Carotid artery health monitoring method and system based on multi-modal sensing fusion

Abstract

The invention relates to the technical field of biomedical sensing, and particularly discloses a carotid artery health monitoring method and system based on multi-modal sensing fusion, and the method comprises the steps: S1, data collection; s2, quality evaluation and adaptive acquisition; s3, preprocessing and feature fusion; s4, establishing a personalized dynamic baseline; s5, performing time sequence prediction and multi-level risk judgment; s6, an intelligent intervention and dynamic linkage module; according to the method, ultrasonic Doppler, photoplethysmography (PPG), body surface bioimpedance, neck pressure time sequence and other multi-mode signals are synchronously collected and fused, and the method can correct noise and disturbance in a single mode by utilizing complementary information among the modes; therefore, the accuracy and the time sequence robustness of hemodynamic parameter identification are remarkably improved. A multi-source cross validation mechanism can reduce the influence on overall judgment when single sensing fails or is abnormal, and the reliability of the overall system is improved.

Description

Technical Field

[0001] This invention belongs to the field of biomedical sensing technology, specifically relating to a carotid artery health monitoring method and system based on multimodal sensor fusion. Background Technology

[0002] The health status of the carotid arteries is closely related to the occurrence and development of cardiovascular diseases. Accurate and timely monitoring of carotid artery health is of great significance for the prevention, early diagnosis, and treatment of cardiovascular diseases. Currently, some progress has been made in the field of carotid artery health monitoring, but many shortcomings still exist.

[0003] Most existing methods for monitoring carotid artery health rely on single-modal physiological signals, such as using only Doppler ultrasound to monitor hemodynamic parameters or simply relying on photoplethysmography (PPG) to obtain basic vital signs such as heart rate. However, single-modal signals have significant limitations in reflecting the complex physiological state of the carotid artery. On the one hand, single signals are susceptible to noise interference. For example, Doppler ultrasound signals may produce spurious motions due to poor probe-skin contact or patient movement, leading to distorted blood flow waveforms; PPG signals are also affected by changes in ambient light and skin color, reducing signal quality. On the other hand, the information provided by single-modal signals is limited, making it difficult to comprehensively and accurately reflect the physiological characteristics and potential lesions of the carotid artery. For instance, relying solely on blood flow velocity information cannot comprehensively assess important parameters such as carotid artery elasticity and resistance, thus limiting the accurate assessment of carotid artery health.

[0004] In response to this, this application proposes a carotid artery health monitoring method and system based on multimodal sensor fusion to solve the above-mentioned problems. Summary of the Invention

[0005] The purpose of this invention is to provide a carotid artery health monitoring method and system based on multimodal sensor fusion to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Carotid artery health monitoring methods based on multimodal sensor fusion include:

[0008] Multimodal raw physiological signals from the carotid artery and related signs were collected to obtain the raw dataset;

[0009] The original dataset is subjected to signal quality evaluation and adaptive acquisition parameter adjustment to remove or mark low-quality segments and obtain a high-quality dataset.

[0010] The high-quality dataset is subjected to temporal denoising, segmentation, steady-state and transient separation, and multi-source feature extraction. The extracted multimodal features are then time-aligned and fused at the feature level and decision level to generate a fused feature set.

[0011] A user-personalized dynamic baseline model is constructed using the initial monitoring window of the fused feature set, and updated in real time in subsequent monitoring using a forgetting factor and incremental learning mechanism to obtain the dynamic baseline model and baseline deviation sequence.

[0012] The fused feature set and the dynamic baseline model are input into a deep learning-based time-series prediction and classification model to calculate risk prediction scores and obtain multi-level risk discrimination results based on preset and individualized thresholds.

[0013] Based on the multi-level risk assessment results, differentiated intelligent intervention and linkage strategies are executed to generate personalized intervention suggestions, graded reminder messages, and trigger corresponding linkage action sequences. At the same time, the handling feedback from users and experts is encrypted and desensitized before being sent back for subsequent model iterations to obtain intervention execution records and training feedback.

[0014] The multimodal raw physiological signals include ultrasound Doppler blood flow waveforms, photoplethysmography (PPG), body surface bioimpedance signals, neck pressure time-series signals, and basic vital signs signals synchronized with them.

