Near-infrared brain imaging map feature-based emotional disorder recognition control model acquisition method and system, controller and control method

Abstract

An emotional disorder recognition control model acquisition method and system based on near-infrared brain imaging map characteristics, a controller and a control method, and the method comprises the following steps: step 100, fNIRS-based multi-dimensional nerve marker acquisition and analysis, step 200, an integrated handheld diagnostic device, step 300, a dual-mode high-speed intelligent analysis system, step 300, a near-infrared brain imaging map characteristic analysis system, and step 300, a near-infrared brain imaging map characteristic analysis system. And step 400, a closed-loop AI training and database iteration system realizes a link of collection-analysis-diagnosis-display by fusing technologies of signal collection, data preprocessing and normalization, characteristic parameter extraction, intelligent report display and the like, so that the transformation of mood disorder screening from subjectivity to objectivity and from experience driving to data driving is promoted, and the mood disorder screening efficiency is improved. Therefore, the embedded analysis system for the rehabilitation medical scene is provided.

Description

TECHNICAL FIELD

[0001] The present application relates to a mood disorder recognition control model acquisition method and system, a controller and a control method, in particular to a mood disorder recognition control model acquisition method and system, a controller and a control method based on near-infrared brain imaging atlas features. BACKGROUND

[0002] In the field of rehabilitation medicine, the physical and mental function recovery of patients not only depends on the progress of physical therapy, but also is closely related to the psychological and emotional state. Clinical studies have shown that patients recovering from neurological damage, such as stroke, spinal cord injury, and chronic pain, may develop mood disorders such as depression and anxiety. Such negative emotions can trigger abnormal release of inflammatory factors through activation of the hypothalamic-pituitary-adrenal axis (HPA axis), which may lead to delayed motor function recovery, decreased treatment compliance, and increased risk of recurrence.

[0003] However, the current rehabilitation efficacy evaluation system focuses too much on physiological indicators such as muscle strength and joint range of motion. The screening of mood disorders relies heavily on patient self-reporting scales such as PHQ-9 and GAD-7, which are easily influenced by social desirability bias and cognitive impairment. Developing a biological sensing technology that can objectively quantify the emotional state during rehabilitation and achieve dynamic detection has become a key scientific problem in improving the precision of rehabilitation treatment.

[0004] Functional near-infrared spectroscopy (fNIRS) is a non-invasive and non-radioactive brain function imaging technology that has received widespread attention. This technology monitors the dynamic changes in cerebral cortex blood oxygen concentration and can reflect the activity patterns of specific brain regions under emotional stimulation or task conditions. Previous studies have shown that the fNIRS atlas features of mood disorder patients in the prefrontal cortex and other brain regions are significantly different from those of normal people, indicating the potential for auxiliary recognition.

[0005] In addition, in the current medical scenario, fNIRS data interpretation relies heavily on manual analysis by professionals, with a single detection taking more than 30 minutes and only outputting basic indicators such as local brain region activation intensity, which cannot meet the needs of dynamic monitoring and immediate feedback in rehabilitation treatment.

[0006] The following technical defects exist:

[0007] I. Dependence on patient self-reporting scales

[0008] Current clinical diagnosis generally relies on self-reporting scales such as PHQ-9 and GAD-7, which are greatly influenced by patient subjective will and response bias. Patients may distort the results due to concealment of symptoms or cognitive impairment, reducing the objectivity and reliability of diagnosis.

[0009] II. High threshold for traditional neuroimaging devices

[0010] Although functional magnetic resonance imaging (fMRI) can provide high spatial resolution, it is expensive, has high operation and maintenance requirements, and is significantly limited by the use scenario. For patients with claustrophobia, children, and groups with poor compliance, the detection tolerance of fMRI is poor, and the clinical applicability is limited,

[0011] III. fNIRS data interpretation relies on manual work

[0012] Existing near-infrared brain imaging (fNIRS) mainly relies on professional personnel to manually complete data preprocessing and interpretation in clinical practice, and lacks automation and standardization. Most devices still need to rely on U disk and other media to transfer data, and the single detection process takes more than 30 minutes, and the output results are usually limited to basic indicators such as local brain region activation intensity, which makes it difficult to form a systematic diagnostic reference,

[0013] IV. Lack of standardized database and grassroots promotion

[0014] Whether it is fMRI or electroencephalogram (EEG), the existing method involves a complex operation process, which requires the intervention of professional engineers or senior medical technicians, making it difficult to popularize in grassroots medical institutions. In addition, there is currently a lack of a unified neurobiological marker database for emotional disorders, and the diagnostic threshold lacks evidence-based medical support, which affects its clinical transformation and promotion,

[0015] The present application develops a signal processing method to reconstruct the reflection pulse of the surface and the bottom, thereby predicting the density of the thin asphalt overlay layer during the compaction process, and proposes a surface moisture influence control method, which can realize real-time demonstration of the density change during the compaction process. SUMMARY

[0016] The object of the present application is a method for obtaining an emotional disorder recognition control model based on near-infrared brain imaging atlas features,

[0017] The object of the present application is a system for obtaining an emotional disorder recognition control model based on near-infrared brain imaging atlas features,

[0018] The object of the present application is an emotional disorder recognition controller based on near-infrared brain imaging atlas features,

[0019] The object of the present application is an emotional disorder recognition control method based on near-infrared brain imaging atlas features.

[0020] In order to overcome the above technical shortcomings, the object of the present application is to provide a method and system for obtaining an emotional disorder recognition control model based on near-infrared brain imaging atlas features, a controller and a control method, thereby realizing an embedded analysis system for rehabilitation medical scenarios.

[0021] To achieve the above object, the technical scheme adopted by the present application is:

[0022] A mood disorder recognition control model acquisition method based on near-infrared brain imaging atlas features, the steps of which are:

[0023] The steps are:

[0024] Step 100: fNIRS-based multi-dimensional neural marker acquisition and analysis,

[0025] Step 200: integrated handheld diagnostic device,

[0026] Step 300: dual-mode high-speed intelligent analysis system,

[0027] Step 400: closed-loop AI training and database iteration system.

[0028] Due to the design of the above steps, by integrating signal acquisition, data preprocessing and normalization, feature parameter extraction, and intelligent report display technologies, the link of "acquisition-analysis-diagnosis-display" is realized, which promotes the transformation of mood disorder screening from subjective to objective, from experience-driven to data-driven, to meet the needs of multi-scenario rehabilitation medicine, and thus an embedded analysis system for rehabilitation medical scenarios is provided.

[0029] The present application designs a running state equation group in CPU, which includes the following contents:

[0030] Step 100 specifically includes the following contents:

[0031] A near-infrared light source array is used to cover the core brain regions of emotion regulation, including the prefrontal lobe, temporal lobe, frontal lobe, central sulcus, parietal lobe, and occipital lobe,

[0032] The system monitors the dynamic changes of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) concentrations in real time and outputs raw light intensity data,

[0033] The concentration change curve is obtained by Beer-Lambert Law (MBLL) conversion,

[0034] The feature extraction module analyzes the neural markers in the following dimensions:

[0035] First, spatial dimension: calculate the functional connection strength between brain regions, such as the correlation weight of prefrontal lobe-edge system,

[0036] Second, time dimension: extract nonlinear features from the hemoglobin concentration time-varying curve, including Hurst index and sample entropy,

[0037] Third, frequency domain dimension: power spectrum analysis of low-frequency oscillation (0.01-0.1 Hz) to extract frequency energy distribution characteristics,

[0038] The processed multi-dimensional features are input to a subsequent diagnosis module for mood disorder identification.

[0039] Step 200 specifically includes the following:

[0040] It includes an fNIRS signal acquisition module, an embedded processor, a display screen, and a power management module,

[0041] After the patient wears the headgear, the device automatically completes signal acquisition, and the embedded processor calculates twelve features, i.e., functional connectivity network density, oxygen metabolism time-varying nonlinear index, locally, and inputs the results into the diagnosis model,

[0042] The model outputs mood disorder risk level and subtype probability, and then the report generation module presents the visual results on the display screen, including risk level prompt, abnormal brain region heat map, and subtype suggestion.

[0043] Step 300 specifically includes the following:

[0044] A dual-mode data transmission architecture is used to support efficient data acquisition and processing:

[0045] First, wired mode: a 5Gbps channel is established through a USB-C interface, with an end-to-end delay of less than 1ms, suitable for ICU or operating room environments with strong electromagnetic interference,

[0046] Second, wireless mode: integrates a 2.4GHz custom protocol based on IEEE 802.11ax, supports dynamic bandwidth allocation and AES-256 encryption, with a peak rate of up to 800Mbps, suitable for mobile deployment scenarios such as community follow-up and pediatric ward,

[0047] The system has an automatic triggering mechanism: when the terminal and the fNIRS device establish a wired connection or successfully pair wirelessly, the data flow pipeline is automatically activated for real-time transmission, eliminating the need for manual copying steps,

[0048] The embedded processor synchronously calculates spatial dimension, i.e., prefrontal-amygdala functional connectivity, temporal dimension (HbO / HbR sample entropy), and frequency domain dimension (θ frequency band power ratio) indicators during the acquisition process, and generates a diagnosis report within 2 minutes, including risk level, abnormal brain region heat map, and subtype probability.

[0049] Step 400 specifically includes the following:

[0050] It consists of three parts: database construction, model training, and edge deployment,

[0051] First, standardized database construction: the system automatically collects detection data from the handheld terminal, that is, twelve types of features and clinical diagnosis labels, and completes the grading annotation of depression and anxiety according to the DSM-5 standard, and stores them into the mood disorder brain function biomarker library,

[0052] Second, dynamic AI optimization mechanism: the model training module constructs an emotional disorder classifier based on gradient boosting tree (GBDT), the input is 12-dimensional neural features, and the output is the probability distribution of disorder subtypes, for newly labeled data, the system adopts an incremental learning strategy to add it to the training set and update the model, thereby improving the generalization ability and subtype discrimination accuracy of the model,

[0053] Basic empowerment strategy: the updated model is quantized and pruned to reduce the weight of the terminal, so that the terminal can call the latest model in an offline state, and the user only needs to perform a single key operation to complete detection and diagnosis, which is suitable for community hospitals and basic medical scenarios, and the data uploading, model training and distribution form a closed loop to ensure continuous iteration of the model and stable improvement of the diagnosis performance.

