Blood pressure measurement error self-correction method and system based on AI voice interaction

Abstract

The invention discloses a blood pressure measurement error self-correction method and system based on AI voice interaction, and the method comprises the steps: synchronously obtaining a user voice signal and blood pressure original data through a multi-channel voice collection module and a distributed blood pressure sensing array, and extracting related voice feature parameters through a multi-sensor voice collaborative filtering algorithm; inputting the data and blood pressure original data into a blood pressure measurement deviation learning model to obtain an initial deviation estimation value, generating a dynamic compensation coefficient through an ambulatory blood pressure error compensation model in combination with physiological and environmental parameters, and fusing multi-dimensional data by a multi-modal blood pressure correction analysis engine to execute error correction operation; finally, a closed-loop feedback mechanism is constructed through an error self-correction module based on AI voice interaction, and real-time self-correction of errors in the measurement process is achieved. According to the system, association between voice and blood pressure data is fully mined through multi-unit collaborative operation, the accuracy and adaptability of blood pressure measurement are improved, and the system is suitable for precise blood pressure monitoring in a non-professional scene.

Description

Technical Field

[0001] This invention relates to the field of blood pressure measurement error correction technology, and in particular to a blood pressure measurement error self-correction method and system based on AI voice interaction. Background Technology

[0002] The accuracy of blood pressure measurement is directly related to the reliability of health monitoring and disease diagnosis. With the widespread use of home blood pressure measurement devices, measurement errors in non-professional scenarios are becoming increasingly prominent. The physiological information contained in voice signals has not been fully explored. Traditional blood pressure measurement systems mostly rely on single-sensor data acquisition and lack dynamic adjustment mechanisms based on user interaction and feedback, making it difficult to cope with errors caused by complex factors such as individual physiological differences and environmental interference. Therefore, there is an urgent need to construct an error self-correction scheme that integrates multi-source information and intelligent algorithms. Through the synergy of AI voice interaction technology and multimodal data processing, real-time error correction during blood pressure measurement can be achieved, meeting the needs for precise and convenient health monitoring.

[0003] Existing technologies have two significant drawbacks: First, traditional blood pressure measurement methods do not effectively integrate the correlation between voice features and blood pressure data, relying solely on single sensor data for measurement and calculation. This fails to utilize potential information such as vascular status and heart rate changes reflected in the voice signal, resulting in insufficient adaptability to individual physiological differences and difficulty in eliminating systematic biases caused by variations in vascular elasticity and circulatory status. Second, existing error correction mechanisms lack dynamic adaptive capabilities, often employing fixed compensation coefficients or simple linear correction models. They do not fully consider dynamic changes in the measurement environment and real-time feedback adjustments during the measurement process, making them unable to cope with the dynamic influence of environmental factors such as temperature, humidity, and electromagnetic interference. Furthermore, the lack of closed-loop feedback self-correction logic makes it difficult to optimize correction parameters in real time during continuous measurement, thus limiting the timeliness and accuracy of error correction. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a method and system for self-correction of blood pressure measurement errors based on AI voice interaction.

[0005] The technical solution adopted in this invention is a blood pressure measurement error self-correction method based on AI voice interaction, comprising the following steps: S1, synchronously acquiring user voice signals and raw blood pressure data through a multi-channel voice acquisition module and a distributed blood pressure sensor array to establish a spatiotemporally aligned data acquisition link; S2, using a multi-sensor voice collaborative filtering algorithm to suppress noise in the mixed interference voice signal and extract voice feature parameters related to blood pressure changes; S3, inputting the filtered voice feature parameters and raw blood pressure data into a blood pressure measurement deviation learning model to mine the nonlinear mapping relationship between the two and output an initial deviation estimate; S4, based on user physiological feature parameters and measurement environment parameters, adaptively adjusting the initial deviation estimate through a dynamic blood pressure error compensation model to generate a dynamic compensation coefficient; S5, inputting the dynamic compensation coefficient into a multimodal blood pressure correction analysis engine to integrate voice feature feedback and blood pressure data trend analysis and perform multi-dimensional error correction calculations; S6, receiving the correction results through an AI voice interaction-based blood pressure measurement error self-correction module, constructing a closed-loop feedback mechanism, and performing real-time self-correction of errors during the measurement process.