[0015] The acquisition of multimodal raw physiological signals from the carotid artery and related signs to obtain a raw dataset includes: performing hard synchronization or software time synchronization on each modality signal in the multimodal raw physiological signals according to a unified time reference to ensure time alignment of each modality signal and obtain a synchronized raw dataset.

[0016] Preferably, the signal quality evaluation is based on signal-to-noise ratio, morphological consistency, contact pressure threshold, and artifact detection. It is used to calculate the quality score for each time window in the original dataset and dynamically adjust the sampling rate, gain, or trigger re-acquisition based on the quality score, thereby obtaining a high-quality dataset.

[0017] Preferably, the multi-source feature extraction includes: extracting peak systolic velocity, end-diastolic velocity, pulsatility index, waveform slope, heart rate variability, spectral energy distribution, and blood flow morphology index, and using principal component analysis, time-frequency transformation, and attention mechanism to perform weighted fusion of the features to obtain a fused feature set.

[0018] Preferably, the dynamic baseline model is represented by an adaptive time-series model with a forgetting factor, and the formula for calculating the forgetting factor α is as follows:

[0019] α = exp(−λ·Δt)

[0020] Where λ is the time decay constant and Δt is the adjacent update interval, used to achieve smooth migration of the baseline model and robust response to mutations.

[0021] Preferably, the multi-level risk assessment is based on a dual threshold strategy that combines the group threshold obtained from regional big data statistics with the user's personalized baseline deviation, in order to simultaneously satisfy both the universality of the group and the individual differences. The multi-level risk level includes at least four levels: low risk, attention, medium risk and high risk.

[0022] Preferably, the linkage strategy includes:

[0023] When the risk level is determined to be intermediate, short-term behavioral and medication adherence suggestions are automatically generated and pushed to the user.

[0024] When a high-risk condition is identified, an encrypted alert containing summary clinical information is simultaneously sent to the user, pre-set family members, and designated medical responsible parties, and a hospital emergency alert is initiated.

[0025] All uploaded processing feedback and original or derived data are encrypted and de-identified before transmission for privacy protection and for model iteration training.

[0026] Preferably, the method further includes a closed-loop model iteration step based on expert annotation: after the expert reviews the system warning and provides the handling annotation, the annotation and the corresponding data form a training sample and after verification, the time series prediction and classification model is updated incrementally to continuously improve the model's discrimination accuracy and reduce the false positive and false negative rates.

[0027] Another aspect of the present invention is to provide a carotid artery health monitoring system based on multimodal sensor fusion, comprising:

[0028] The data acquisition module is used to collect multimodal raw physiological signals from the carotid artery and related signs to obtain raw datasets.

[0029] The quality assessment and adaptive acquisition module is used to assess the signal quality of the original dataset and adjust the adaptive acquisition parameters to remove or label low-quality segments and obtain a high-quality dataset.

[0030] The preprocessing and feature fusion module is used to perform temporal denoising, segmentation, steady-state and transient separation, and multi-source feature extraction on high-quality datasets. It also performs time alignment and feature-level and decision-level fusion on the extracted multimodal features to generate a fused feature set.

[0031] The personalized dynamic baseline establishment module is used to construct a user-personalized dynamic baseline model using the initial monitoring window of the fused feature set, and to update it in real time in subsequent monitoring using a forgetting factor and incremental learning mechanism to obtain a dynamic baseline model.

[0032] The temporal prediction and multi-level risk discrimination module is used to input the fused feature set and dynamic baseline model into the deep learning-based temporal prediction and classification model to calculate the risk prediction score and obtain multi-level risk discrimination results based on preset and individualized thresholds.

[0033] The intelligent intervention and dynamic linkage module is used to execute differentiated intelligent intervention and linkage strategies based on multi-level risk assessment results. It generates personalized intervention suggestions, tiered reminder messages, and triggers corresponding linkage action sequences. At the same time, it encrypts and desensitizes the feedback from users and experts before sending it back for subsequent model iterations to obtain intervention execution records and training feedback.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] (1) This invention simultaneously acquires and fuses multimodal signals such as ultrasound Doppler, photoplethysmography (PPG), body surface bioimpedance, and neck pressure timing. This method can utilize the complementary information between each mode to correct noise and disturbances in a single mode, thereby significantly improving the accuracy and temporal robustness of hemodynamic parameter identification. The multi-source cross-validation mechanism can reduce the impact of single sensor failure or abnormality on the overall judgment and improve the reliability of the overall system.