[0054] The present application designs, first, an fNIRS and EEG joint acquisition scheme:

[0055] In addition to using fNIRS signals alone, a multi-modal alternative scheme is provided: EEG electrodes are integrated into the headgear at the same time, arranged according to the international 10-20 system, and brain electric potential signals are collected,

[0056] After the EEG data is wave-trapped and de-noised by independent component analysis, it is multi-modally fused with the hemoglobin dynamic features extracted by fNIRS,

[0057] The fusion method can include: feature-level concatenation, decision-level weighting (ensemble), or joint modeling based on graph convolution network (GCN),

[0058] This alternative scheme can enhance the classification robustness when the fNIRS signal-to-noise ratio is low,

[0059] Second, different feature extraction methods:

[0060] In addition to Hirst index, sample entropy and other nonlinear features, the following alternative indicators can also be extracted:

[0061] 1. Time domain: based on detrended fluctuation analysis (DFA), Lyapunov index,

[0062] 2. Frequency domain: wavelet packet energy distribution, empirical mode decomposition (EMD) features,

[0063] 3. Spatial domain: graph theory indicators such as degree centrality, characteristic path length, clustering coefficient,

[0064] These indicators can replace or supplement the original twelve categories of features to adapt to different populations or clinical scenarios.

[0065] Third, handheld terminal replacement form:

[0066] In addition to integrated handheld devices, the system can also use modular architecture:

[0067] The acquisition unit, i.e. the headgear + fNIRS module, is directly connected to the analysis unit, i.e. the smartphone or tablet, through Bluetooth Low Energy or Wi-Fi,

[0068] The acquisition module only completes the raw signal acquisition, and the feature extraction and diagnosis model running are undertaken by the external mobile terminal,

[0069] This scheme can reduce the weight of the device and is suitable for follow-up scenarios for children and the elderly,

[0070] Fourth, different data transmission and encryption mechanisms:

[0071] In addition to USB-C and 2.4GHz wireless methods, alternative solutions can also use:

[0072] 1. Fiber interface (SFP module) to achieve higher bandwidth transmission,

[0073] 2. 5G mobile communication module for remote real-time monitoring,

[0074] 3. End-to-end encryption method can use SM4 or TLS1.3 instead of AES-256 to adapt to different regional regulatory requirements,

[0075] Fifth, alternative AI training methods:

[0076] In addition to gradient boosting tree (GBDT) classifiers, the following methods can also be used:

[0077] 1. Support vector machine (SVM) or random forest (RF) for multi-class mood disorder classification,

[0078] 2. Time series modeling method based on Transformer (such as temporal attention network) for extracting long-term dynamic features,

[0079] 3. Semi-supervised or self-supervised learning framework to reduce the need for manually labeled data,

[0080] These methods can be selected according to the size of the data and the computing resources,

[0081] Sixth, model deployment and update alternatives:

[0082] In addition to edge OTA delivery, models can also be updated through the following methods:

[0083] 1. Deploy the model in the cloud, and the handheld terminal can call the cloud inference service in real time.

[0084] 2. By embedding an FPGA or NPU module in the terminal, more complex deep neural network models can be run.

[0085] Using a federated learning mechanism, each terminal only uploads the model gradients to the server after training locally, ensuring data privacy.

[0086] This invention designs a progressive architecture of data collection, analysis, and diagnosis to achieve objective detection of mood disorders. Figure 3 This is a system block diagram showing the collaborative workflow of each module, where: A is the near-infrared data acquisition device, B is the handheld terminal (including B1 signal preprocessing, B2 feature analysis, B3 emotion diagnosis, B4 report generation, and B5 display screen), and C is the AI ​​training service (including C1 training data storage, C2 model training, and C3 new model output).

[0087] The steps are as follows:

[0088] I. Diagnostic Process for Mood Disorders: Figure 4 This is a flowchart of the diagnostic process for mood disorders.

[0089] First, data origin and synchronization

[0090] 1. fNIRS raw input: dual-wavelength light intensities I(λ1,t), I(λ2,t), each source-detector pair forming one channel; sampling rate 10–50Hz.

[0091] 2. EEG raw input: 21-channel electrode potential V based on the international 10–20 system e (t), the reference electrode is set to average reference or dual-mammary gland reference; sampling rate 250–1000Hz (default 500Hz),

[0092] 3. Clock synchronization: A (data acquisition unit) and B (handheld terminal) are calibrated via the NTP protocol; frame-level timestamp alignment error ≤ 1ms.

[0093] Second, fNIRS preprocessing (B1)

[0094] Optical-to-Concentration Conversion (MBLL)

[0095]

[0096] 1. Parameters: E is the extinction coefficient matrix; L = 30–35 mm; DPF = 5.5–6.5 (adaptively adjusted according to the subject's age and head circumference).

[0097] 2. Filtering: FIR bandpass filtering (0.01–0.2Hz) is used to preserve physiologically relevant frequencies, while high-pass filtering (cutoff frequency 0.01Hz) is used to remove baseline drift.

[0098] 3. Motion artifacts: Artifacts are identified using a dual-threshold detection method combining amplitude thresholding and first-order difference, and corrected using short-time-window spline interpolation.

[0099] 4. Short probe distance regression: Using a short probe distance channel as a regressor, superficial tissue signal interference is eliminated in the generalized linear model (GLM).

[0100] 5. Channel Quality Scoring: Taking into account signal-to-noise ratio (SNR), coefficient of variation (CV), and inter-channel correlation consistency, channels below a certain threshold are marked as low-weighted.

[0101] Third, EEG preprocessing (B1)

[0102] 1. Re-reference: Use average reference or dual mastoid reference; use notch filtering (50 / 60Hz) to remove power line interference, and perform bandpass filtering (1–45Hz) to retain effective EEG frequency bands.

[0103] 2. ICA Artifact Removal: Independent Component Analysis (ICA) is applied to separate and remove noise components from electrooculography (EOG, such as blinking), electromyography (EMG), electrocardiography (ECG), and other channels.

[0104] 3. Trial cleaning: Remove bad trials with severe noise, retain good trials; perform baseline correction (using the first 10 seconds at rest as the baseline interval).

[0105] Fourth, windowing and baseline

[0106] 1. Online sliding window: A sliding window (W) with a length of 20–30 seconds and a step size of 2–5 seconds (S) is used; simultaneously, a circular buffer for maintaining individual baselines is maintained for a duration of no less than 120 seconds.

[0107] 2. Paradigm Task: The data acquisition process consists of "baseline segment + stimulation segment + recovery segment"; if there is no paradigm task, only sliding window processing is performed.

[0108] Fifth, Feature Analysis (B2)

[0109] 5.1 fNIRS features

[0110] 1. Spatial dimension: including the mean ROI (Region of Interest), the left and right hemisphere lateralization indices (LI), and the connection weights between the prefrontal cortex and the limbic system.

[0111] 2. Time dimension: This includes indicators such as Hurst exponent, sample entropy, multi-scale entropy, detrended volatility analysis (DFA) slope, peak latency, and half-peak width.

[0112] 3. Frequency domain dimension: including indicators such as the energy proportion of the 0.01–0.1Hz frequency band and the peak frequency of the power spectrum.

[0113] 4. Connectivity dimension: including correlation coefficient, partial correlation coefficient, coherence, phase-locked value (PLV) matrix, and graph theory-based metrics (degree, clustering coefficient, efficiency, betweenness centrality, and modularity).

[0114] 5.2 EEG Features

[0115] 1. Power Spectrum Characteristics: Calculates the power spectral density (PSD) for each frequency band, including delta wave (0.5–4Hz), theta wave (4–8Hz), alpha wave (8–13Hz), beta wave (13–30Hz), and gamma wave (30–45Hz). It can also further calculate the normalized ratio of the power in each frequency band.

[0116] 2. Emotion-related indicators: A typical indicator is frontal alpha asymmetry (FAA), which is calculated by subtracting the logarithm of the left frontal alpha power from the logarithm of the right frontal alpha power.

[0117] 3. Connectivity characteristics: including weighted phase lag index (wPLI), coherence and partial coherence indices; effective connectivity indices (such as Granger causality or dynamic transfer function DTF) can also be extracted when needed.

[0118] 4. Event-related characteristics: When a task paradigm exists, the peak values ​​(such as P2 and N2) and their latencies of event-related potential (ERP) components can be extracted.

[0119] 5.3 Quality and Confidence

[0120] 1. Channel and Electrode Quality Weights: A quality weight α_i∈[0,1] is assigned to each channel or electrode, and a weighted average method is used during feature aggregation.

[0121] 2. Window-level confidence: At the sliding window level, considering factors such as effective channel rate, signal-to-noise ratio (SNR), and artifact ratio, the confidence index q of the window is calculated.

[0122] Sixth, Emotion Diagnosis (B3)

[0123] 1. Input features: fNIRS top map (64×64×2), fNIRS connection matrix M_n, EEG connection matrix M_e, and statistical feature vector F.

[0124] 2. Model Structure: A three-branch fusion structure is adopted, including CNN (processing the top image), GCN (processing M_n, M_e), and GBDT or MLP (processing F). Attention weights ω_n, ω_e, and ω_s are introduced into the fusion layer, and the constraint ω_n + ω_e + ω_s = 1 is applied.

[0125] 3. Dynamic weighting mechanism: The attention weight ω is adaptively allocated based on quality metrics (confidence q, effective channel rate) and historical performance.

[0126] 6.1 Probability and Uncertainty

[0127]

[0128] 6.2 Smoothing and Alarm Logic

[0129] R t =α·max(p_t)+(1-α)·R_{t-1}, 0<α<1

[0130] 1. Alarm Trigger Mechanism: When the risk score R_t is greater than or equal to the threshold θ and the fluctuation index σ is less than or equal to the threshold τ for K consecutive time windows, the system determines the state as "high risk" and triggers an alarm. Typical parameters are set to K=3, θ=0.8, and τ=0.2. These parameters can be modified according to the actual application scenario.

[0131] 2. Subtype discrimination function: The system further outputs probability vectors for mood disorder subtypes, covering categories such as depression, anxiety, bipolar, and neutral, to assist in refined classification and subsequent intervention.

[0132] Seventh, Report Generation and Readability (B4 / B5)

[0133] 1. Report Fields: Output results include risk level, subtype probability, confidence interval, heatmap of abnormal brain regions based on fNIRS, and EEG / fNIRS connectivity plot.

[0134] 2. Interpretability: The model possesses interpretability analysis capabilities, including Grad-CAM-based top-graph visualization, channel or electrode importance ranking, and key edge contributions identified through the graph convolutional network (GCN) attention mechanism.

[0135] 3. Edge-side latency: During the edge-side inference phase, the inference latency for a single time window is less than 300 milliseconds; the generation and rendering latency of the complete report is less than 2 minutes.

[0136] 8. End-side data flow table (input → processing → output → delay)

[0137]

[0138] II. AI Training Service and Database Iteration System (Cloud-based C1–C3): Figure 5 System flowchart for AI training services and database iteration system.

[0139] Second, data standardization and storage (C1)

[0140] 1. Data Unit: The stored data unit includes the original summary ID, feature vectors F_n and F_e, connection matrices M_n and M_e, top graph hash value, quality metric (q_n, q_e), device parameters, task paradigm, individual baseline, label y, and timestamp.

[0141] 2. Anonymization: Direct identity information is removed, and the subject's ID is hashed; data transmission is encrypted using the TLS 1.3 protocol, and storage is encrypted using the AES-256 algorithm.

