Description: The penetration and reflection characteristics of brain tissue to ultrasound are complex, and the high attenuation of the skull may lead to insufficient imaging signal quality and affect the decoding accuracy. At the same time, there are potential safety concerns due to the prolonged effect of ultrasound, such as excessive power that may cause tissue heating or microbubble damage. Response: (1) At the signal level, optimize the ultrasound imaging parameters (frequency, emission angle, etc.) and use signal enhancement algorithms (such as image reconstruction, denoising and template matching) to improve the signal-to-noise ratio; If necessary, consider multimodal fusion, such as in combination with EEG to assist in correcting the ultrasound signal. (2) In terms of safety, strictly abide by international ultrasound safety standards (such as mechanical index MI and temperature rise index TI limits), continuously monitor the temperature and structural changes of the affected tissues in animal experiments, and adjust the strategy immediately if there is an abnormality. The safety monitoring module is introduced to calculate the cumulative ultrasonic dose and estimated temperature rise in real time, and the safety threshold is included in the ACPU decision-making constraints. Long-term continuous stimulation is replaced by a low-dose and multi-frequency stimulation regimen to reduce the energy accumulation of a single session. At the same time, the project will work with medical imaging experts to evaluate the possible biological effects of ultrasound and control the safety red line.

• Risk 2: Insufficient accuracy of brain semantic decoding – Explanation: The decoding of high-level semantic states of the brain (such as emotion, purpose) is a very challenging problem in the field of AI, and individual differences and environmental noise may lead to insufficient accuracy of the decoding model and misjudgment. If the ACPU incorrectly interprets the normal state as abnormal, it will cause unnecessary irritation. Response: (1) Increase the diversity and quantity of training data: In the algorithm development stage, collect as much as possible data on the semantic correspondence of brain signals and behaviors, including different individuals, different situations, and different emotional levels, so as to improve the generalization ability of the model. (2) Adopt active learning strategy: In the actual use of the system, the decoding model is constantly adjusted according to the doctor's feedback and the patient's subjective report. (3) Set up multi-layer redundancy check: The hierarchical structure of the DIKWP model is used to cross-verify the reliability of the results, for example, multiple independent features of the information layer need to point to the same emotional conclusion to trigger the judgment of the knowledge layer; Before making a decision at the Wisdom layer, the metacognitive module checks the support of the low-level evidence, and if the evidence is insufficient, it postpones the decision to request more data. (4) Allow manual intervention: In closed-loop clinical applications, human safety officers (such as clinicians) are introduced to monitor the decoding results in real time, and when the system interprets the semantics that are obviously inconsistent with the actual situation (the doctor can observe and verify from the patient's performance), he has the right to suspend or correct the system decoding to avoid false stimulation. Through the above measures, the false positive rate is reduced to an acceptable range as much as possible, and even if a false positive is made, it will not directly harm the patient.

• Risk 3: Closed-loop control is unstable or deviates from expectations – Description: Adaptive closed-loop systems may exhibit unexpected dynamic behaviors, such as oscillations, divergence, or falling into a local suboptimal state, leading to regulatory failure or even worsening of patient status. This can be caused by inaccurate models, excessive time delays, poorly designed control strategies, etc. Solution: (1) Strict control theory analysis: When designing stimulus control strategies, control theory experts are introduced to analyze the stability of the closed-loop system, including establishing a mathematical model to solve the closed-loop characteristic equation and analyzing the influence of parameters on stability. Make sure that the control gain and feedback mechanism chosen will allow the system to converge. If necessary, a conservative progressive control strategy can be used, such as observing feedback with only a small stimulus for the first time, and gradually approaching the target. (2) Implement real-time monitoring and oscillation detection: The ACPU metacognitive module monitors the closed-loop behavior pattern, and if it detects repeated fluctuations up and down in the same state (oscillation signs) or the error does not decrease but rises (divergence signs), it will trigger a safe shutdown or switch to manual control, and mark the problem for post-event analysis and improvement. (3) In the simulation and animal experiment stages, the dynamic response of the closed-loop under different model parameters is extensively tested, and the algorithm is adjusted to enhance the robustness of the system to uncertainty. For example, the introduction of adaptive gain or fuzzy control methods allows the control force to be adjusted with the error and rate of change to prevent overshoot and back-and-forth oscillation. (4) Preset emergency exit mechanism: If the system fails to converge to a reasonable range within a certain period of time (such as more than double the expected control time), it will automatically stop the current closed-loop process, issue an alert and wait for manual decision. Through the above means, the closed-loop control process is stable and reliable, and the risk of deviation is reduced.