[0006] Furthermore, the expression for the multi-sensor speech collaborative filtering algorithm is as follows: ,in: This is the filtered speech signal. The original voice signal collected by the i-th sensor. For spatiotemporal adaptive weighting coefficients, As the sensor trust factor, The gradient value of the speech feature. The interference suppression coefficient is... It is the minimum constant. For the number of sensors, For frequency parameters, This is a time parameter.

[0007] Furthermore, the expression for the blood pressure measurement bias learning model is: ,in, This is the initial deviation estimate. For the model weight vector, This is the coupled feature matrix of speech features and blood pressure data. For the set of speech feature parameters, This is a collection of raw blood pressure data. This is a matrix of physiological and environmental influencing factors. This is the original blood pressure measurement value. A set of environmental parameters This is the deviation adjustment coefficient. It is a nonlinear mapping function. For measuring time windows, This is for the Hadamard product operation.

[0008] Furthermore, the expression for the dynamic blood pressure error compensation model is: ,in, For dynamic compensation coefficients, As a proportional adjustment factor, The function is influenced by physiological characteristics. A set of user physiological characteristic parameters For heart rate-related parameters, It is a secondary compensation factor. As a dynamic environmental factor, For ambient temperature parameters, For ambient humidity parameters, For constraint coefficients, For the k-th state influence factor, The number of state factors. The derivative adjustment coefficient is... This represents the rate of change of deviation.

[0009] Furthermore, the correction calculation expression of the multimodal blood pressure correction analysis engine is as follows: ,in, This is the corrected blood pressure value. As a voice-blood pressure co-correction factor, Let q be the q-th multimodal influence parameter. To affect the number of parameters, To correct the threshold, The activation coefficient, It is a non-linear activation function.

[0010] Furthermore, the feedback adjustment expression of the blood pressure measurement error self-correction module based on AI voice interaction is as follows: This is the estimated value for the next round of deviation. For feedback coefficients, For reference blood pressure values, These are newly acquired speech feature parameters. For feedback constraint constants, For the iterative decay factor, This is the estimated deviation value from the previous round. This represents the number of iterations.

[0011] Further, step S3 includes the following sub-steps: S31, the filtered speech feature parameters are split into multi-dimensional features according to the frequency domain, time domain, and Mel-frequency cepstral domain to form a structured feature set, and the original blood pressure data is processed by time-series segmentation to obtain data segments of equal duration; S32, the structured feature set and the blood pressure data segments are paired one-to-one to construct a training sample set including input features and output labels, wherein the input features are a combination vector of multi-dimensional speech features and original blood pressure data, and the output labels are preset deviation reference values; S33, the training sample set is divided into a training set, a validation set, and a test set according to a preset ratio, the blood pressure measurement deviation learning model is iteratively trained using the training set, the model hyperparameters are adjusted in real time using the validation set, and the deviation estimation accuracy of the model is verified using the test set; S34, based on the trained model, the real-time collected speech feature parameters and original blood pressure data are input, and the initial deviation estimate value corresponding to the measurement scenario is output through nonlinear transformation operations within the model.

[0012] Further, S4 includes the following sub-steps: S41, collecting physiological characteristic parameters of the user such as age, weight, and vascular elasticity coefficient, as well as environmental parameters of the measurement environment such as temperature, humidity, and electromagnetic interference intensity; quantifying and encoding various parameters to form a standardized parameter vector; S42, inputting the standardized physiological characteristic parameters and environmental parameters into the parameter adaptation layer of the dynamic blood pressure error compensation model; generating a parameter weight matrix that matches the initial deviation estimate through feature mapping operations within the layer; S43, adjusting the initial deviation estimate component by component based on the parameter weight matrix; updating the adjustment range in real time according to the parameter change trend through the model's built-in dynamic adjustment mechanism to eliminate the error influence caused by physiological differences and environmental changes; S44, merging the adjustment results of each component into a unified dynamic compensation coefficient through the integration operation of the model output layer to ensure the adaptability of the compensation coefficient to the current measurement scenario.

[0013] Further, S5 includes the following sub-steps: S51, inputting the dynamic compensation coefficient, voice feature parameters, raw blood pressure data, and initial deviation estimate into the feature fusion module of the multimodal blood pressure correction analysis engine, and allocating the weight ratio of each input data through an attention mechanism to strengthen the representation of calibration information; S52, at the engine's computation layer, performing preliminary correction calculations on the raw blood pressure data based on preset multi-dimensional correction rules and combined with the dynamic compensation coefficient, while constructing correction constraints using voice feature feedback information; S53, extracting the time-series trend of the preliminarily corrected blood pressure data through the engine's trend analysis module, comparing the change patterns of historical correction data, identifying abnormal correction results, and correcting them; S54, inputting the corrected correction results into the engine's output calibration module, and generating the final multi-dimensional error correction result after accuracy verification and format standardization processing.