[0142] 3. Quality Control: Perform data integrity checks (field missing rate), physiological rationality checks (numerical range, drift trend, heart rate coupling), noise ratio analysis, and channel effectiveness evaluation.

[0143] 4. Version Management: Assign DataVersion and SchemaVersion identifiers to each uploaded data to ensure traceability and version consistency.

[0144] Second, training data construction and stratified sampling

[0145] 1. Label Alignment: Based on the DSM-5 criteria, the samples were labeled with the type of mood disorder (e.g., depression, anxiety, etc.), and the scale score ranges (e.g., PHQ-9, GAD-7 scale score ranges) were recorded.

[0146] 2. Layered strategy: Modifications are made according to data collection center, gender, age, device type, and multiple strategies. Third, model training (C2).

[0147] 1. Multi-branch deep model: Employs a multi-branch fusion architecture, including Convolutional Neural Networks (CNN, for processing top graphs), Graph Convolutional Networks (GCN, for processing connection matrices M_n and M_e), and Multilayer Perceptrons (MLP, for processing statistical features). This model structure maintains homology with the edge-side inference structure to facilitate knowledge distillation. 2. Base learner: Uses Gradient Boosting Decision Trees (GBDT) for lightweight classification of 12 core neural features, used for edge deployment and baseline comparison.

[0148] 3. Hyperparameter search: Hyperparameter search is performed using Bayesian optimization methods or the Optuna framework, and model evaluation is conducted by combining cross-validation and time buffer strategies.

[0149] 4. Model Calibration: The model output is calibrated using temperature scaling and Platt scaling methods to generate more reliable confidence estimates, and reliability metrics such as expected calibration error (ECE) and maximum calibration error (MCE) are provided. 5. Fairness and Robustness Assessment: ROC-AUC and ECE metrics are evaluated according to population subgroups, and robustness tests are conducted against noise disturbances and missing channels.

[0150] Fourth, incremental learning and active learning

[0151] 1. Incremental Retraining: After new samples are manually verified and added to the database, the system triggers a retraining process based on time (weekly) and sample size thresholds. To avoid the forgetting effect, Elastic Weight Consolidation (EWC) and a replay buffer mechanism are used to enhance stability.

[0152] 2. Active Learning: Based on the uncertainty index σ of the prediction results and the distribution drift measures (PSI and KL divergence), high-value samples are dynamically selected to prioritize entering the labeling process, thereby improving model training efficiency.

[0153] 3. Semi-supervised learning: Introducing consistency regularization and pseudo-label methods on unlabeled data improves the model's representation learning ability and generalization performance.

[0154] Fifth, model registration, deployment, and rollback (C3→B3)

[0155] 1. Stratified sampling is performed based on task type, with the dataset divided into three parts: 70% training set, 15% validation set, and 15% test set. 2. Class imbalance handling: To address the class imbalance problem, methods including focal loss, cost-sensitive learning, and resampling techniques (SMOTE oversampling and stratified undersampling) are employed.

[0156] 3. Model Registration: The system registers and manages each model, recording information including the model version number (ModelVersion), the corresponding training data version, key performance metrics, and interpretability outputs (Grad-CAM heatmap and feature importance ranking results).

[0157] 4. Gradual Deployment: A gradual deployment strategy will be adopted, with the deployment percentage increasing from 5% to 25% and then to 100%. During this process, A / B testing or control experiments will be conducted, and online performance metrics, including AUC, ECE, and alarm rate, will be continuously monitored.

[0158] 5. Compressed Deployment: During the model deployment phase, model compression optimization is supported, including quantization (INT8), pruning, and knowledge distillation. The new version automatically replaces the old model after passing the on-device self-check, while the system retains a rollback channel to ensure availability and security.

[0159] Sixth, data and model drift monitoring

[0160] 1. Data Drift Monitoring: Continuous monitoring of channel and electrode statistics triggers an alert when the Population Stability Index (PSI) exceeds 0.2. Simultaneously, structural changes in the connectivity graph are monitored, using spectral radius and modularity as evaluation metrics.

[0161] 2. Concept Drift Monitoring: When online model performance metrics show a downward trend (e.g., AUC decreases or ECE increases), the system triggers a retraining process and audits relevant features to ensure model performance stability. 3. Log and Compliance Audit: All events throughout the system's operation are logged to ensure that data and model processing meet compliance and traceability requirements.

[0162] VII. Database and Field Examples

[0163]

[0164] This solution ensures that the processing chain from raw data to diagnostic results remains continuous, traceable, and rollbackable, and continuously improves the robustness and interpretability of the model in real-world application scenarios through a closed-loop learning mechanism.

[0165] This invention designs a method for identifying mood disorders based on near-infrared brain imaging (fNIRS) technology. During the detection process, the subject wears a headgear containing a near-infrared light source and a detector array. The detector array covers brain regions related to mood regulation, such as the prefrontal cortex, dorsolateral prefrontal cortex, frontal lobe, and temporal lobe. The NIRS device simultaneously acquires the concentration change curves of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR), outputting analog signals. These signals are converted into digital signals by a near-infrared data acquisition unit and transmitted to a handheld terminal. The signal preprocessing module of the handheld terminal performs motion artifact removal, baseline correction, multivariate Bell's Law (MBLL) transformation, and filtering. Subsequently, the feature extraction module calculates functional connectivity strength (prefrontal-amygdala correlation) in the spatial dimension, calculates nonlinear indices of concentration dynamics (sample entropy, Hearst exponent) in the temporal dimension, and extracts low-frequency oscillation power spectrum (0.01–0.1 Hz) in the frequency domain. These multidimensional features constitute a neural marker matrix for subsequent mood disorder analysis.

[0166] This invention designs a handheld mood disorder identification device, comprising: an fNIRS signal acquisition module, an embedded processor, an interactive display screen, and a power supply module. When the patient wears the headgear, the device automatically initiates signal acquisition. The embedded processor performs multi-dimensional feature calculations locally, including 12 parameters such as functional connectivity network density and oxygen metabolism time-varying index. The mood diagnosis module classifies these features based on a pre-trained model, outputting the mood disorder risk level and possible subtypes. The diagnostic results are converted into graphic and textual formats by a visualization report generation module, including: risk level indication, abnormal brain region heatmap, and feature curve display. Users can view the report in real time on the display screen; the entire process requires no additional computer or specialized software support.

[0167] This invention designs a high-speed data processing system. The handheld terminal is equipped with a USB-C interface and a 2.4GHz wireless module, supporting both wired and wireless transmission. In wired mode, the USB-C establishes a 5Gbps channel with a data transmission latency of less than 1ms, suitable for high-interference environments such as ICUs. In wireless mode, a customized protocol based on IEEE 802.11ax is used, supporting AES-256 encryption, with a peak rate of up to 800Mbps. Upon arrival at the handheld terminal, the system automatically triggers a real-time analysis process, eliminating the need for manual copying. The embedded processor performs dynamic HbO / HbR analysis, functional connectivity calculation, and time-frequency index extraction, generating a diagnostic report within approximately 2 minutes. The report includes: risk level (low, medium, high), abnormal brain region localization (abnormal activation in the left prefrontal cortex), probability of mood disorder subtype (depression / anxiety, etc.), and outputs corresponding heatmaps and connectivity strength matrices.

[0168] This invention designs a closed-loop AI training system, comprising a data acquisition terminal, a cloud database, and a model training service. The data acquisition terminal is the aforementioned handheld terminal, which collects multidimensional features (including 12 categories such as hemoglobin dynamic curves and functional connectivity indicators) along with clinical diagnostic labels and uploads them to the cloud database. The database is uniformly structured and stored according to the DSM-5 standard, constructing a brain functional biomarker library for mood disorders. The model training service performs incremental learning based on the database: using a gradient boosting tree (GBDT) and a deep neural network classifier, it trains the input features and outputs the probability distribution of disorder subtypes. When new data is added to the database, the system automatically triggers retraining, updates the classification model, and records version information. The compressed and quantized new model is delivered to the handheld terminal via OTA, allowing it to access the latest algorithm even offline. During clinical use, diagnostic reports and user feedback are returned to the database as training data for the next round, achieving closed-loop updates and ensuring continuous model optimization.

[0169] This invention designs a system for acquiring an emotion disorder identification and control model based on near-infrared brain imaging atlas features, comprising the following:

[0170] Unit 10 was established based on the multidimensional neural marker acquisition and analysis using fNIRS.

[0171] Unit 20 is established based on an integrated handheld diagnostic device, unit 30 is established based on a dual-mode high-speed intelligent analysis system, and unit 40 is established based on a closed-loop AI training and database iteration system.

[0172] This invention designs an emotion disorder recognition controller based on near-infrared brain imaging atlas features, comprising the following: a control model for emotion disorder recognition based on near-infrared brain imaging atlas features is stored in the controller, and the control model for emotion disorder recognition based on near-infrared brain imaging atlas features is obtained according to the above-mentioned method for obtaining the control model for emotion disorder recognition based on near-infrared brain imaging atlas features.

[0173] A method for identifying and controlling mood disorders based on near-infrared brain imaging atlas features includes the following: applying a mood disorder identification controller based on near-infrared brain imaging atlas features in a CPU for control.

[0174] The technical advantages of this invention are as follows:

[0175] First, objectivity is enhanced.

[0176] This invention directly reflects the state of brain function by real-time acquisition of fNIRS signals (dynamic changes in HbO / HbR concentration) and multi-dimensional feature extraction, reducing bias caused by subjective patient responses (such as PHQ-9 and GAD-7), thereby improving the objectivity and reliability of diagnosis.

[0177] Second, integration and portability

[0178] The handheld diagnostic device proposed in this invention integrates an fNIRS analysis module, an embedded processor, and a display screen into one unit. Its overall size and weight are significantly lower than traditional fMRI and other imaging devices. The detection process requires no additional computer or specialized software support, making it convenient for use in outpatient clinics, community hospitals, and mobile settings.

[0179] Third, real-time performance and high efficiency

[0180] The device supports dual-mode data transmission via USB-C wired (5Gbps) and 2.4GHz wireless (IEEE 802.11ax, AES-256 encryption), establishing an automatically triggered data path and eliminating the need for manual copying. A single detection process can complete feature extraction and report generation within 2 minutes, significantly shortening the time compared to the traditional process relying on manual interpretation (over 30 minutes). Fourth, the multi-dimensional indicator system of this invention simultaneously analyzes neural features in three dimensions: spatial, temporal, and frequency domains, including 12 types of features such as functional connectivity strength, sample entropy, Hearst exponent, and frequency band power distribution. This is more systematic and comprehensive than existing fNIRS methods that only output local activation intensity, providing abnormal brain region heatmaps and subtype identification results.