• Risk 4: ACPU implementation complexity and performance bottlenecks – Description: ACPU needs to perform a large number of calculations, such as deep learning inference, symbolic reasoning, and real-time control decisions, and may face performance bottlenecks, resulting in real-time operation. At the same time, its development complexity is high, and multi-module collaboration is prone to bugs. Response: (1) In terms of architecture, adopt modular and parallel design. Take full advantage of multi-core CPU and GPU parallelism, process decoding, decision-making pipelines in parallel, and use FPGA/ASIC to accelerate critical parts when necessary. The computing power requirements of each module are evaluated in advance, and the algorithm is optimized to an acceptable complexity. For example, compressing depth models, employing heuristic rules to reduce search space, and so on. (2) Strengthen software engineering management: strict version control, equal emphasis on module unit testing and integration testing, formal verification of key algorithms and interface designs, and reduce logical loopholes. (3) Formulate a downgrade strategy: In special cases, if the ACPU calculation is too slow or error, you can switch to simplified mode. For example, only some key layers of inference are performed, or a preset fixed stimulation program can be enabled to ensure that the basic functions of the system are not interrupted. (4) Synchronization with hardware R&D: Once the software prototype verification is successful, we will start the adaptation with domestic chips as soon as possible, and offload part of the computing to special hardware to improve the speed. At the same time, we should consider algorithm alternatives when resources are limited, such as replacing deep networks with shallow networks on embedded platforms. Through both technology and management, we ensure that the ACPU is controllable, measurable, and operates reliably within the required performance range.

• Risk 5: Clinical Translation and Ethical Risks – Description: Although technically feasible, actual clinical application may still encounter risks such as personnel cooperation, ethical approval, policies and regulations, etc. For example, patients and doctors have trust in "artificial awareness" interventions, ethics committees have safety concerns about adaptive AI decision-making, and medical device regulatory authorities have requirements for algorithm transparency and controllability. Response: (1) Strengthen multidisciplinary cooperation: Clinicians and ethicists are invited to participate in the project process to guide and evaluate the program to ensure that the design is in line with ethics and clinical norms. For example, an independent ethics monitoring team should be established to review experimental protocols, human-computer interfaces, privacy protection, etc. (2) Improve system controllability and transparency: As mentioned above, we provide a human monitoring interface and detailed explanations, which will enhance clinical trust in the system. At the same time, do a good job of informed communication with patients, explain the working principle and safety guarantee of the system in plain language, and let patients voluntarily participate and feedback their experience. (3) In the clinical trial stage, start from low-risk groups and adjuvant treatments. For example, it should be used first in severe patients (if other methods are ineffective), as an adjunct rather than the only means, to gradually accumulate safety data. When approved by regulatory agencies, detailed test reports and source code audits can be provided to prove that the system behavior is predictable and the risks are controllable. (4) Pay attention to the dynamics of laws and regulations: track domestic and foreign policies on AI medical and brain-computer interfaces, and lay out the necessary registration and certification processes in advance to ensure that the project results are promoted under the premise of compliance. Through these measures, we are confident that we can reduce the risk to a reasonable level and enable the innovative technology to move responsibly to the clinic.

In summary, we have made a comprehensive review of the potential risks of the project and formulated a thorough response plan. During the implementation of the project, the risk assessment will be reviewed regularly, and the countermeasures will be adjusted and refined according to the actual progress. Through proactive risk management, projects are completed on schedule and with high quality to the greatest extent.

Expected Outcomes and Appraisal Indicators

After the completion of this project, it is expected that a series of important achievements will be achieved in theoretical innovation, technological breakthrough and application demonstration. We will set assessment indicators from both qualitative and quantitative perspectives to ensure that the results are verifiable and measurable.

1. Theoretical and methodological achievements:

• Theoretical framework of semantic closed-loop brain-computer interface: A set of theoretical models of brain-computer-consciousness semantic dynamic interaction has been formed, and 1-2 high-level academic papers have been published to describe the application principles and effects of the DIKWP model in the brain-computer interface scenario. Assessment indicators: at least 1 paper published in SCI Zone 1 journal (or SCI Zone 1 journal of the Chinese Academy of Sciences in China), and he has been cited 20 >times; 1 paper at the International Summit Conference.