[0014] A blood pressure measurement error self-correction system based on AI voice interaction is disclosed. This system, applied to a blood pressure measurement error self-correction method based on AI voice interaction, includes: a multi-source data synchronous acquisition unit, used to establish a spatiotemporal alignment link with a distributed blood pressure sensing submodule through a multi-channel voice acquisition submodule, synchronously acquiring user voice signals and raw blood pressure data, and transmitting the data to a signal preprocessing unit; a signal preprocessing unit, communicatively connected to the multi-source data synchronous acquisition unit, incorporating a multi-sensor voice collaborative filtering algorithm module to perform noise suppression processing on the voice signal and extract voice feature parameters, transmitting the processed voice feature parameters and raw blood pressure data to a deviation estimation unit; and a deviation estimation unit, communicatively connected to both the signal preprocessing unit and the compensation adjustment unit, equipped with a blood pressure measurement deviation learning model to mine... The nonlinear mapping relationship between sound features and blood pressure data outputs an initial deviation estimate to the compensation adjustment unit. The compensation adjustment unit, which communicates with the deviation estimation unit and the correction analysis unit, generates dynamic compensation coefficients based on physiological features and environmental parameters through a dynamic blood pressure error compensation model, and transmits the compensation coefficients and the initial deviation value to the correction analysis unit. The correction analysis unit communicates with the multimodal blood pressure correction analysis engine, the compensation adjustment unit, and the feedback correction unit, integrates multi-dimensional data to perform error correction calculations, and outputs the correction results to the feedback correction unit. The feedback correction unit forms a closed-loop connection with the correction analysis unit and the multi-source data synchronous acquisition unit, receives the correction results through an AI voice interaction-based blood pressure measurement error self-correction module, and provides real-time feedback on the data acquisition and correction process to perform dynamic self-correction of errors.

[0015] Beneficial Effects: This invention proposes a blood pressure measurement error self-correction method and system based on AI voice interaction. Through multi-channel voice acquisition and synchronous data acquisition from distributed blood pressure sensors, combined with a multi-sensor voice collaborative filtering algorithm, it accurately extracts voice features related to blood pressure, breaking the limitations of traditional single-sensor data measurement. It fully explores the inherent correlation between voice signals and blood pressure data, overcoming the shortcomings of existing technologies that fail to integrate such correlation information and are insufficiently adapted to individual physiological differences. An initial deviation estimate is obtained by mining nonlinear mapping relationships through a blood pressure measurement deviation learning model. Then, an adaptive compensation coefficient is generated by combining physiological characteristics and environmental parameters through a dynamic blood pressure error compensation model, replacing the traditional fixed compensation or simple linear correction mode. This effectively addresses the error impact caused by dynamic environmental changes and individual physiological differences. A multimodal blood pressure correction analysis engine integrates multi-source data to perform multi-dimensional correction calculations. Combined with a closed-loop feedback mechanism based on AI voice interaction, it achieves real-time self-correction of errors during measurement, solving the problems of lack of dynamic adaptive capability and closed-loop feedback logic, as well as limitations in correction timeliness and accuracy in existing technologies. This significantly improves the accuracy and stability of blood pressure measurement, providing reliable technical support for accurate blood pressure monitoring in non-professional scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall steps of the method of the present invention; Figure 2 This is a flowchart of method step S3 of the present invention; Figure 3 This is a flowchart of method step S4 of the present invention; Figure 4 This is a flowchart of step S5 of the method of the present invention; Figure 5 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, a blood pressure measurement error self-correction method based on AI voice interaction includes the following steps: S1, through the multi-channel voice acquisition module and the distributed blood pressure sensor array, synchronously acquires user voice signals and raw blood pressure data, and establishes a spatiotemporally aligned data acquisition link; Specifically, step S1 establishes a spatiotemporally aligned data acquisition link between a multi-channel voice acquisition module consisting of eight evenly distributed voice acquisition channels and a distributed blood pressure sensor array comprising 12 sensor nodes, enabling synchronous acquisition of user voice signals and raw blood pressure data. The voice acquisition channels employ a 16kHz sampling rate and 24-bit quantization precision, with each channel spaced 15 cm apart, covering a 360-degree area around the user's head to ensure comprehensive voice signal capture. The distributed blood pressure sensor array has a 2 cm spacing between its sensor nodes, evenly fitted to the inside of the measuring cuff, and uses a 500Hz sampling frequency to acquire blood pressure fluctuation data. Through GPS timestamp synchronization technology, the acquisition time deviation between the voice acquisition channels and the sensor nodes is controlled within 10 microseconds, establishing a spatiotemporal alignment mechanism to ensure that the acquisition time point of each set of voice signals strictly corresponds to the acquisition time point of the raw blood pressure data. Meanwhile, this step uses a data caching module to store the synchronously collected data in real time. The cache capacity is set to 16GB, which supports continuous data collection for 4 hours. During the collection process, the integrity of the data is verified in real time through a check code mechanism to ensure that the transmission error rate of each frame of data is less than 0.001%, providing basic data support with strong spatiotemporal consistency and high integrity for subsequent filtering, deviation estimation and correction calculations.