[0036] (2) This invention establishes a dynamic baseline based on an initial monitoring window, a forgetting factor, and incremental learning. This strategy can adaptively adjust the "normal" range according to an individual's long-term physiological changes, avoiding misjudging long-term stable differences as abnormalities. This makes risk assessment more aligned with individual physiological characteristics, improving the accuracy and individual suitability of assessment in long-term monitoring scenarios. By using a deep learning model based on time-series prediction to perform short-term prediction and risk scoring on fused features, abnormal trends can be identified and graded for early warning before clinical symptoms become apparent, thus gaining a valuable time window for medical intervention. The graded early warning mechanism allows response measures to be rationally allocated according to priority and urgency, improving the timeliness of response and intervention effectiveness in acute events.

[0037] (3) In this invention, the micro spring array can adapt to the differences in neck contours and dynamic body movement scenarios of different users, continuously maintain the consistency of sensor contact, reduce the proportion of low-quality data caused by poor fit, improve the continuity and integrity of long-term monitoring, and the spring adjustment function reduces the adaptive re-acquisition frequency caused by fit problems by dynamically optimizing the sensor contact state. This reduces the power consumption of the device and the frequency of operation prompts to the user. While improving the data acquisition efficiency, it further optimizes the user experience of long-term wear. Attached Figure Description

[0038] Figure 1 This is a flowchart of the carotid artery health monitoring method based on multimodal sensor fusion according to the present invention;

[0039] Figure 2 This is a block diagram of the carotid artery health monitoring system based on multimodal sensor fusion according to the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1:

[0042] Neckband-style, full-function, multimodal wearable carotid artery monitoring and intervention system:

[0043] Please see Figure 1 and Figure 2 As shown, this embodiment uses a neck ring device to simultaneously acquire ultrasound Doppler signals, neck pressure array, body surface bioimpedance, and linked wrist PPG and basic vital signs; after signal quality assessment, adaptive acquisition, and feature fusion, a user-personalized dynamic baseline is established; a hybrid deep learning temporal prediction model is used for multi-level risk assessment; differentiated interventions and medical linkages are triggered for medium / high risk, and expert annotation feedback is used for incremental model iteration.

[0044] The hardware parameters and acquisition configuration used include:

[0045] Ultrasonic probe: High-frequency linear array probe, frequency 7.5MHz, pulse repetition frequency (PRF) for Doppler measurement is set to 4kHz, and the processed output is the instantaneous Doppler envelope (sampling rate 1kHz).

[0046] Pressure sensor array: 16×4 micro-piezoresistive dot matrix, sampling rate 200Hz, contact pressure resolution 0.1kPa, to detect whether it is in contact with the artery.

[0047] Miniature spring array: 20 micro-actuators, capable of ±5° angle fine adjustment, action response time <150ms, maximum thrust 0.8N / unit.

[0048] PPG (wrist): Dual wavelengths (850nm, 940nm), sampling rate 200Hz, ADC 16-bit.

[0049] Bioimpedance (neck): AC stimulation 50kHz, sampling rate 500Hz, impedance resolution 0.01Ω.

[0050] Synchronization and Communication: All modes achieve time alignment through a unified hardware time base (NTP-like hardware clock) or software time synchronization, and are uploaded to the mobile device via BLE5.2 or Wi-Fi (maximum data packet latency <200ms).

[0051] Power and Wearing: Battery capacity 3600mAh, typical continuous monitoring can support 30 hours (typical usage scenario); recommended device wearing time 12–20 hours / day.

[0052] Data Acquisition:

[0053] Initial monitoring period (baseline establishment): Baseline data collection for 7 days during the asymptomatic / stable period, with continuous sampling for 2 hours daily (one hour each in the morning and evening is recommended) → forming the initial dataset.