[0181] Fifth, closed-loop AI optimization mechanism

[0182] This invention constructs a multi-center standardized database, combined with DSM-5 labels, to continuously accumulate biomarker data on mood disorders. Through incremental learning and iterative model updates, the diagnostic threshold and classification accuracy improve as the data scale expands, ensuring that the diagnostic system has an evidence-based medical foundation.

[0183] Sixth, ease of use and grassroots promotion

[0184] The lightweighted model can be directly deployed to handheld terminals, which can access the latest algorithms even offline. Users can obtain diagnostic results with just a single click, thus lowering the barrier to entry and making it suitable for promotion in primary healthcare settings and resource-limited locations.

[0185] Seventh, scalability and compatibility

[0186] The architecture of this invention supports multimodal extensions (such as fNIRS+EEG fusion), alternative feature extraction methods (wavelet energy, graph theory metrics), alternative AI classifiers (SVM, Transformer, etc.), and different deployment methods (cloud inference, federated learning), and has strong technical adaptability and scalability. Attached Figure Description

[0187] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0188] Figure 1 This is a flowchart illustrating a method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features, according to the present invention.

[0189] Figure 2This is a schematic diagram of the structure of an emotion disorder recognition controller based on near-infrared brain imaging atlas features according to the present invention.

[0190] Figure 3 A system block diagram showing the collaborative workflow of each module.

[0191] Figure 4 This is a flowchart of the diagnostic process for mood disorders.

[0192] Figure 5 System flowchart for AI training services and database iteration system. Detailed Implementation

[0193] According to the examination guidelines, terms such as “having,” “comprising,” and “including” used in this invention should be understood to mean without dispensing the presence or addition of one or more other elements or combinations thereof.

[0194] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0195] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0196] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0197] The present invention will be further described below with reference to embodiments. These embodiments are intended to illustrate the present invention and not to further limit the present invention.

[0198] A method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features, the first embodiment of which comprises the following steps:

[0199] Step 100: Multidimensional neural marker acquisition and analysis based on fNIRS

[0200] Step 200: Integrated handheld diagnostic device,

[0201] Step 300: Dual-mode high-speed intelligent analysis system.

[0202] Step 400: Closed-loop AI training and database iteration system.

[0203] In this embodiment, step 100 specifically includes the following:

[0204] A near-infrared light source array was used to cover the core brain regions for emotion regulation, including the prefrontal cortex, temporal lobe, frontal lobe, central sulcus, parietal lobe, and occipital lobe.

[0205] The system monitors the dynamic changes in the concentrations of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) in real time and outputs raw light intensity data.

[0206] The concentration change curve was obtained through Bell-Lambert Law (MBLL) conversion.

[0207] The feature extraction module analyzes neural markers according to the following dimensions:

[0208] First, the spatial dimension: calculating the strength of functional connectivity between brain regions, such as the relevant weights of the prefrontal-limbic system.

[0209] Second, the time dimension: nonlinear features, including the Hurst exponent and sample entropy, are extracted from the time-varying curve of hemoglobin concentration.

[0210] Third, frequency domain dimension: power spectrum analysis is performed on low-frequency oscillations (0.01–0.1Hz) to extract frequency domain energy distribution characteristics.

[0211] The processed multidimensional features are input into the subsequent diagnostic module for the identification of mood disorders.

[0212] In this embodiment, step 200 specifically includes the following:

[0213] Includes: fNIRS signal acquisition module, embedded processor, display screen, and power management module.

[0214] After the patient puts on the headgear, the device automatically completes signal acquisition. The embedded processor calculates twelve features locally, namely functional connectivity network density and oxygen metabolism time-varying nonlinear exponent, and inputs the results into the diagnostic model.

[0215] The model outputs the risk level and subtype probability of mood disorders, and then the report generation module presents the visualization results on the screen, including: risk level prompts, abnormal brain region heat maps and subtype suggestions.

[0216] In this embodiment, step 300 specifically includes the following:

[0217] A dual-mode data transmission architecture to support efficient data acquisition and processing:

[0218] First, wired mode: A 5Gbps channel is established via the USB-C interface, with an end-to-end latency of less than 1ms, making it suitable for ICU or operating room environments with strong electromagnetic interference.

[0219] Second, wireless mode: It integrates a customized 2.4GHz protocol based on IEEE 802.11ax, supports dynamic bandwidth allocation and AES-256 encryption, and has a peak rate of up to 800Mbps, making it suitable for mobile deployment scenarios such as community follow-up and pediatric wards.

[0220] The system features an automatic triggering mechanism: when the terminal successfully establishes a wired connection or wireless pairing with the fNIRS device, the data stream pipeline is automatically activated, enabling real-time transmission and eliminating the need for manual copying.

[0221] During the acquisition process, the embedded processor simultaneously calculates spatial dimension indicators, namely prefrontal-amygdala functional connectivity, temporal dimension (HbO / HbR sample entropy), and frequency domain dimension (θ band power ratio), and generates a diagnostic report within 2 minutes, which includes risk level, abnormal brain region heat map, and subtype probability.

[0222] In this embodiment, step 400 specifically includes the following:

[0223] It consists of three parts: database construction, model training, and edge deployment.

[0224] First, standardized database construction: The system automatically collects detection data from handheld terminals, including twelve categories of features and clinical diagnostic labels, and completes the classification and labeling of depression and anxiety according to the DSM-5 standard, storing it in the brain functional biomarker library for mood disorders.

[0225] Second, the dynamic AI optimization mechanism: The model training module constructs an emotion disorder classifier based on Gradient Boosting Tree (GBDT). The input is 12-dimensional neural features, and the output is the probability distribution of disorder subtypes. For newly added labeled data, the system adopts an incremental learning strategy to add it to the training set and update the model, thereby improving the model's generalization ability and subtype discrimination accuracy.

[0226] Grassroots empowerment strategy: The updated model is distributed to handheld terminals after lightweight processing such as parameter quantization and pruning, so that the terminals can call the latest model even when offline. Users only need to perform a single-click operation to complete the detection and diagnosis. It is suitable for community hospitals and primary medical scenarios. Data upload, model training and distribution form a closed loop to ensure continuous model iteration and stable improvement of diagnostic performance.

[0227] In this embodiment, the first method is a combined fNIRS and EEG acquisition scheme:

[0228] In addition to using fNIRS signals alone, a multimodal alternative is provided: EEG electrodes are integrated into the headgear and arranged according to the international 10-20 system to collect EEG potential signals.

[0229] After notch filtering and artifact removal by independent component analysis, EEG data were fused with hemoglobin dynamic features extracted by fNIRS for multimodal analysis.

[0230] Fusion methods may include: feature-level concatenation, decision-level ensemble, or joint modeling based on graph convolutional networks (GCNs).

[0231] This alternative approach can enhance classification robustness even when the fNIRS signal-to-noise ratio is low.

[0232] Second, different feature extraction methods:

[0233] In addition to nonlinear features such as the Hearst exponent and sample entropy, the following alternative indicators can also be extracted:

[0234] 1. Time domain: Based on detrended volatility analysis (DFA) and the Lyapunov index.

[0235] 2. Frequency Domain: Wavelet packet energy distribution, Empirical Mode Decomposition (EMD) characteristics.

[0236] 3. Spatial domain: Graph theory metrics such as degree centrality, characteristic path length, and clustering coefficient.

[0237] These indicators can replace or supplement the original twelve categories of characteristics to adapt to different populations or clinical scenarios.

[0238] Third, alternative forms of handheld terminals:

[0239] In addition to integrated handheld devices, this system can also adopt a modular architecture:

[0240] The data acquisition unit (headgear + fNIRS module) and the analysis unit (smartphone or tablet) are directly connected via Bluetooth Low Energy or Wi-Fi.

[0241] The acquisition module only completes the raw signal acquisition; feature extraction and diagnostic model operation are handled by an external mobile terminal.

[0242] This solution reduces equipment weight, making it suitable for follow-up scenarios involving children and the elderly.

[0243] Fourth, different data transmission and encryption mechanisms:

[0244] In addition to USB-C and 2.4GHz wireless, alternative solutions include:

[0245] 1. Fiber optic interface (SFP module) enables higher bandwidth transmission.

[0246] 2. 5G mobile communication module, used for remote real-time monitoring.

[0247] 3. End-to-end encryption can use the Chinese national standard SM4 or TLS 1.3 instead of AES-256 to adapt to the regulatory requirements of different regions.

[0248] Fifth, alternative AI training methods:

[0249] In addition to the Gradient Boosting Tree (GBDT) classifier, the following can also be used:

[0250] 1. Support Vector Machine (SVM) or Random Forest (RF) for multi-class emotional disorder classification; 2. Transformer-based temporal modeling methods (such as Temporal Attention Networks) for extracting long-term dynamic features.

[0251] 3. Semi-supervised or self-supervised learning frameworks are used to reduce the need for manually labeled data.

[0252] These methods can be selected based on the amount of data and available computing resources.

[0253] Sixth, alternative solutions for model deployment and updates:

[0254] In addition to edge OTA delivery, the model can also be updated in the following ways:

[0255] 1. Deploy the model in the cloud, and the handheld terminal can call the cloud inference service in real time.

[0256] 2. By embedding an FPGA or NPU module in the terminal, more complex deep neural network models can be run.

[0257] Using a federated learning mechanism, each terminal only uploads the model gradients to the server after local training, ensuring data privacy. In this embodiment, an objective detection of mood disorders is achieved through a progressive architecture of collection-analysis-diagnosis. Figure 3 This is a system block diagram showing the collaborative workflow of each module, where: A is the near-infrared data acquisition device, B is the handheld terminal (including B1 signal preprocessing, B2 feature analysis, B3 emotion diagnosis, B4 report generation, and B5 display screen), and C is the AI ​​training service (including C1 training data storage, C2 model training, and C3 new model output).

[0258] The specific implementation is as follows:

[0259] I. Diagnostic Process for Mood Disorders: Figure 4This is a flowchart of the diagnostic process for mood disorders.

[0260] First, data origin and synchronization

[0261] 1. fNIRS raw input: dual-wavelength light intensities I(λ1,t), I(λ2,t), each source-detector pair forming one channel; sampling rate 10–50Hz.

[0262] 2. EEG raw input: 21-channel electrode potential V based on the international 10–20 system e (t), the reference electrode is set to average reference or dual-papillary reference; the sampling rate is 250–1000Hz (default 500Hz); 3. Clock synchronization: A (collector) and B (handheld terminal) are calibrated via NTP protocol; frame-level timestamp alignment error ≤1ms.

[0263] Second, fNIRS preprocessing (B1)

[0264] Optical-to-Concentration Conversion (MBLL)

[0265]

[0266] 1. Parameters: E is the extinction coefficient matrix; L = 30–35 mm; DPF = 5.5–6.5 (adaptively adjusted according to the subject's age and head circumference).

[0267] 2. Filtering: FIR bandpass filtering (0.01–0.2Hz) is used to preserve physiologically relevant frequencies, while high-pass filtering (cutoff frequency 0.01Hz) is used to remove baseline drift.