• DIKWP Semantic Mathematics and Brain Mapping Model: A mapping relationship model between the five layers of semantics and physiological signals in DIKWP is established to provide clear mathematical definitions and empirical support. Assessment indicators: submit no less than 1 national invention patent to protect the intellectual property rights of the semantic mapping model and decoding/encoding method; And in the field of conference reports, received positive feedback from peer experts.

2. Core technology and system:

• Artificial Consciousness ACPU Prototype System: Developed and completed the ACPU software system with DIKWP artificial awareness architecture, which can run in real time on the laboratory computing platform and realize closed-loop control with ultrasound equipment. Assessment index: ACPU decoding-decision-control delay< 100ms (meeting the requirements of physiological real-time); Semantic decoding accuracy >85% (for a defined set of typical brain states); Stimulus control accuracy (deviation in the time required to reach the target state) <± 20%. The system has passed internal testing, and the demonstration success rate is >90%.

• Closed-loop ultrasound brain-computer interface prototype device: Build a prototype device with software and hardware integration, including ultrasonic transmitting and receiving module, ACPU computing unit, monitoring interface, etc. Assessment indicators: The device successfully closed-loop controlled at least 3 different brain state scenarios in the simulation experiment, and the state deviation relative to the non-feedback control in each scenario was reduced by >=30% (e.g., the comparison of the reduction of anxiety index). The localization rate of key components of the device is >=80%, which lays the foundation for subsequent independent controllability.

3. Application Verification and Effectiveness:

• Efficacy in animal experiments: Validation of the effect of closed-loop interventions on rodent models of anxiety. Assessment indicators: Compared with the control group, the anxiety behavior indicators of the experimental group received closed-loop ultrasound intervention were reduced by >40% (such as the prolonged time of entering the open arm of the elevated cross maze, etc.), and the stress-related physiological indicators (heart rate, stress hormones) returned to normal levels faster. There was a statistically significant difference between groups (P<0.05). No animals had permanent neurological deficits or pathological damage, and the safety assessment was passed.

• Preliminary clinical effect: preclinical test results in a small number of subjects. Assessment measures: For example, patients with severe anxiety who participated in the trial had an average reduction of >30% in Beck Anxiety Scale (BAI) score compared with baseline after 2 weeks of intervention with this system (equivalent to an effect size of moderate or above, Cohen's d > 0.8); Patients with PTSD had a >25% reduction in PCL-5 (Post-Traumatic Stress Checklist) scores. At the same time, the patient's self-reported symptom relief was consistent with the improvement of objective physiological indicators. No serious discomfort or side effects were reported in all participants, with only a slight reversible discomfort <=1 patient. These data will indicate that the system has preliminary efficacy and safety feasibility in humans.

4. Support Tools & Standards:

• Draft Brain Semantic Interaction Protocol Standard: Compile and complete one technical report on the Brain Semantic Interaction Protocol (BSCP) Specification, covering the framework of the protocol, message format, semantic label definition, etc. (no less than 50 pages). Assessment indicators: The report will be reviewed by internal and external experts of the project, and submitted to the relevant standardization organizations or societies for project discussion, and strive to become the prototype of the industry standard.

• Open-source software and datasets: The proposed open-source simulation platform software, some non-sensitive algorithm code, and sample datasets generated during simulation and experiments (desensitized). Assessment indicators: publish at least 1 open source library on GitHub or domestic open source platform, and get 100+ stars; One dataset related to brain-computer semantic decoding was published, including no less than 1,000 sets of annotated samples, for academic research. This increases the impact of the project and advances the field.

5. Talent training and team building:

• Postgraduate training: Through the implementation of the project, 2-3 doctoral/master's students with interdisciplinary backgrounds will be trained to accumulate experience in the interdisciplinary fields of artificial intelligence, brain science and ultrasound engineering. Assessment indicators: students with project-related topics successfully graduated and published papers; At least one of them can independently undertake sub-project research and make a name for himself in the industry.

• Team influence: The project team has established a leading position in the domestic brain-computer interface and artificial intelligence academic circles. Assessment indicators: Invited to give a report at more than 1 important academic conference in China (such as BCI conference, artificial intelligence conference); The project results have won 1 provincial and ministerial science and technology awards or honors (such as natural science awards, technological invention awards); Project-related patents enter the substantive transformation process (e.g., signing a license or co-development intention with the enterprise).