[0054] Continuous monitoring period: The device automatically collects data continuously during preset time periods every day (default: cyclical collection within 14 hours during the waking period; the sampling interval during the sleep period can be reduced to save power). The collected data includes: Doppler envelope waveform, PPG raw waveform, impedance waveform, pressure array sequence, heart rate, and respiratory rate.

[0055] Triggered resampling: When the pressure sensor array determines that the fit quality is below the threshold (quality score Qp<0.6), it automatically triggers micro-spring adjustment and executes an adaptive resampling process when the user does not respond.

[0056] Signal quality assessment and adaptive acquisition:

[0057] Signal Quality Index (SQI) Construction:

[0058] SQI = 0.4⋅SQI US +0.25⋅SQI PPG +0.2⋅SQI IMP +0.15⋅SQI PRS

[0059] Each sub-item is normalized to [0,1] based on narrowband signal-to-noise ratio, morphological matching degree, baseline drift, and artifact ratio.

[0060] Adaptive rule: If SQI < 0.65 → the system increases gain / sampling rate or activates spring fine-tuning; if SQI < 0.5 is sent three times in a row → a voice prompt is pushed to guide the user to adjust their posture.

[0061] Preprocessing, feature extraction and fusion:

[0062] Preprocessing: Doppler envelope denoising (bandpass 0.5–50Hz, downsampling from 1kHz to 200Hz for feature extraction), PPG baseline removal, Kalman filtering to eliminate motion artifacts, impedance denoising and extraction of impedance change rate.

[0063] Feature extraction:

[0064] Ultrasound: Peak systolic velocity (PSV) (cm / s), end-diastolic velocity (EDV) (cm / s), pulsatility index PI = (PSV - EDV) / mean; waveform slope, rise time.

[0065] PPG: Pulse shape index, pulse propagation time (PTT), spectral energy (0.5–4 Hz).

[0066] Impedance: Cardiac output-related estimation (ΔZ), respiratory coupling component.

[0067] Pressure array: Thermal map of contact area, mean and variance of pressure.

[0068] Feature fusion: First, perform time alignment (unify to 200Hz), then perform feature-level fusion (principal component / ICA or attention weighting) to obtain a fused feature set.

[0069] Personalized baseline establishment:

[0070] Initial baseline data window: 7 days of initial sampling were used as the baseline training samples (total sampling amount of 14 hours per person).

[0071] Dynamic update mechanism: Baseline Bt is updated over time:

[0072] Bt+1=α⋅Bt+(1−α)⋅Fnew, where the forgetting factor α=exp(−λΔt);

[0073] If we take λ = 0.05 / day, and Δt = 1 day, then α = e^-0.05 = 0.9512.

[0074] Incremental learning: When new labeled samples with expert annotations appear, incremental fine-tuning (small step size, small learning rate) is used to update the individual model.

[0075] Time series forecasting and multi-level risk assessment:

[0076] Model structure: Hybrid model – 1D-CNN (local temporal features) → two-layer LSTM (128 units) → Transformer encoder (4 heads, 2 layers) → ensemble output. The model simultaneously outputs prediction scores and uncertainty estimates (based on MCDropout).

[0077] Risk scoring and thresholds:

[0078] Risk prediction score S∈[0,1].

[0079] Low risk: S < 0.25; High risk: 0.25 ≤ S < 0.45; Medium risk: 0.45 ≤ S < 0.7; High risk: S ≥ 0.7. The thresholds can be determined based on a weighted average of regional big data and individual baselines.

[0080] Intelligent intervention and linkage:

[0081] Intermediate-risk action sequence: Push behavioral / medication compliance reminders to users, strengthen short-term monitoring (increase sampling rate), and request users to answer a systematic questionnaire (symptoms / medication).

[0082] High-risk action sequence: Simultaneously push encrypted alarm summaries (including: the PSV / PI curve of the most recent 30 minutes, trigger timestamp, and location information (if authorized)) to the user, preset family members, and medical personnel in charge, and can initiate hospital emergency warning interface calls (if authorized by the user).

[0083] Privacy: AES-256 encryption and field anonymization are performed before upload. Expert review results are used for model iteration.