[0268] 3. Motion artifacts: Artifacts are identified using a dual-threshold detection method combining amplitude thresholding and first-order difference, and corrected using short-time-window spline interpolation.

[0269] 4. Short probe distance regression: Using a short probe distance channel as a regressor, superficial tissue signal interference is eliminated in the generalized linear model (GLM).

[0270] 5. Channel Quality Scoring: Taking into account signal-to-noise ratio (SNR), coefficient of variation (CV), and inter-channel correlation consistency, channels below a certain threshold are marked as low-weighted.

[0271] Third, EEG preprocessing (B1)

[0272] 1. Re-reference: Use average reference or dual mastoid reference; use notch filtering (50 / 60Hz) to remove power line interference, and perform bandpass filtering (1–45Hz) to retain effective EEG frequency bands.

[0273] 2. ICA Artifact Removal: Independent Component Analysis (ICA) is applied to separate and remove noise components from electrooculography (EOG, such as blinking), electromyography (EMG), electrocardiography (ECG), and other channels.

[0274] 3. Trial cleaning: Remove bad trials with severe noise, retain good trials; perform baseline correction (using the first 10 seconds at rest as the baseline interval).

[0275] Fourth, windowing and baseline

[0276] 1. Online sliding window: A sliding window (W) with a length of 20–30 seconds and a step size of 2–5 seconds (S) is used; simultaneously, a circular buffer for maintaining individual baselines is maintained for a duration of no less than 120 seconds.

[0277] 2. Paradigm Task: The acquisition process consists of "baseline segment + stimulation segment + recovery segment"; if there is no paradigm task, only sliding window processing is performed. Fifth, Feature Analysis (B2)

[0278] 5.1 fNIRS features

[0279] 1. Spatial dimension: including the mean ROI (Region of Interest), the left and right hemisphere lateralization indices (LI), and the connection weights between the prefrontal cortex and the limbic system.

[0280] 2. Time dimension: This includes indicators such as Hurst exponent, sample entropy, multi-scale entropy, detrended volatility analysis (DFA) slope, peak latency, and half-peak width.

[0281] 3. Frequency domain dimension: including energy proportion in the 0.01–0.1Hz band, peak power spectrum frequency, etc. 4. Connectivity dimension: including correlation coefficient, partial correlation coefficient, coherence, phase-locked value (PLV) matrix, and graph theory-based metrics (degree, clustering coefficient, efficiency, betweenness centrality, and modularity).

[0282] 5.2 EEG Features

[0283] 1. Power Spectrum Characteristics: Calculates the power spectral density (PSD) for each frequency band, including delta wave (0.5–4Hz), theta wave (4–8Hz), alpha wave (8–13Hz), beta wave (13–30Hz), and gamma wave (30–45Hz). It can also further calculate the normalized ratio of the power in each frequency band.

[0284] 2. Emotion-related indicators: A typical indicator is frontal alpha asymmetry (FAA), which is calculated by subtracting the logarithm of the left frontal alpha power from the logarithm of the right frontal alpha power.

[0285] 3. Connectivity characteristics: including weighted phase lag index (wPLI), coherence and partial coherence indices; effective connectivity indices (such as Granger causality or dynamic transfer function DTF) can also be extracted when needed. 4. Event correlation characteristics: when a task paradigm exists, the peak values ​​(such as P2, N2) and their latencies of event-related potential (ERP) components can be extracted.

[0286] 5.3 Quality and Confidence 1. Channel and Electrode Quality Weights: A quality weight α_i∈[0,1] is assigned to each channel or electrode, and a weighted average method is used during feature aggregation.

[0287] 2. Window-level confidence: At the sliding window level, considering factors such as effective channel rate, signal-to-noise ratio (SNR), and artifact ratio, the confidence index q of the window is calculated. Sixth, emotion diagnosis (B3).

[0288] 1. Input features: fNIRS top map (64×64×2), fNIRS connection matrix M_n, EEG connection matrix M_e, and statistical feature vector F.

[0289] 2. Model Structure: A three-branch fusion structure is adopted, including CNN (processing the top image), GCN (processing M_n, M_e), and GBDT or MLP (processing F). Attention weights ω_n, ω_e, and ω_s are introduced into the fusion layer, and the constraint ω_n + ω_e + ω_s = 1 is applied.

[0290] 3. Dynamic weighting mechanism: The attention weight ω is adaptively allocated based on quality metrics (confidence q, effective channel rate) and historical performance.

[0291] 6.1 Probability and Uncertainty

[0292]

[0293] 6.2 Smoothing and Alarm Logic

[0294] R t =α·max(p_t)+(1-α)·R_{t-1}, 0<α<1

[0295] 1. Alarm Trigger Mechanism: When the system detects that the risk score R_t is greater than or equal to the threshold θ and the fluctuation index σ is less than or equal to the threshold τ within K consecutive time windows, it determines the state as "high risk" and triggers an alarm. Typical parameters are set to K=3, θ=0.8, and τ=0.2. These parameters can be modified according to the actual application scenario. 2. Subtype Disorder Differentiation Function: The system further outputs probability vectors of mood disorder subtypes, covering categories such as depression, anxiety, bipolar disorder, and neutral disorder, to assist in refined classification and subsequent intervention.

[0296] Seventh, Report Generation and Readability (B4 / B5)

[0297] 1. Report Fields: Output results include risk level, subtype probability, confidence interval, heatmap of abnormal brain regions based on fNIRS, and EEG / fNIRS connectivity plot.

[0298] 2. Interpretability: The model possesses interpretability analysis capabilities, including Grad-CAM-based top-graph visualization, channel or electrode importance ranking, and key edge contributions identified through the graph convolutional network (GCN) attention mechanism.

[0299] 3. Edge-side latency: During the edge-side inference phase, the inference latency for a single time window is less than 300 milliseconds; the generation and rendering latency of the complete report is less than 2 minutes.

[0300] 8. End-side data flow table (input → processing → output → delay)

[0301]

[0302] II. AI Training Service and Database Iteration System (Cloud-based C1–C3): Figure 5 System flowchart for AI training services and database iteration system.

[0303] Second, data standardization and storage (C1)

[0304] 1. Data Unit: The stored data unit includes the original summary ID, feature vectors F_n and F_e, connection matrices M_n and M_e, top graph hash value, quality metric (q_n, q_e), device parameters, task paradigm, individual baseline, label y, and timestamp.

[0305] 2. Anonymization: Direct identity information is removed, and the subject's ID is hashed; data transmission is encrypted using the TLS 1.3 protocol, and storage is encrypted using the AES-256 algorithm.

[0306] 3. Quality Control: Perform data integrity checks (field missing rate), physiological rationality checks (numerical range, drift trend, heart rate coupling), noise ratio analysis, and channel effectiveness evaluation.

[0307] 4. Version Management: Assign DataVersion and SchemaVersion identifiers to each uploaded data to ensure traceability and version consistency.

[0308] Second, training data construction and stratified sampling

[0309] 1. Label Alignment: Based on the DSM-5 criteria, the samples were labeled with the type of mood disorder (e.g., depression, anxiety, etc.), and the scale score ranges (e.g., PHQ-9, GAD-7 scale score ranges) were recorded.

[0310] 2. Layered Strategy: Adjustments are made based on data collection center, gender, age, device type, and multiple other strategies.

[0311] Third, model training (C2)

[0312] 1. Multi-branch Deep Model: Employing a multi-branch fusion architecture, including Convolutional Neural Networks (CNN, for processing top graphs), Graph Convolutional Networks (GCN, for processing connection matrices M_n and M_e), and Multilayer Perceptrons (MLP, for processing statistical features). This model structure maintains homology with the edge-side inference structure to facilitate knowledge distillation.

[0313] 2. Base Learner: Gradient Boosting Decision Tree (GBDT) is used to perform lightweight classification of 12 core neural features, which is used for edge deployment and comparison with baseline.

[0314] 3. Hyperparameter search: Hyperparameter search is performed using Bayesian optimization methods or the Optuna framework, and model evaluation is conducted by combining cross-validation and time buffer strategies.

[0315] 4. Model Calibration: The model output is calibrated using temperature scaling and Platt scaling methods to generate more reliable confidence estimates, and reliability metrics such as expected calibration error (ECE) and maximum calibration error (MCE) are provided.

[0316] 5. Fairness and robustness assessment: ROC-AUC and ECE indicators were assessed according to population subgroups, and robustness tests were conducted against noise disturbances and missing channels.

[0317] Fourth, incremental learning and active learning

[0318] 1. Incremental Retraining: After new samples are manually verified and added to the database, the system triggers a retraining process based on time (weekly) and sample size thresholds. To avoid the forgetting effect, Elastic Weight Consolidation (EWC) and a replay buffer mechanism are used to enhance stability.

[0319] 2. Active Learning: Based on the uncertainty index σ of the prediction results and the distribution drift measures (PSI and KL divergence), high-value samples are dynamically selected to prioritize entering the labeling process, thereby improving model training efficiency.

[0320] 3. Semi-supervised learning: Introducing consistency regularization and pseudo-label methods on unlabeled data improves the model's representation learning ability and generalization performance.

[0321] Fifth, model registration, deployment, and rollback (C3→B3)

[0322] 1. Stratified sampling is performed based on task type, with the dataset divided into three parts: 70% training set, 15% validation set, and 15% test set. 2. Class imbalance handling: To address the class imbalance problem, methods including focal loss, cost-sensitive learning, and resampling techniques (SMOTE oversampling and stratified undersampling) are employed.

[0323] 3. Model Registration: The system registers and manages each model, recording information including the model version number (ModelVersion), the corresponding training data version, key performance metrics, and interpretability outputs (Grad-CAM heatmap and feature importance ranking results).

[0324] 4. Gradual Deployment: A gradual deployment strategy will be adopted, with the deployment percentage increasing from 5% to 25% and then to 100%. During this process, A / B testing or control experiments will be conducted, and online performance metrics, including AUC, ECE, and alarm rate, will be continuously monitored.

[0325] 5. Compressed Deployment: During the model deployment phase, model compression optimization is supported, including quantization (INT8), pruning, and knowledge distillation. The new version automatically replaces the old model after passing the on-device self-check, while the system retains a rollback channel to ensure availability and security.

[0326] Sixth, data and model drift monitoring

[0327] 1. Data Drift Monitoring: Continuous monitoring of channel and electrode statistics triggers an alert when the Population Stability Index (PSI) exceeds 0.2. Simultaneously, structural changes in the connectivity graph are monitored, using spectral radius and modularity as evaluation metrics.

[0328] 2. Concept Drift Monitoring: When online model performance metrics show a downward trend (e.g., AUC decreases or ECE increases), the system triggers a retraining process and audits relevant features to ensure model performance stability. 3. Log and Compliance Audit: All events throughout the system's operation are logged to ensure that data and model processing meet compliance and traceability requirements.