[0019] S2, a multi-sensor speech collaborative filtering algorithm is used to suppress noise in the speech signal with mixed interference and extract speech feature parameters related to blood pressure changes; Specifically, step S2 employs a multi-sensor speech collaborative filtering algorithm to suppress noise in speech signals containing mixed environmental noise, electromagnetic interference, and human physiological noise, accurately extracting speech feature parameters related to blood pressure changes. This algorithm is based on synchronous data from eight speech acquisition channels. Through a spatiotemporal adaptive weight allocation mechanism, it assigns a dynamic weight coefficient of 0.1 to 0.9 to the original speech signal of each channel. The weight coefficient is adjusted in real time according to the signal-to-noise ratio (SNR) of each channel. Channels with an SNR higher than 30dB are assigned a weight no less than 0.7, and channels with an SNR lower than 15dB are assigned a weight no greater than 0.3. The algorithm evaluates the reliability of the data from each channel using a sensor trust factor, with a value ranging from 0.05 to 0.95. It combines speech feature gradient values ​​to enhance effective signal components and simultaneously uses an interference suppression coefficient to specifically suppress environmental noise. The interference suppression coefficient is dynamically adjusted according to the noise frequency characteristics; the suppression coefficient for 50Hz power frequency interference is set to 0.92, and the suppression coefficient for high-frequency noise above 1kHz is set to 0.85. After filtering, the signal-to-noise ratio of the speech signal is improved to over 45dB, and the noise suppression ratio reaches 35dB. Subsequently, multi-dimensional feature parameters in the frequency domain, time domain, and Mel-Cepstral domain are extracted from the filtered speech signal, including 128-dimensional Mel-Cepstral coefficients, 24-dimensional time-domain zero-crossing rate, and 32-dimensional frequency-domain power spectral density, totaling 184-dimensional speech feature parameters. The extraction time of all feature parameters is controlled within 20 milliseconds to ensure that real-time processing requirements are met.

[0020] S3: Input the filtered speech feature parameters and the original blood pressure data into the blood pressure measurement deviation learning model, explore the nonlinear mapping relationship between the two, and output the initial deviation estimate. Specifically, step S3 inputs the 184-dimensional speech feature parameters extracted in step S2 and the raw blood pressure data into the blood pressure measurement deviation learning model. Through deep training of the model, the nonlinear mapping relationship between the two is mined, and an initial deviation estimate is output. This model employs a 3-layer hidden neural network architecture. The input layer has 200 nodes (including 184-dimensional speech feature parameters and 16-dimensional statistical features of the raw blood pressure data), the first hidden layer has 128 nodes, the second has 64, the third has 32, and the output layer has 1 node, corresponding to the initial deviation estimate. During model training, a training dataset of 100,000 labeled samples is constructed. The samples include user data from different ages (18-80 years old, in 5-year intervals), weights (45-120 kg, in 10 kg intervals), and blood pressure levels (systolic blood pressure 90-180 mmHg, diastolic blood pressure 60-110 mmHg, in 10 mmHg intervals), as well as samples collected under different environments (temperature 10-35℃, humidity 30%-85%). During training, batch gradient descent was employed with a batch size of 64 and an initial learning rate of 0.001, decreasing by 0.1 every 100 iterations for a total of 1000 iterations. The mean squared error of the trained model was controlled within 0.8, and the coefficient of determination (R²) was no less than 0.95. In real-time operation, the model received the speech feature parameters and raw blood pressure data from the current acquisition period. After a 25-millisecond computation period, it output an initial deviation estimate, with the absolute error of the estimate controlled within 2 mmHg, providing accurate deviation data for subsequent dynamic compensation.