[0084] AI Model Training and Testing Design (Datasets and Hyperparameters):

[0085] Cohort: N=120 subjects (age 55–78, mixed male and female), of whom 30 had radiographically confirmed carotid artery stenosis (≥50%), and the rest were controls / occasional symptom subjects.

[0086] Monitoring duration: 30 days per person on average (3600 participant-days in total); 180 abnormal events were marked by human experts (after ultrasound / clinical review).

[0087] Training set / validation / test split: 5-fold cross-validation, stratified sampling by subjects (to ensure uniform distribution of stenotic patients).

[0088] Hyperparameters: learning rate 1e-4 (Adam), batch size 128, epoch 50, early stopping patience = 8, dropout = 0.3.

[0089] Data augmentation: time scaling, noise addition, slight phase perturbation to avoid overfitting.

[0090] Evaluation metrics: Sensitivity, Specificity, average number of false alarms (24h), leadtime (minutes), mean signal quality (SQI), and user comfort.

[0091] Actual sample data obtained in Example 1:

[0092] Number of anomalies (positive examples): 180.

[0093] The model detected 148 true positives (TP) and 32 false negatives (FN). Sensitivity Sens = TP / (TP + FN) = 148 / (148 + 32) = 148 / 180 = 0.822 (82.2%).

[0094] The average number of false alarms recorded during the non-event monitoring period was 0.5 per 24 hours (the average number of false alarms was obtained from all subjects and all monitoring days).

[0095] Average early detection time (for TP): 9.0 minutes (meaning the model can provide an average warning of 9 minutes before the event is clinically confirmed).

[0096] SQI_mean (arithmetic mean of all monitoring windows according to the aforementioned weighted formula) = 0.90.

[0097] User comfort (questionnaire, N=120, Likert 1–5): average score 4.2.

[0098] As shown above, by simultaneously acquiring and fusing multimodal signals such as ultrasound Doppler, photoplethysmography (PPG), surface bioimpedance, and neck pressure timing, this method can utilize complementary information between modes to correct for noise and disturbances in a single mode, thereby significantly improving the accuracy and temporal robustness of hemodynamic parameter identification. The multi-source cross-validation mechanism can reduce the impact of single sensor failure or abnormality on the overall judgment, improving the reliability of the overall system.

[0099] The miniature spring array can adapt to the different neck contours and dynamic body movements of different users, continuously maintaining consistent sensor contact, reducing the proportion of low-quality data caused by poor fit, and improving the continuity and integrity of long-term monitoring. The spring adjustment function dynamically optimizes the sensor contact state, reducing the frequency of adaptive re-sampling caused by fit problems, which reduces device power consumption and reduces the need for frequent operation prompts to users. While improving data acquisition efficiency, it further optimizes the user experience for long-term wear.

[0100] Example 2:

[0101] Patch-type lightweight multimodal wearable system:

[0102] This embodiment, while retaining multimodal fusion (PPG, neck impedance, pressure array) and AI dynamic baseline / linkage logic, reduces or weakens the continuous ultrasound process (only retaining the micro-ultrasound Doppler array or low-power echo sampling) in exchange for longer battery life, lighter wearability, and lower cost, making it suitable for long-term continuous daily monitoring of grassroots populations.

[0103] Hardware parameters and acquisition configuration:

[0104] Ultrasound: Employs miniature low-power Doppler units (replacing traditional linear arrays) with a sampling rate of 500Hz, primarily used for supplementary flow velocity estimation (significantly reducing power consumption).

[0105] Pressure sensor array: 8×4 dot matrix, sampling rate 150Hz.

[0106] PPG: Single wavelength 940nm, sampling rate 200Hz.

[0107] Impedance: 50kHz, sampling rate: 300Hz.

[0108] Battery and battery life: 5000mAh battery (or optimized solid-state power supply), typical battery life of 72 hours (continuous monitoring mode).

[0109] Wearing comfort: The patch-type adhesive tape or lightweight neckband design gives users a slightly higher comfort rating than Example 1 (but slightly lower ultrasonic accuracy).

[0110] Data Acquisition and Baseline Construction:

[0111] Initial baseline period: 7 days, with 1.5 hours of continuous sampling per day (more in line with the usage habits of grassroots staff).