The above indicators will be used as an important basis for project acceptance. Among them, the quantitative indicators strive to be clear, specific and measurable; Qualitative indicators are judged by expert review and material review. We will establish a data collection and validation methodology for each metric to ensure that the final project outcome stands up to scrutiny. For example, keep all the original experimental records for verification, establish a patient follow-up mechanism to evaluate the sustainability of efficacy, etc. By achieving the above results, the project will not only make a breakthrough in science, but also produce prototypes of technologies and products with practical application prospects, and promote the development of related industries.

Clinical demonstration application pathway

The ultimate goal of this project is to realize the transformation from laboratory results to clinical practice. We have planned a step-by-step clinical demonstration path to ensure that the new technology is safely and effectively applied to the target population and lay the foundation for large-scale rollout.

Phase 1: Simulated Clinical Scenario Validation (Year 3-4) – Build a near-real clinical scenario in the laboratory and test the system end-to-end. For example, simulating a psychiatric ward or psychotherapy room environment, having volunteers wearing ultrasound probes (healthy simulated patients) interact with clinicians, and the system monitors their mood changes in real time and gives appropriate interventions. Through this scenario exercise, the shortcomings of the system in the human-computer interaction process are found, such as patient cooperation, doctor operation interface friendliness, etc. Improve the design based on feedback in a timely manner to ensure that the system process is in line with clinical practice. For example, if a patient is found to be unable to sit still with ultrasound when anxious is intense, we may need to develop a portable wearable ultrasound device that allows the patient to move freely; If doctors want more manual control options, we can add manual mode capabilities, etc.

Phase 2: Small Sample Clinical Trial Study (Year 4-5) – 10-20 eligible patients with anxiety disorder or PTSD are selected for a single-arm initial clinical trial following ethical approval and regulatory filing. The location is available at the mental health center of the partner hospital. Each patient receives a multi-week system adjuvant therapy: conventional psychotherapy or pharmacotherapy plus our closed-loop ultrasound intervention. Detailed data were recorded, including brain signals, system decisions, stimulation parameters, as well as patients' subjective feelings and doctors' observation records during each session. The changes in the patients' symptom scale were evaluated before and after treatment and at follow-up, and the differences before and after the intervention were compared. If possible, introduce a control group (e.g., psychotherapy only) for comparison. The goal is to verify the safety of the system, patient tolerability, and trends in symptomatic improvement. If the results are positive (the rate of symptom improvement is significantly higher than that of the control), it can provide a basis for the next step of larger-scale validation; If the results are not satisfactory, you can also analyze the reasons such as the need to improve the algorithm or the need to limit the subtype for patient screening, etc., and try again after the system is upgraded.

Phase 3: Expansion of Strategic Collaboration and Multi-center Trial (Post-5 Phase) – After the formal phase of the project, we plan to move forward to Phase II/III clinical trials, which will require more units and larger samples. To this end, we will actively strive for follow-up scientific research or industrial transformation funds, and work with medical device companies, hospitals and regulatory agencies to carry out multi-center clinical trials. The psychiatric departments of 3-5 tertiary hospitals across the country were selected, led by experts who had participated in the project, and the improved and mature system was used in a controlled trial of more than 100 patients, and the efficacy and safety were strictly evaluated. If the trial is successful, we will cooperate with the application for special approval of innovative medical devices by the State Food and Drug Administration, and strive to make the product enter clinical use as soon as possible. At the same time, explore the application expansion: in addition to anxiety/PTSD, this system may be useful for patients with other brain diseases, such as depression (regulating emotional brain regions), addiction disorders (regulating reward circuits), etc., and can be piloted in a small area to summarize the efficacy. In terms of technology promotion, we have cooperated with local hospitals to establish demonstration centers, train clinical personnel to use the system, collect data on the effects of the system in different regions, and continuously optimize the system algorithm and use specifications.

Vision: If the above steps are successfully completed, we are expected to achieve the clinical transformation of this semantic closed-loop ultrasound brain-computer interface system within 5-8 years. At that time, patients with anxiety or PTSD may use this system as a routine adjunctive treatment in hospital or community care. For example, a typical scenario is for a patient to wear a specially designed ultrasound brain-computer interface headset connected to a portable device with a built-in ACPU. Under the guidance of a psychotherapist, if the patient has a strong stress response, the device will automatically detect and administer gentle ultrasound stimulation, and the patient's excessive tension will be quickly relieved, so as to better cooperate with the treatment. This immediate, individualized approach to intervention is expected to significantly improve treatment outcomes and patients' quality of life. From a broader perspective, the success of this system will prove the feasibility of this new paradigm of brain semantic interaction, provide valuable experience for the future development of a wider range of purpose-driven brain-computer interfaces (such as consciousness arousal in patients with impaired consciousness, motor purpose decoding in stroke patients, etc.), and promote the development of the integration of brain science and artificial intelligence.