[0021] S4, based on user physiological characteristic parameters and measurement environment parameters, adaptively adjusts the initial deviation estimate through a dynamic blood pressure error compensation model to generate dynamic compensation coefficients; Specifically, step S4, based on the user's eight physiological characteristic parameters and six measurement environment parameters, adaptively adjusts the initial deviation estimate output in step S3 using a dynamic blood pressure error compensation model to generate dynamic compensation coefficients. The physiological characteristic parameters include age, weight, height, vascular elasticity coefficient, heart rate, historical mean blood pressure, body mass index, and blood viscosity. Each parameter is quantized and encoded into an 8-dimensional standardized vector, with values ​​mapped to the 0-1 range. The measurement environment parameters include temperature, humidity, air pressure, electromagnetic interference intensity, light intensity, and wind speed, which are quantized and encoded into a 6-dimensional standardized vector, also mapped to the 0-1 range. The model uses a parameter adaptation layer to perform feature mapping on the 14-dimensional standardized parameters, generating a 32-dimensional parameter weight matrix. Each element of the weight matrix ranges from 0.01 to 0.99 and is dynamically adjusted based on the parameter's sensitivity to deviation. During the adjustment process, the model optimizes the initial bias estimate component by component, with the adjustment range of each component updated every 5 milliseconds based on the parameter change trend. The adjustment step size is set to a dynamic range of 0.005 to 0.05 to ensure rapid response to physiological differences and environmental changes. Finally, through weighted summation of the output layer, the adjustment results of each component are merged into a single dynamic compensation coefficient. The compensation coefficient ranges from 0.8 to 1.2. When the environmental change rate exceeds 5% / minute or the deviation of physiological parameters from historical averages exceeds 10%, the adjustment range of the compensation coefficient is increased by 30%, ensuring that the compensation result is highly adapted to the current measurement scenario.

[0022] S5 inputs the dynamic compensation coefficient into the multimodal blood pressure correction analysis engine, integrates voice feature feedback and blood pressure data trend analysis, and performs multi-dimensional error correction calculation; Specifically, step S5 inputs the dynamic compensation coefficient generated in step S4 into the multimodal blood pressure correction analysis engine, integrates voice feature feedback and blood pressure data trend analysis, and performs multi-dimensional error correction calculations. This engine includes four functional modules: feature fusion, computational core, trend analysis, and output calibration. The feature fusion module assigns weights to the dynamic compensation coefficient, 184-dimensional voice feature parameters, 16-dimensional blood pressure raw data statistical features, and initial deviation estimates using an attention mechanism. The weight allocation range for voice feature feedback is 0.3 to 0.6, for blood pressure data trends it is 0.2 to 0.4, and for the dynamic compensation coefficient it is 0.1 to 0.3, thus strengthening the representation of key information through weight optimization. The computational core layer incorporates 12 multi-dimensional correction rules, which, combined with the dynamic compensation coefficient, perform preliminary correction on the raw blood pressure data. Simultaneously, it utilizes the frequency change trends in voice features to construct three correction constraints. The thresholds of these constraints are dynamically adjusted based on historical correction data to ensure the rationality of the preliminary correction results. The trend analysis module extracts the time-series trend of the initially corrected blood pressure data over 10 consecutive acquisition cycles. By comparing the changing patterns of the past 100 sets of historical corrected data, a sliding window algorithm (window size set to 20) is used to identify abnormal correction results. The abnormal identification threshold is set to 3 times the standard deviation, and the identified abnormal results are corrected in reverse. The output calibration module verifies the accuracy of the corrected results. The standard for passing the verification is that the deviation between the corrected result and the reference value is less than 1.5 mmHg. After passing the verification, the format is standardized to generate the final correction result, which includes systolic blood pressure, diastolic blood pressure, and correction confidence (value range 0.8 to 1.0). The entire correction calculation process takes less than 30 milliseconds, meeting the timeliness requirements of real-time measurement.