[0112] Continuous monitoring: Sampling is performed continuously during appropriate time windows (awake periods), with a reduced sampling rate at night to save power.

[0113] Synchronization: A unified clock and software time synchronization are used, with time alignment down to the 1ms level.

[0114] Signal quality and adaptation:

[0115] Use a lighter SQI weighting (weights adjusted to: PPG 0.45, impedance 0.25, pressure 0.2, microsound 0.1).

[0116] Adaptive rule: If SQI < 0.6 → prompt the user to manually adjust the wearing; the system only performs automatic hardware adjustment after multiple low-quality detections (in order to save power and reduce complexity).

[0117] Feature extraction, baseline and discrimination:

[0118] The feature set is slightly simplified (no high-precision PSV, but includes PPG pulse pattern, PTT, impedance ΔZ, and pressure contact diagram).

[0119] The baseline used the same forgetting factor mechanism (λ=0.05 / day).

[0120] Model: Lightweight hybrid model (1D-CNN + single-layer LSTM), the output is also a risk score S, but the threshold is slightly adjusted (the same risk stratification logic is retained).

[0121] Intelligent intervention and linkage:

[0122] The linkage strategy is the same as in Implementation Example 1, but when a high-risk alarm is triggered, local early warning (user / family member) is prioritized, and the decision on whether to activate linkage with medical institutions is made based on user authorization (considering the internet conditions at the grassroots level).

[0123] All data is still encrypted and anonymized.

[0124] Training and Testing (same overall experimental framework as in Example 1)

[0125] Comparisons were conducted using subsets from the same population / time window (60 people using patch-type devices and 60 people using neck rings were randomly selected from the same cohort of 120 people for A / B comparison), ensuring that each category contained a proportional distribution of stenosis patients.

[0126] The model training adopts the same 5-fold CV approach, and the hyperparameters are adjusted for the lightweight model (learning rate 2e-4, batch 64).

[0127] Test results of Example 2:

[0128] 180 abnormal events were marked (as above), TP=140, FN=40 → Sensitivity=140 / 180=0.778 (77.8%).

[0129] False alarms average 0.7 times per 24 hours.

[0130] Average early detection time: 6.0 minutes.

[0131] SQI_mean=0.86.

[0132] The average user comfort score is 4.4 (the patch is lighter).

[0133] Comparative example:

[0134] Comparison device (baseline / existing control):

[0135] To demonstrate the technical effectiveness of this invention, a commonly available single-modal PPG wearable device was selected as a control (only wrist PPG+HRV analysis, without direct neck blood flow measurement). The control device underwent comparative testing under the same cohort / time window (simulating real-world scenarios).

[0136] Compare with the equipment test values:

[0137] 180 abnormal events were marked, TP=126→Sensitivity=126 / 180=0.70(70%).

[0138] False alarms average 1.0 times / 24h.

[0139] Average early detection time: 2.5 minutes.

[0140] SQI_mean=0.72.

[0141] User comfort rating averaged 4.6 (lightest and best battery life).

[0142] Mass score and sensitivity calculation:

[0143] Signal Quality Index (SQI) calculation (for a specific time window):

[0144] The measured sub-items are: SQIUS=0.92, SQIPPG=0.88, SQIIMP=0.85, SQIPRS=0.90.

[0145] Weighted formula (weight in Example 1): SQI = 0.4⋅0.92 + 0.25⋅0.88 + 0.2⋅0.85 + 0.15⋅0.90 SQI = 0.893 → SQI ≈ 0.89 (high quality).

[0146] The statistical method for average false alarms / 24h is as follows:

[0147] The total number of false alarms (i.e., the number of times an event that was not manually labeled but was judged as "medium / high risk" by the model) was counted across all monitoring dates for all participants, resulting in the total false alarm M.

[0148] Total participant-days = 3600 (120 people × 30 days).

[0149] Average false alarms / 24h = M / 3600.

[0150] Leadtime calculation:

[0151] For each TP, calculate the difference (in minutes) between the model's first warning time and the manual / clinical event confirmation time, and take the arithmetic mean of the TP set.