Of course, in the process of moving towards clinical application, we will continue to pay attention to and abide by ethical and legal guidelines, respect the privacy and wishes of patients, and ensure that the technology is used wisely and not abused. Maintain communication with regulatory authorities and provide project data in a timely manner to facilitate the introduction of a reasonable regulatory framework to regulate the use of such AI brain-computer interface products. We believe that through a robust and far-sighted clinical transformation path, the results of this project will not only stay in papers and laboratories, but truly enter medical practice, bring good news to the majority of patients, and promote the innovation and upgrading of China's brain-computer interface industry and medical technology.

conclusion

This project is oriented to the cutting-edge research needs of high-spatiotemporal resolution ultrasound brain-computer interface, and takes the DIKWP artificial consciousness model proposed by Professor Yucong Duan as the theoretical cornerstone, and innovatively applies it to the closed-loop brain-computer interface system. Through the research protocol described in this paper, we will construct a closed-loop ultrasound brain-computer interface prototype system integrating the DIKWP artificial consciousness model to achieve strong semantic perception of brain state, purpose-oriented stimulus regulation, and the system's own autonomous evolutionary regulation ability, and finally form a self-explanatory, flexible and intelligent brain-computer interaction platform. This system will play an exemplary role in the intervention of neuropsychiatric diseases such as anxiety disorders and PTSD, provide patients with a new intelligent treatment pathway, and also open up a way to explore the integration of human semantics and machine intelligence.

Through the implementation of this project, it is expected to make major breakthroughs in the following aspects: firstly, the content of the intersection of artificial consciousness and brain-computer interface is theoretically enriched, and the semantic interaction paradigm of brain-computer-consciousness trinity is proposed, which expands the application boundary of the DIKWP model; Secondly, in terms of technology, it has developed an internationally leading prototype of semantic closed-loop brain-computer interface, overcome the key problems of high-precision decoding and intelligent control, form independent intellectual property rights and standards and specifications, and seize the commanding heights of the future industry. Finally, the application of the system verifies the intervention value of the system for typical mental illnesses, and brings new ideas for clinical treatment. It is foreseeable that the success of this project will enable China to take a leading position in the emerging frontier field of semantic brain-computer interaction, accelerate the paradigm shift from "signal interaction" to "semantic interaction", and lay a milestone foundation for the realization of a new generation of highly intelligent and human-oriented brain-computer interface.

In summary, this project focuses on the high spatiotemporal resolution ultrasound brain-computer interface, integrates the DIKWP model of artificial consciousness, and advances with the methodology of systems engineering. With a passion for scientific exploration and the benefit of patients, we have made sufficient theoretical preparations and technical reserves. Through close multidisciplinary cooperation, strict project management and risk control, we are confident that we will achieve the set goals on time and with high quality. The final result will not only be a prototype system and a series of paper patents, but also a powerful implementation of the great vision of "human-machine integrated intelligence". We believe that the development of this project will promote the deep integration of brain science, artificial intelligence and medical engineering, and its impact will continue to deepen, contributing to China's scientific and technological innovation and people's health.

References:

1. Duan, Y., et al. Research Report on the DIKWP Artificial Consciousness Model, 2023., et al.

2. Yucong Duan . DIKWP Artificial Consciousness Model and Its Application. Xinhuanet, 2025., etc.

3. Guo, Z., Duan, Y., et al. Neurophysiological basis and mechanism of reticulated DIKWP model, 2025., et al.

4. Barksdale, B., et al. “Low-intensity transcranial focused ultrasound… clinical trial.” Molecular Psychiatry, 2025.。

5. Kosnoff, J., et al. “Transcranial focused ultrasound to V5… attention.” Nat. Commun., 2024.。

6. Yoo, S., et al. “Decoding motor plans using… interface.” Nat. Neurosci., 2023.。

7. Zeng, X., et al. “Reduction of Anxiety-Related Symptoms… Ultrasound… Vagus Nerve.” JMIR Neurotech, 2025.。

8. Focused Ultrasound Foundation. “Anxiety – Focused Ultrasound could transform care.” 2024.。