[0023] The S6 receives correction results through an AI-based voice interaction-based blood pressure measurement error self-correction module, constructs a closed-loop feedback mechanism, and performs real-time self-correction of errors during the measurement process.

[0024] Specifically, step S6 receives the correction result from step S5 through the AI-based voice interaction-based blood pressure measurement error self-correction module, constructing a closed-loop feedback mechanism to achieve real-time self-correction of errors during the measurement process. This module includes a voice interaction unit, a feedback adjustment unit, and a parameter update unit. The voice interaction unit uses natural language processing technology, receiving user voice feedback information (such as adjustments to measurement posture, changes in body state, etc.) through eight voice acquisition channels. The voice recognition accuracy is no less than 98%, and the recognition response time is controlled within 50 milliseconds, converting the voice feedback into 6-dimensional control parameters. The feedback adjustment unit dynamically adjusts the key parameters for the next round of measurement based on the deviation between the correction result and the reference blood pressure value, combined with the control parameters generated by the voice feedback. The adjusted parameters include the weighting coefficients of the voice acquisition channels, the sampling frequency of the blood pressure sensor array, the initial weights of the deviation learning model, and the adjustment step size of the compensation model. The parameter update unit transmits the adjusted parameters in real-time to the multi-source data synchronous acquisition unit, signal preprocessing unit, deviation estimation unit, and compensation adjustment unit, achieving closed-loop parameter updates throughout the entire process. The update frequency is consistent with the measurement cycle, i.e., once every 2 seconds. When the deviation between the calibration result and the reference value exceeds 2 mmHg, the intensity of the feedback adjustment increases by 40%, and the voice interaction unit actively outputs a prompt (1 second long) to guide the user to adjust the measurement status. When the calibration deviation for three consecutive measurement cycles is less than 1 mmHg, the intensity of the feedback adjustment decreases by 20% to maintain parameter stability. This closed-loop feedback mechanism ensures that the error correction for each measurement round is based on the calibration result of the previous round and the user's real-time feedback, keeping the response delay of error self-correction within 100 milliseconds and the error accumulation rate during continuous measurement less than 0.5% / hour, significantly improving the long-term stability and accuracy of blood pressure measurement.

[0025] Preferably, the expression for the multi-sensor speech collaborative filtering algorithm is: ,in: This is the filtered speech signal. The original voice signal collected by the i-th sensor. For spatiotemporal adaptive weighting coefficients, As the sensor trust factor, The gradient value of the speech feature. The interference suppression coefficient is... It is the minimum constant. For the number of sensors, For frequency parameters, This is a time parameter.

[0026] Specifically, the multi-sensor speech collaborative filtering algorithm integrates raw speech signals collected by multiple sensors to achieve noise suppression and effective feature extraction. During implementation, eight sensors are used, covering a 360-degree acquisition range around the user to ensure comprehensive capture of speech information. During algorithm operation, the spatiotemporal adaptive weight coefficients are dynamically adjusted based on the signal-to-noise ratio (SNR) of the real-time signals collected by each sensor, ranging from 0.1 to 0.9. Sensors with an SNR higher than 30dB are assigned a weight no less than 0.7, while sensors with an SNR lower than 15dB are assigned a weight no greater than 0.3, to highlight the contribution of highly reliable sensor data. The sensor trust factor ranges from 0.05 to 0.95, combining speech feature gradient values ​​to enhance speech signal components related to blood pressure changes. Simultaneously, environmental noise is specifically suppressed using interference suppression coefficients: a suppression coefficient of 0.92 is set for 50Hz power frequency interference, and a suppression coefficient of 0.85 is set for high-frequency noise above 1kHz. The minimum constant is set to 1e-6 to avoid a zero denominator. This algorithm constructs a two-dimensional spatiotemporal processing dimension using a time parameter t (sampling interval of 10 microseconds) and a frequency parameter f (range of 20Hz-8kHz). It performs frame-by-frame processing on the raw speech signal of each sensor, with the processing time of a single frame controlled within 2 milliseconds. After processing, the signal-to-noise ratio of the speech signal is improved to over 45dB, and the noise suppression ratio reaches 35dB. This lays the foundation for subsequent extraction of accurate speech feature parameters related to blood pressure and ensures that the physiological state-related information contained in the speech signal is not destroyed by noise interference.