[0152] Indicators / Solutions Example 1 (Neck-ring type multimodal) Example 2 (Patch Multimodal) Comparison device (single-modal PPG) Subjects & Monitoring Days 120 people, 30 days / person (total 3600 participant-days) Same as above (group comparison: 60 people each) Same as above Label abnormal events (positive examples) 180 times 180 times 180 times True positive TP was detected (number of examples) 148 140 126 Sensitivity 82.2% 77.8% 70.0% Average number of false alarms (24h) 0.50 times / 24h 0.70 times / 24h 1.00 times / 24h Average leadtime 9.0 minutes 6.0 minutes 2.5 minutes Mean signal quality (SQI) 0.90 0.86 0.72 User comfort level (1–5) 4.2 4.4 4.6 Typical battery life (continuous monitoring) 30 hours 72 hours >7 days (low power) System advantages High accuracy, multimodal cross-validation, fewer false alarms, and longer lead time Portable, long battery life, low cost, and easier to implement at the grassroots level It offers the best comfort and battery life, but has a weak ability to directly measure hemodynamics.

[0153] As can be seen from the above, the dynamic baseline establishment strategy proposed in this method, based on the initial monitoring window, forgetting factor, and incremental learning, can adaptively adjust the "normal" range according to the individual's long-term physiological changes, avoid misjudging long-term stable differences of individuals as abnormal, thereby making risk discrimination more in line with individual physiological characteristics and improving the discrimination accuracy and individual adaptability in long-term monitoring scenarios.

[0154] By using a time-series prediction-based deep learning model to perform short-term predictions and risk scoring on fused features, abnormal trends can be identified and tiered for early warning before clinical symptoms become apparent, thus gaining valuable time for medical intervention. The tiered early warning mechanism allows for the rational allocation of response measures according to priority and urgency, improving the timeliness and effectiveness of interventions in acute events.

[0155] Multimodal fusion, joint feature-level and decision-level discrimination, and a dual-threshold strategy based on population thresholds and personalized baselines can suppress false alarms while ensuring sensitivity and reducing unnecessary medical intervention requests. Combined with a multi-level linkage strategy, medical resources can be prioritized for high-risk events that truly require urgent treatment, thereby improving the efficiency of medical resource utilization and reducing the secondary burden caused by the system.

[0156] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A carotid health monitoring method based on multi-modal sensor fusion, characterized in that, The method comprises the following steps: Collecting multi-modal original physiological signals from the carotid artery and related signs to obtain an original data set; Evaluating the signal quality of the original data set and adjusting the acquisition parameters adaptively to eliminate or mark low-quality segments and obtain a high-quality data set; Performing time series denoising, segmentation, steady-state and transient separation, and multi-source feature extraction on the high-quality data set, and performing time alignment and feature-level and decision-level fusion on the extracted multi-modal features to generate a fusion feature set; Using the initial monitoring window of the fusion feature set to construct a user's personalized dynamic baseline model, and updating it in real time with a forgetting factor and an incremental learning mechanism in subsequent monitoring to obtain a dynamic baseline model; Inputting the fusion feature set and the dynamic baseline model into a deep learning-based time series prediction and classification model to calculate a risk prediction score and divide it into multiple risk levels according to pre-set and individualized thresholds; According to the multi-level risk discrimination results, differential intelligent intervention and linkage strategy is executed to generate individualized intervention suggestions, hierarchical reminder messages and trigger corresponding linkage action sequences, while the user's and expert's treatment feedback is encrypted and desensitized before being returned for subsequent model iteration to obtain intervention execution records and training feedback.

2. The carotid health monitoring method based on multi-modal sensor fusion according to claim 1, characterized in that, The multi-modal original physiological signals include ultrasonic Doppler blood flow waveform, photoplethysmography (PPG), body surface bioimpedance signal, neck pressure time series signal, and basic vital sign signal synchronized with them in time; The multi-modal original physiological signals from the carotid artery and related signs are collected to obtain an original data set, which includes hard synchronization or software time correction of each modality signal in the multi-modal original physiological signals according to a unified time reference, to ensure the time alignment of each modality signal and obtain a synchronized original data set.