[0329] VII. Database and Field Examples

[0330]

[0331] This solution ensures that the processing chain from raw data to diagnostic results remains continuous, traceable, and rollbackable, and continuously improves the robustness and interpretability of the model in real-world application scenarios through a closed-loop learning mechanism.

[0332] This embodiment provides a method for identifying mood disorders based on near-infrared brain imaging (fNIRS) technology. During the detection process, the subject wears a headgear containing a near-infrared light source and a detector array. The detector array covers brain regions related to mood regulation, such as the prefrontal cortex, dorsolateral prefrontal cortex, frontal lobe, and temporal lobe. The NIRS device simultaneously acquires the concentration change curves of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR), outputting analog signals. These signals are converted into digital signals by a near-infrared data acquisition unit and transmitted to a handheld terminal. The signal preprocessing module of the handheld terminal performs motion artifact removal, baseline correction, multivariate Bell's Law (MBLL) transformation, and filtering. Subsequently, the feature extraction module calculates functional connectivity strength (prefrontal-amygdala correlation) in the spatial dimension, calculates nonlinear indices of concentration dynamics (sample entropy, Hearst exponent) in the temporal dimension, and extracts low-frequency oscillation power spectrum (0.01–0.1 Hz) in the frequency domain. These multidimensional features constitute a neural marker matrix for subsequent mood disorder analysis.

[0333] This embodiment provides a handheld mood disorder identification device, including: an fNIRS signal acquisition module, an embedded processor, an interactive display screen, and a power supply module. When the patient wears the headgear, the device automatically starts signal acquisition. The embedded processor performs multi-dimensional feature calculations locally, including 12 parameters such as functional connectivity network density and oxygen metabolism time-varying index. The mood diagnosis module classifies the above features based on a pre-trained model, outputting the mood disorder risk level and possible subtypes. The diagnostic results are converted into graphic and textual forms by a visualization report generation module, including: risk level prompts, abnormal brain region heat maps, and feature curve displays. Users can view the report on the display screen in real time. The entire process requires no additional computer or professional software support.

[0334] This embodiment provides a high-speed data processing system. The handheld terminal is equipped with a USB-C interface and a 2.4GHz wireless module, supporting both wired and wireless transmission. In wired mode, the USB-C establishes a 5Gbps channel with a data transmission latency of less than 1ms, suitable for high-interference environments such as ICUs. In wireless mode, a customized protocol based on IEEE 802.11ax is used, supporting AES-256 encryption, with a peak rate of up to 800Mbps. Upon arrival at the handheld terminal, the system automatically triggers a real-time analysis process, eliminating the need for manual copying. The embedded processor performs dynamic HbO / HbR analysis, functional connectivity calculation, and time-frequency index extraction, generating a diagnostic report within approximately 2 minutes. The report includes: risk level (low, medium, high), abnormal brain region localization (abnormal activation in the left prefrontal cortex), probability of mood disorder subtype (depression / anxiety, etc.), and outputs corresponding heatmaps and connectivity strength matrices.

[0335] This embodiment provides a closed-loop AI training system, including a data acquisition terminal, a cloud database, and a model training service. The data acquisition terminal is the aforementioned handheld terminal. The collected multidimensional features (including 12 categories such as hemoglobin dynamic curves and functional connectivity indicators) along with clinical diagnostic labels are uploaded to the cloud database. The database uses a unified structured storage according to the DSM-5 standard to construct a brain function biomarker library for mood disorders. The model training service performs incremental learning based on the database: it uses a gradient boosting tree (GBDT) and a deep neural network classifier to train the input features and outputs the probability distribution of disorder subtypes. When new data is added to the database, the system automatically triggers retraining, updates the classification model, and records version information. The compressed and quantized new model is delivered to the handheld terminal via OTA, allowing it to access the latest algorithm even offline. During clinical use, diagnostic reports and user feedback are returned to the database as training data for the next round, achieving closed-loop updates and ensuring continuous model optimization.

[0336] A system for acquiring an emotion disorder identification and control model based on near-infrared brain imaging atlas features includes the following: a unit 10 for establishing a model based on the acquisition and analysis of multi-dimensional neural markers based on fNIRS.

[0337] Unit 20 is established based on an integrated handheld diagnostic device, unit 30 is established based on a dual-mode high-speed intelligent analysis system, and unit 40 is established based on a closed-loop AI training and database iteration system.

[0338] A controller for identifying mood disorders based on near-infrared brain imaging atlas features includes the following: a control model for identifying mood disorders based on near-infrared brain imaging atlas features is stored in the controller.

[0339] In this embodiment, the control model for emotion disorder recognition based on near-infrared brain imaging atlas features is obtained according to the above-described method for obtaining the control model for emotion disorder recognition based on near-infrared brain imaging atlas features.

[0340] Step 100: Multidimensional neural marker acquisition and analysis based on fNIRS

[0341] Step 200: Integrated handheld diagnostic device,

[0342] Step 300: Dual-mode high-speed intelligent analysis system.

[0343] Step 400: Closed-loop AI training and database iteration system.

[0344] A method for identifying and controlling mood disorders based on near-infrared brain imaging atlas features includes the following: applying a mood disorder identification controller based on near-infrared brain imaging atlas features in a CPU for control.

[0345] This embodiment provides a method for identifying mood disorders based on near-infrared brain imaging (fNIRS) technology. During the detection process, the subject wears a headgear containing a near-infrared light source and a detector array. The detector array covers brain regions related to mood regulation, such as the prefrontal cortex, dorsolateral prefrontal cortex, frontal lobe, and temporal lobe. The NIRS device simultaneously acquires the concentration change curves of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR), outputting analog signals. These signals are converted into digital signals by a near-infrared data acquisition unit and transmitted to a handheld terminal. The signal preprocessing module of the handheld terminal performs motion artifact removal, baseline correction, multivariate Bell's Law (MBLL) transformation, and filtering. Subsequently, the feature extraction module calculates functional connectivity strength (prefrontal-amygdala correlation) in the spatial dimension, calculates nonlinear indices of concentration dynamics (sample entropy, Hearst exponent) in the temporal dimension, and extracts low-frequency oscillation power spectrum (0.01–0.1 Hz) in the frequency domain. These multidimensional features constitute a neural marker matrix for subsequent mood disorder analysis.

[0346] This embodiment provides a handheld mood disorder identification device, including: an fNIRS signal acquisition module, an embedded processor, an interactive display screen, and a power supply module. When the patient wears the headgear, the device automatically starts signal acquisition. The embedded processor performs multi-dimensional feature calculations locally, including 12 parameters such as functional connectivity network density and oxygen metabolism time-varying index. The mood diagnosis module classifies the above features based on a pre-trained model, outputting the mood disorder risk level and possible subtypes. The diagnostic results are converted into graphic and textual forms by a visualization report generation module, including: risk level prompts, abnormal brain region heat maps, and feature curve displays. Users can view the report on the display screen in real time. The entire process requires no additional computer or professional software support.

[0347] This embodiment provides a high-speed data processing system. The handheld terminal is equipped with a USB-C interface and a 2.4GHz wireless module, supporting both wired and wireless transmission. In wired mode, the USB-C establishes a 5Gbps channel with a data transmission latency of less than 1ms, suitable for high-interference environments such as ICUs. In wireless mode, a customized protocol based on IEEE 802.11ax is used, supporting AES-256 encryption, with a peak rate of up to 800Mbps. Upon arrival at the handheld terminal, the system automatically triggers a real-time analysis process, eliminating the need for manual copying. The embedded processor performs dynamic HbO / HbR analysis, functional connectivity calculation, and time-frequency index extraction, generating a diagnostic report within approximately 2 minutes. The report includes: risk level (low, medium, high), abnormal brain region localization (abnormal activation in the left prefrontal cortex), probability of mood disorder subtype (depression / anxiety, etc.), and outputs corresponding heatmaps and connectivity strength matrices.

[0348] This embodiment provides a closed-loop AI training system, including a data acquisition terminal, a cloud database, and a model training service. The data acquisition terminal is the aforementioned handheld terminal. The collected multidimensional features (including 12 categories such as hemoglobin dynamic curves and functional connectivity indicators) along with clinical diagnostic labels are uploaded to the cloud database. The database uses a unified structured storage according to the DSM-5 standard to construct a brain function biomarker library for mood disorders. The model training service performs incremental learning based on the database: it uses a gradient boosting tree (GBDT) and a deep neural network classifier to train the input features and outputs the probability distribution of disorder subtypes. When new data is added to the database, the system automatically triggers retraining, updates the classification model, and records version information. The compressed and quantized new model is delivered to the handheld terminal via OTA, allowing it to access the latest algorithm even offline. During clinical use, diagnostic reports and user feedback are returned to the database as training data for the next round, achieving closed-loop updates and ensuring continuous model optimization.

[0349] The above embodiments are merely one implementation of the method and system for obtaining an emotional disorder identification and control model based on near-infrared brain imaging atlas features, as well as the controller and control method provided by the present invention. Any other modifications to the solution provided by the present invention, including adding or reducing components or steps, or applying the present invention to other technical fields similar to the present invention, shall all fall within the protection scope of the present invention.

Claims

1. A method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features, characterized by the following steps: The steps are as follows: Step 100: Multidimensional neural marker acquisition and analysis based on fNIRS Step 200: Integrated handheld diagnostic device, Step 300: Dual-mode high-speed intelligent analysis system. Step 400: Closed-loop AI training and database iteration system.

2. The method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features according to claim 1, characterized in that: A near-infrared light source array was used to cover the core brain regions for emotion regulation, including the prefrontal cortex, temporal lobe, frontal lobe, central sulcus, parietal lobe, and occipital lobe. The system monitors the dynamic changes in the concentrations of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) in real time and outputs raw light intensity data. The concentration change curve was obtained through Bell-Lambert Law (MBLL) conversion. The feature extraction module analyzes neural markers according to the following dimensions: First, the spatial dimension: calculating the strength of functional connectivity between brain regions, such as the relevant weights of the prefrontal-limbic system. Second, the time dimension: nonlinear features, including the Hurst exponent and sample entropy, are extracted from the time-varying curve of hemoglobin concentration. Third, frequency domain dimension: power spectrum analysis is performed on low-frequency oscillations (0.01–0.1Hz) to extract frequency domain energy distribution characteristics. The processed multidimensional features are input into the subsequent diagnostic module for the identification of mood disorders.

3. The method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features according to claim 1, characterized in that: include: The device consists of an fNIRS signal acquisition module, an embedded processor, a display screen, and a power management module. After the patient wears the headgear, the device automatically completes signal acquisition. The embedded processor calculates twelve features locally, namely functional connectivity network density and oxygen metabolism time-varying nonlinear exponent, and inputs the results into the diagnostic model. The model outputs the risk level and subtype probability of mood disorders, and then the report generation module presents the visualization results on the screen, including: risk level prompts, abnormal brain region heat maps and subtype suggestions.