[0027] Preferably, the expression for the blood pressure measurement bias learning model is: ,in, This is the initial deviation estimate. For the model weight vector, This is the coupled feature matrix of speech features and blood pressure data. For the set of speech feature parameters, This is a collection of raw blood pressure data. This is a matrix of physiological and environmental influencing factors. This is the original blood pressure measurement value. A set of environmental parameters This is the deviation adjustment coefficient. It is a nonlinear mapping function. For measuring time windows, This is for the Hadamard product operation.

[0028] Specifically, the blood pressure measurement bias learning model is used to mine the nonlinear mapping relationship between speech feature parameters and raw blood pressure data to output an initial bias estimate. During the model training phase, a dataset of 100,000 labeled samples is constructed. These samples include different age ranges (18-80 years), different weight ranges (45-120 kg), and blood pressure ranges (90-180 mmHg systolic blood pressure, 60-110 mmHg diastolic blood pressure). The dataset also includes data on different environmental conditions (10-35℃ temperature, 30%-85% humidity). The model weight vector is set to 128 dimensions and iteratively optimized using a gradient descent algorithm. The initial learning rate is 0.001, decreasing by 0.1 every 100 iterations. After 1000 iterations, the mean squared error of the model is controlled within 0.8. The coupling feature matrix between voice features and blood pressure data is 184×16-dimensional (corresponding to 184-dimensional voice features and 16-dimensional statistical features of raw blood pressure data). The physiological-environmental influencing factor matrix is ​​16×14-dimensional (including 8 physiological features and 6 environmental parameters). The bias adjustment coefficient is set to 0.05, and the measurement time window is set to 5 seconds to ensure the capture of short-term blood pressure fluctuations. The nonlinear mapping function uses the ReLU activation function. By integrating the signal within the 5-second time window, the dynamic correlation between voice features and blood pressure data is captured in real time. The model computation time is controlled within 25 milliseconds, and the absolute error of the initial bias estimate does not exceed 2 mmHg, providing an accurate bias basis for subsequent dynamic compensation and effectively avoiding measurement bias caused by individual physiological differences.

[0029] Preferably, the expression for the dynamic blood pressure error compensation model is: ,in, For dynamic compensation coefficients, As a proportional adjustment factor, The function is influenced by physiological characteristics. A set of user physiological characteristic parameters For heart rate-related parameters, It is a secondary compensation factor. As a dynamic environmental factor, For ambient temperature parameters, For ambient humidity parameters, For constraint coefficients, For the k-th state influence factor, The number of state factors. The derivative adjustment coefficient is... This represents the rate of change of deviation.

[0030] Specifically, the dynamic blood pressure error compensation model adaptively adjusts the initial deviation estimate by combining user physiological characteristics and measurement environment parameters to generate dynamic compensation coefficients. The model input includes eight physiological characteristic parameters (age, weight, height, vascular elasticity coefficient, etc.) and six environmental parameters (temperature, humidity, air pressure, etc.). All parameters are quantized and mapped to a standardized vector within the 0-1 range. The proportional adjustment factor is set to 0.7, the quadratic compensation factor is set to 0.3, and the physiological characteristic influence function is constructed based on the user's vascular elasticity coefficient and heart rate correlation parameters. The heart rate correlation parameter sampling frequency is 500Hz to capture the impact of heart rate fluctuations on blood pressure measurement in real time. The environmental dynamic factors are constructed by combining temperature and humidity parameters. The temperature parameter acquisition accuracy is ±0.1℃, the humidity parameter acquisition accuracy is ±1%, the constraint coefficient is set to 0.5, the number of state factors is set to 12 (including key physiological and environmental influence dimensions), and the derivative adjustment coefficient is set to 0.2. The model senses the dynamic changes in deviation in real time by calculating the rate of change of deviation (sampling interval of 50 milliseconds). When the rate of change of the environment exceeds 5% / minute or the deviation of physiological parameters from the historical average exceeds 10%, the adjustment range of the compensation coefficient is automatically increased by 30%. The dynamic compensation coefficient ranges from 0.8 to 1.2, and the calculation time for a single round is controlled within 15 milliseconds, ensuring that the compensation coefficient can quickly adapt to different measurement scenarios, effectively offset the errors caused by physiological differences and environmental changes, and improve the pertinence and accuracy of deviation correction.

[0031] Preferably, the correction calculation expression of the multimodal blood pressure correction analysis engine is: ,in, This is the corrected blood pressure value. As a voice-blood pressure co-correction factor, Let q be the q-th multimodal influence parameter. To affect the number of parameters, To correct the threshold, The activation coefficient, It is a non-linear activation function.