3. The multi-modal sensor fusion based carotid health monitoring method as claimed in claim 1, wherein, The signal quality evaluation is based on signal-to-noise ratio, shape consistency, contact pressure threshold, and artifact detection to calculate the quality score of each time window in the original data set and dynamically adjust the sampling rate, gain, or trigger reacquisition according to the quality score, thereby obtaining a high-quality data set.

4. The multi-modal sensor fusion based carotid health monitoring method as claimed in claim 1, wherein, The multi-source feature extraction includes extracting systolic peak flow rate, diastolic end flow rate, beat index, waveform slope, heart rate variability, spectral energy distribution, and blood flow morphology index, and using principal component analysis, time-frequency transformation, and attention mechanism to weight and fuse the features to obtain a fusion feature set.

5. The multi-modal sensor fusion based carotid health monitoring method as claimed in claim 1, wherein, The dynamic baseline model is represented by an adaptive time series model with a forgetting factor, and the calculation formula of the forgetting factor α is: α = exp(-λ·Δt) Where λ is the time series decay constant, and Δt is the adjacent update interval, which is used to realize the smooth migration of the baseline model and the robust response to mutations.

6. The multi-modal sensor fusion based carotid health monitoring method as claimed in claim 1, wherein, The multi-level risk discrimination is based on a dual threshold strategy combining the group threshold obtained from regional big data statistics with the user's personalized baseline deviation, which can meet the group universality and individual difference, and the multi-level risk levels include at least low risk, attention, medium risk, and high risk.

7. The multi-modal sensor fusion based carotid health monitoring method as claimed in claim 1, wherein, The linkage strategy includes: When it is determined to be medium risk, automatically generate short-term behavior and drug compliance suggestions and push them to the user; When determined as high risk, the user, the preset family members and the designated medical responsibility party are pushed an encrypted alert containing summary clinical information and an emergency warning is initiated.

8. The multi-modal sensor fusion based carotid health monitoring method as claimed in claim 1, wherein, The method further comprises an expert-labeled closed-loop model iteration step: after the expert reviews the system warning and provides treatment labeling, the labeling and corresponding data are used to form a training sample, which is verified and then updated in an incremental manner to the time series prediction and classification model, so as to continuously improve the discrimination accuracy of the model and reduce the false positive and false negative rates.

9. A carotid health monitoring system based on multi-modal sensor fusion, characterized in that, Comprise: A data acquisition module for acquiring multi-modal original physiological signals from the carotid artery and related signs to obtain an original data set; A quality evaluation and adaptive acquisition module for signal quality evaluation and adaptive acquisition parameter adjustment of the original data set to eliminate or mark low-quality segments and obtain a high-quality data set; A preprocessing and feature fusion module for time series denoising, segmentation, steady-state, transient-state separation and multi-source feature extraction of the high-quality data set, and time alignment and feature-level and decision-level fusion of the extracted multi-modal features to generate a fusion feature set; An individualized dynamic baseline establishment module for constructing a user individualized dynamic baseline model using the initial monitoring window of the fusion feature set, and updating in real time with a forgetting factor and an incremental learning mechanism in subsequent monitoring to obtain a dynamic baseline model; A time series prediction and multi-level risk discrimination module for inputting the fusion feature set and the dynamic baseline model into a deep learning-based time series prediction and classification model to calculate a risk prediction score and obtain a multi-level risk discrimination result according to a preset and individualized threshold; An intelligent intervention and dynamic linkage module for executing differentiated intelligent intervention and linkage strategies according to the multi-level risk discrimination result, generating individualized intervention suggestions, hierarchical reminder messages and triggering corresponding linkage action sequences, and transmitting the treatment feedback of the user and the expert after encryption and desensitization for subsequent model iteration to obtain intervention execution records and training feedback. An intelligent intervention and dynamic linkage module for executing differentiated intelligent intervention and linkage strategies according to the multi-level risk discrimination result, generating individualized intervention suggestions, hierarchical reminder messages and triggering corresponding linkage action sequences, and transmitting the treatment feedback of the user and the expert after encryption and desensitization for subsequent model iteration to obtain intervention execution records and training feedback.

Images