4. The method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features according to claim 1, characterized in that: A dual-mode data transmission architecture to support efficient data acquisition and processing: First, wired mode: A 5Gbps channel is established via the USB-C interface, with an end-to-end latency of less than 1ms, making it suitable for ICU or operating room environments with strong electromagnetic interference. Second, wireless mode: It integrates a customized 2.4GHz protocol based on IEEE 802.11ax, supports dynamic bandwidth allocation and AES-256 encryption, and has a peak rate of up to 800Mbps, making it suitable for mobile deployment scenarios such as community follow-up and pediatric wards. The system features an automatic triggering mechanism: when the terminal successfully establishes a wired connection or wireless pairing with the fNIRS device, the data stream pipeline is automatically activated, enabling real-time transmission and eliminating the need for manual copying. During the acquisition process, the embedded processor simultaneously calculates spatial dimension indicators, namely prefrontal-amygdala functional connectivity, temporal dimension (HbO / HbR sample entropy), and frequency domain dimension (θ band power ratio), and generates a diagnostic report within 2 minutes, which includes risk level, abnormal brain region heat map, and subtype probability.

5. The method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features according to claim 1, characterized in that: It consists of three parts: database construction, model training, and edge deployment. First, standardized database construction: The system automatically collects detection data from handheld terminals, including twelve categories of features and clinical diagnostic labels, and completes the classification and labeling of depression and anxiety according to the DSM-5 standard, storing it in the brain functional biomarker library for mood disorders. Second, the dynamic AI optimization mechanism: The model training module constructs an emotion disorder classifier based on Gradient Boosting Tree (GBDT). The input is 12-dimensional neural features, and the output is the probability distribution of disorder subtypes. For newly added labeled data, the system adopts an incremental learning strategy to add it to the training set and update the model, thereby improving the model's generalization ability and subtype discrimination accuracy. Grassroots empowerment strategy: The updated model is distributed to handheld terminals after lightweight processing such as parameter quantization and pruning, so that the terminals can call the latest model even when offline. Users only need to perform a single-click operation to complete the detection and diagnosis. It is suitable for community hospitals and primary medical scenarios. Data upload, model training and distribution form a closed loop to ensure continuous model iteration and stable improvement of diagnostic performance.

6. The method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features according to claim 1, characterized in that: First, the combined fNIRS and EEG acquisition scheme: In addition to using fNIRS signals alone, a multimodal alternative is provided: EEG electrodes are integrated into the headgear and arranged according to the international 10-20 system to collect EEG potential signals. After notch filtering and artifact removal by independent component analysis, EEG data were fused with hemoglobin dynamic features extracted by fNIRS for multimodal analysis. Fusion methods may include: feature-level concatenation, decision-level ensemble, or joint modeling based on graph convolutional networks (GCNs). This alternative approach can enhance classification robustness even when the fNIRS signal-to-noise ratio is low. Second, different feature extraction methods: In addition to nonlinear features such as the Hearst exponent and sample entropy, the following alternative indicators can be extracted:

1. Time domain: based on detrended volatility analysis (DFA) and the Lyapunov exponent.

2. Frequency Domain: Wavelet packet energy distribution, Empirical Mode Decomposition (EMD) characteristics.

3. Spatial domain: Graph theory metrics such as degree centrality, characteristic path length, and clustering coefficient. These indicators can replace or supplement the original twelve categories of characteristics to adapt to different populations or clinical scenarios. Third, alternative forms of handheld terminals: In addition to integrated handheld devices, this system can also adopt a modular architecture: The data acquisition unit (headgear + fNIRS module) and the analysis unit (smartphone or tablet) are directly connected via Bluetooth Low Energy or Wi-Fi. The acquisition module only completes the raw signal acquisition; feature extraction and diagnostic model operation are handled by an external mobile terminal. This solution reduces equipment weight, making it suitable for follow-up scenarios involving children and the elderly. Fourth, different data transmission and encryption mechanisms: In addition to USB-C and 2.4GHz wireless, alternative solutions include:

1. Fiber optic interface (SFP module) enables higher bandwidth transmission.

2. 5G mobile communication module, used for remote real-time monitoring.

3. End-to-end encryption can use the Chinese national standard SM4 or TLS 1.3 instead of AES-256 to adapt to the regulatory requirements of different regions. Fifth, alternative AI training methods: In addition to the Gradient Boosting Tree (GBDT) classifier, the following can also be used:

1. Use Support Vector Machine (SVM) or Random Forest (RF) to classify multiple emotion disorders.

2. Transformer-based temporal modeling methods (such as temporal attention networks) are used to extract long-term dynamic features.

3. Semi-supervised or self-supervised learning frameworks are used to reduce the need for manually labeled data. These methods can be selected based on the amount of data and available computing resources. Sixth, alternative solutions for model deployment and updates: In addition to edge OTA delivery, the model can also be updated in the following ways:

1. Deploy the model in the cloud, and the handheld terminal can call the cloud inference service in real time.

2. By embedding an FPGA or NPU module in the terminal, more complex deep neural network models can be run. Using a federated learning mechanism, each terminal only uploads the model gradients to the server after training locally, ensuring data privacy.

7. The method for obtaining an emotion disorder identification and control model based on near-infrared brain imaging atlas features according to claim 1, characterized in that: Objective detection of mood disorders is achieved through a progressive architecture of data acquisition, analysis, and diagnosis. Figure 3 shows the system block diagram of the collaborative workflow of each module, where: A is the near-infrared data acquisition device, B is the handheld terminal (including B1 signal preprocessing, B2 feature analysis, B3 emotion diagnosis, B4 report generation, and B5 display screen), and C is the AI ​​training service (including C1 training data storage, C2 model training, and C3 new model output). The steps are as follows: I. Diagnostic Process for Mood Disorders: Figure 4 is a system flowchart for the diagnostic process of mood disorders. First, data origin and synchronization 1. fNIRS raw input: dual-wavelength light intensities I(λ1,t), I(λ2,t), each source-detector pair forming one channel; sampling rate 10–50Hz.

2. EEG raw input: 21-channel electrode potential V based on the international 10–20 system e (t), the reference electrode is set to average reference or dual-mammary gland reference; sampling rate 250–1000Hz (default 500Hz), 3. Clock synchronization: A (data acquisition unit) and B (handheld terminal) are calibrated via the NTP protocol; Frame-level timestamp alignment error ≤1ms Second, fNIRS preprocessing (B1), optical-to-concentration conversion (MBLL).

1. Parameters: E is the extinction coefficient matrix; L = 30–35 mm; DPF = 5.5–6.5 (adaptively adjusted according to the subject's age and head circumference).

2. Filtering: FIR bandpass filtering (0.01–0.2Hz) is used to preserve physiologically relevant frequencies, while high-pass filtering (cutoff frequency 0.01Hz) is used to remove baseline drift.

3. Motion artifacts: Artifacts are identified using a dual-threshold detection method combining amplitude thresholding and first-order difference, and corrected using short-time-window spline interpolation.

4. Short probe distance regression: Using a short probe distance channel as a regressor, superficial tissue signal interference is eliminated in the generalized linear model (GLM).

5. Channel Quality Scoring: Taking into account signal-to-noise ratio (SNR), coefficient of variation (CV), and inter-channel correlation consistency, channels below a certain threshold are marked as low-weighted. Third, EEG preprocessing (B1) 1. **Reference:** Use average reference or dual mastoid reference; employ notch filtering (50 / 60Hz) to remove power line interference and bandpass filtering (1–45Hz) to retain effective EEG frequencies.

2. **ICA Artifact Removal:** Apply independent component analysis (ICA) to separate and remove noise components from electrooculography (EOG, such as blinking), electromyography (EMG), electrocardiography (ECG), and other channels.

3. **Trial Washing:** Eliminate heavily noisy bad trials, retaining good trials; perform baseline correction (using the 10s before rest as the baseline interval). Fourth, windowing and baseline 1. Online sliding window: A sliding window (W) with a length of 20–30 seconds and a step size of 2–5 seconds (S) is used; a circular buffer for maintaining the individual baseline is maintained simultaneously for a duration of no less than 120 seconds.

2. Paradigm task: The acquisition process consists of "pre-baseline segment + stimulation segment + recovery segment"; if there is no paradigm task, only sliding window processing is performed. Fifth, Feature Analysis (B2) 5.1 fNIRS features 1. Spatial dimension: including the mean ROI (Region of Interest), the left and right hemisphere lateralization indices (LI), and the connection weights between the prefrontal cortex and the limbic system.

2. Time dimension: This includes indicators such as Hurst exponent, sample entropy, multi-scale entropy, detrended volatility analysis (DFA) slope, peak latency, and half-peak width.

3. Frequency domain dimension: including indicators such as the energy proportion of the 0.01–0.1Hz frequency band and the peak frequency of the power spectrum.

4. Connectivity dimension: including correlation coefficient, partial correlation coefficient, coherence, phase-locked value (PLV) matrix, and graph theory-based metrics (degree, clustering coefficient, efficiency, betweenness centrality, and modularity). 5.2 EEG Features 1. Power Spectrum Characteristics: Calculates the power spectral density (PSD) for each frequency band, including delta wave (0.5–4Hz), theta wave (4–8Hz), alpha wave (8–13Hz), beta wave (13–30Hz), and gamma wave (30–45Hz). It can also further calculate the normalized ratio of the power in each frequency band.

2. Emotion-related indicators: A typical indicator is frontal alpha asymmetry (FAA), which is calculated by subtracting the logarithm of the left frontal alpha power from the logarithm of the right frontal alpha power.

3. Connectivity characteristics: including weighted phase lag index (wPLI), coherence and partial coherence indices; effective connectivity indices (such as Granger causality or dynamic transfer function DTF) can also be extracted when needed.

4. Event-related characteristics: When a task paradigm exists, the peak values ​​(such as P2 and N2) and their latencies of event-related potential (ERP) components can be extracted. 5.3 Quality and Confidence 1. Channel and Electrode Quality Weights: A quality weight α_i∈[0,1] is assigned to each channel or electrode, and a weighted average method is used during feature aggregation.

2. Window-level confidence: At the sliding window level, considering factors such as effective channel rate, signal-to-noise ratio (SNR), and artifact ratio, the confidence index q of the window is calculated. Sixth, emotion diagnosis (B3).

1. Input features: fNIRS top map (64×64×2), fNIRS connection matrix M_n, EEG connection matrix M_e, and statistical feature vector F.

2. Model Structure: A three-branch fusion structure is adopted, including CNN (processing the top image), GCN (processing M_n, M_e), and GBDT or MLP (processing F). Attention weights ω_n, ω_e, and ω_s are introduced into the fusion layer, and the constraint ω_n + ω_e + ω_s = 1 is applied.

3. Dynamic weighting mechanism: The attention weight ω is adaptively allocated based on quality metrics (confidence q, effective channel rate) and historical performance. 6.1 Probability and Uncertainty p = softmax(f {fusion}(x) ); Confidence level ~ via MC-Dropout: σ = Var({p {(k)} } {k＝1..K} ), 6.2 Smoothing and Alarm Logic R t =α·max(p_t)+(1-α)·R_{t-1},0<α<1 1. Alarm Trigger Mechanism: When the risk score R_t is greater than or equal to the threshold θ and the fluctuation index σ is less than or equal to the threshold τ for K consecutive time windows, the system determines the state as "high risk" and triggers an alarm. Typical parameters are set to K=3, θ=0.8, and τ=0.

2. These parameters can be modified according to the actual application scenario.

2. Subtype discrimination function: The system further outputs probability vectors of mood disorder subtypes, covering categories such as depression, anxiety, bipolar, and neutral, to assist in refined classification and subsequent intervention. Seventh, report generation and readability (B4 / B5) 1. Report Fields: Output results include risk level, subtype probability, confidence interval, heatmap of abnormal brain regions based on fNIRS, and EEG / fNIRS connectivity plot.

2. Interpretability: The model possesses interpretability analysis capabilities, including Grad-CAM-based top-graph visualization, channel or electrode importance ranking, and key edge contributions identified through the graph convolutional network (GCN) attention mechanism.

3. Edge-side latency: During the edge-side inference phase, the inference latency for a single time window is less than 300 milliseconds; the generation and rendering latency of the complete report is less than 2 minutes.

8. End-side data flow table (input → processing → output → delay) II. AI Training Service and Database Iteration System (Cloud Side C1–C3): Figure 5 is a flowchart of the AI ​​Training Service and Database Iteration System. Second, data standardization and storage (C1) 1. Data Unit: The stored data unit includes the original summary ID, feature vectors F_n and F_e, connection matrices M_n and M_e, top graph hash value, quality metric (q_n, q_e), device parameters, task paradigm, individual baseline, label y, and timestamp.

2. Anonymization: Direct identity information is removed, and subject IDs are hashed; data transmission is encrypted using the TLS 1.3 protocol, and storage is encrypted using the AES-256 algorithm.

3. Quality Control: Data integrity checks (field missing rate), physiological rationality checks (numerical range, drift trend, heart rate coupling), noise ratio analysis, and channel effectiveness evaluation are performed.

4. Version Management: Assign DataVersion and SchemaVersion identifiers to each uploaded data to ensure traceability and version consistency. Second, training data construction and stratified sampling 1. Label Alignment: Based on the DSM-5 criteria, the samples were labeled with the type of mood disorder (e.g., depression, anxiety, etc.), and the scale score ranges (e.g., PHQ-9, GAD-7 scale score ranges) were recorded.

2. Layered Strategy: Adjustments are made based on data collection center, gender, age, device type, and multiple other strategies. Third, model training (C2) 1. Multi-branch Deep Model: Employing a multi-branch fusion architecture, including Convolutional Neural Networks (CNN, for processing top graphs), Graph Convolutional Networks (GCN, for processing connection matrices M_n and M_e), and Multilayer Perceptrons (MLP, for processing statistical features). This model structure maintains homology with the edge-side inference structure to facilitate knowledge distillation.

2. Base Learner: Gradient Boosting Decision Tree (GBDT) is used to perform lightweight classification of 12 core neural features, which is used for edge deployment and comparison with baseline.

3. Hyperparameter search: Hyperparameter search is performed using Bayesian optimization methods or the Optuna framework, and model evaluation is conducted by combining cross-validation and time buffer strategies.

4. Model Calibration: The model output is calibrated using temperature scaling and Platt scaling methods to generate more reliable confidence estimates, and reliability metrics such as expected calibration error (ECE) and maximum calibration error (MCE) are provided.

5. Fairness and Robustness Assessment: ROC-AUC and ECE indicators were assessed according to population subgroups. Robustness tests against noise disturbances and missing channels were also conducted. Fourth, Incremental Learning and Active Learning.

1. Incremental Retraining: After new samples are manually verified and added to the database, the system triggers a retraining process based on time (weekly) and sample size thresholds. To avoid the forgetting effect, Elastic Weight Consolidation (EWC) and a replay buffer mechanism are used to enhance stability.

2. Active Learning: Based on the uncertainty index σ of the prediction results and the distribution drift measures (PSI and KL divergence), high-value samples are dynamically selected to prioritize entering the labeling process, thereby improving model training efficiency.

3. Semi-supervised learning: Introducing consistency regularization and pseudo-label methods on unlabeled data improves the model's representation learning ability and generalization performance. Fifth, model registration, deployment, and rollback (C3→B3) 1. Stratified sampling is performed based on task type, with the dataset divided into three parts: 70% training set, 15% validation set, and 15% test set.

2. Class Imbalance Handling: To address the problem of imbalanced class distribution, methods including focal loss, cost-sensitive learning, and resampling techniques (SMOTE oversampling and hierarchical undersampling) are employed.

3. Model Registration: The system registers and manages each model, recording information including the model version number (ModelVersion), the corresponding training data version, key performance metrics, and interpretability outputs (Grad-CAM heatmap and feature importance ranking results).

4. Gradual Deployment: A gradual deployment strategy will be adopted, with the deployment percentage increasing from 5% to 25% and then to 100%. During this process, A / B testing or control experiments will be conducted, and online performance metrics, including AUC, ECE, and alarm rate, will be continuously monitored.

5. Compressed Deployment: During the model deployment phase, model compression optimization is supported, including quantization (INT8), pruning, and knowledge distillation. The new version automatically replaces the old model after passing the on-device self-check, while the system retains a rollback channel to ensure availability and security. Sixth, data and model drift monitoring 1. Data Drift Monitoring: Continuous monitoring of channel and electrode statistics triggers an alert when the Population Stability Index (PSI) exceeds 0.

2. Simultaneously, structural changes in the connectivity graph are monitored, using spectral radius and modularity as evaluation metrics.

2. Concept drift monitoring: When online model performance metrics show a downward trend (e.g., AUC decreases or ECE increases), the system triggers a retraining process and audits relevant features to ensure model performance stability.

3. Logging and Compliance Audit: Log all events throughout the system's operation to ensure that data and model processing meet compliance and traceability requirements. VII. Database and Field Examples This solution ensures that the processing chain from raw data to diagnostic results remains continuous, traceable, and rollbackable. Furthermore, it continuously improves the model's robustness and interpretability in real-world application scenarios through a closed-loop learning mechanism. Alternatively, a method for identifying mood disorders based on near-infrared brain imaging (fNIRS) technology involves the following steps: During the detection process, the subject wears a headgear containing a near-infrared light source and a detector array. The detector array covers brain regions related to mood regulation, including the prefrontal cortex, dorsolateral prefrontal cortex, frontal lobe, and temporal lobe. The NIRS device simultaneously acquires the concentration change curves of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR), outputting analog signals. These signals are converted into digital signals by a near-infrared data acquisition unit and transmitted to a handheld terminal. The handheld terminal's signal preprocessing module performs motion artifact removal, baseline correction, multivariate Bell's Law (MBLL) transformation, and filtering. Subsequently, the feature extraction module calculates functional connectivity strength (prefrontal-amygdala correlation) in the spatial dimension, calculates nonlinear indices of concentration dynamics (sample entropy, Hearst exponent) in the temporal dimension, and extracts low-frequency oscillation power spectrum (0.01–0.1 Hz) in the frequency domain. These multidimensional features constitute a neural marker matrix for subsequent mood disorder analysis. Alternatively, a handheld mood disorder identification device includes: an fNIRS signal acquisition module, an embedded processor, an interactive display screen, and a power supply module. When the patient wears the headgear, the device automatically initiates signal acquisition. The embedded processor performs multi-dimensional feature calculations locally, including 12 parameters such as functional connectivity network density and oxygen metabolism time-varying index. The mood diagnosis module classifies these features based on a pre-trained model, outputting the mood disorder risk level and possible subtypes. The diagnostic results are converted into graphic and textual formats by a visualization report generation module, including: risk level indications, abnormal brain region heatmaps, and feature curve displays. Users can view the report in real time on the display screen. The entire process requires no additional computer or specialized software support. Alternatively, a high-speed data processing system is proposed, with a handheld terminal equipped with a USB-C interface and a 2.4GHz wireless module, supporting both wired and wireless transmission. In wired mode, the USB-C establishes a 5Gbps channel with a data transmission latency of less than 1ms, suitable for high-interference environments such as ICUs. In wireless mode, a customized protocol based on IEEE 802.11ax is used, supporting AES-256 encryption, with a peak rate of up to 800Mbps. Upon data arrival at the handheld terminal, a real-time analysis process is automatically triggered, eliminating the need for manual copying. The embedded processor performs dynamic HbO / HbR analysis, functional connectivity calculation, and time-frequency index extraction, generating a diagnostic report within approximately 2 minutes. The report includes: risk level (low, medium, high), abnormal brain region localization (abnormal activation in the left prefrontal cortex), probability of mood disorder subtype (depression / anxiety, etc.), and outputs corresponding heatmaps and connectivity strength matrices. Alternatively, a closed-loop AI training system includes a data acquisition terminal, a cloud database, and a model training service. The data acquisition terminal is the aforementioned handheld terminal. The collected multidimensional features (including 12 categories such as hemoglobin dynamic curves and functional connectivity indicators) along with clinical diagnostic labels are uploaded to the cloud database. The database uses a unified structured storage according to the DSM-5 standard to construct a brain functional biomarker library for mood disorders. The model training service performs incremental learning based on the database: using Gradient Boosting Tree (GBDT) and a deep neural network classifier, the system trains the input features and outputs the probability distribution of the disorder subtype. When new data is added to the database, the system automatically triggers retraining, updates the classification model, and records version information. The compressed and quantized new model is delivered to the handheld terminal via OTA, allowing it to access the latest algorithm even offline. During clinical use, diagnostic reports and user feedback are returned to the database as training data for the next round, achieving closed-loop updates and ensuring continuous model optimization.

8. A system for acquiring an emotion disorder identification and control model based on near-infrared brain imaging atlas features, comprising the following: Unit 10 was established based on the multidimensional neural marker acquisition and analysis using fNIRS. Based on the integrated handheld diagnostic device, unit 20 is established. Unit 30 is established based on the dual-mode high-speed intelligent analysis system. Unit 40 is established based on the closed-loop AI training and database iteration system.

9. A controller for identifying mood disorders based on near-infrared brain imaging atlas features, comprising the following: a control model for identifying mood disorders based on near-infrared brain imaging atlas features is stored in the controller. The control model for emotion disorder recognition based on near-infrared brain imaging atlas features was obtained according to the above-mentioned method for obtaining the control model for emotion disorder recognition based on near-infrared brain imaging atlas features.

10. A method for identifying and controlling mood disorders based on near-infrared brain imaging atlas features, comprising the following: The CPU is controlled by an emotion disorder recognition controller based on near-infrared brain imaging atlas features, according to claim 9.

Images