[0032] Specifically, the multimodal blood pressure correction analysis engine integrates multi-dimensional data to perform error correction and output accurate blood pressure values. During engine operation, the voice-blood pressure co-correction factor is constructed based on the correlation between 184-dimensional voice feature parameters and the original blood pressure measurement, with a value range of 0.5-1.0; the higher the correlation, the closer the factor value is to 1.0. The number of multimodal influence parameters is set to 12, including key dimensions such as environmental interference, physiological differences, and equipment status. The weight of each parameter is dynamically allocated through an attention mechanism, ranging from 0.05 to 0.2, and the correction threshold is set to 0.1 to avoid over-correction. The activation coefficient is set to 0.8, and a hyperbolic tangent nonlinear activation function is used to nonlinearly transform the partial derivatives of the coupling feature matrix of voice features and blood pressure data, enhancing effective correction information. During engine operation, the original blood pressure data is first initially corrected based on the dynamic compensation coefficient. Then, three constraints (frequency consistency, amplitude stability, and temporal correlation) are constructed in conjunction with voice feature feedback. A sliding window algorithm (window size 20 frames) is used to extract temporal trends, and outliers are identified and corrected by comparing with 100 historical correction data sets. The entire calibration process is divided into five stages: feature fusion, preliminary calibration, trend analysis, anomaly correction, and accuracy verification. The total time is controlled within 30 milliseconds. The deviation between the calibrated blood pressure value and the reference value is less than 1.5 mmHg, ensuring the collaborative calibration effect of multimodal data, giving full play to the complementary advantages of voice features and blood pressure data, and improving the overall accuracy of blood pressure measurement.

[0033] Preferably, the feedback adjustment expression of the blood pressure measurement error self-correction module based on AI voice interaction is: This is the estimated value for the next round of deviation. For feedback coefficients, For reference blood pressure values, These are newly acquired speech feature parameters. For feedback constraint constants, For the iterative decay factor, This is the estimated deviation value from the previous round. This represents the number of iterations.

[0034] Specifically, the feedback adjustment of the blood pressure measurement error self-correction module based on AI voice interaction constructs a closed-loop feedback mechanism to optimize the next round of deviation estimation. The feedback coefficient is set to 0.6, and the reference blood pressure value is determined based on the average of the user's three historical accurate measurement data, with an update cycle of 24 hours. The newly acquired voice feature parameters are 184-dimensional features acquired synchronously during each round of measurement. The feedback constraint constant is set to 0.01 to avoid excessive adjustment amplitude due to an excessively small denominator. The iteration decay factor is set to 0.1, and the number of iterations is set to 5 to ensure the stability and convergence of the feedback adjustment. When the module runs, it first receives the correction result and the user's voice feedback (voice recognition accuracy not less than 98%, response time 50 milliseconds), calculates the deviation between the correction result and the reference blood pressure value, and increases the feedback adjustment intensity by 40% when the deviation is greater than 2 mmHg, and decreases the adjustment intensity by 20% when the deviation is less than 1 mmHg for three consecutive rounds. By multiplying the absolute value of the deviation difference by the iteration decay factor, the adjustment amplitude in the iteration process is gradually reduced to avoid oscillation. At the same time, the deviation estimation value is dynamically adjusted by combining the new mapping relationship between voice features and the corrected blood pressure value. This feedback adjustment mechanism executes every 2 seconds, keeping it synchronized with the measurement cycle. The update time for the next round of deviation estimation is controlled within 10 milliseconds, enabling the deviation estimation to follow the changes in the measurement scenario in real time, forming a closed loop of "measurement-correction-feedback-optimization". This ensures the continuity and timeliness of error self-correction and continuously improves the stability of blood pressure measurement.

[0035] Preferred, such as Figure 2 As shown, step S3 includes the following sub-steps: S31, the filtered speech feature parameters are split into multi-dimensional features according to the frequency domain, time domain, and Mel-frequency cepstral domain to form a structured feature set. At the same time, the original blood pressure data is processed by time-series segmentation to obtain data segments of equal duration; S32, the structured feature set and the blood pressure data segments are paired one-to-one to construct a training sample set including input features and output labels. The input features are a combination vector of multi-dimensional speech features and original blood pressure data, and the output labels are preset deviation reference values; S33, the training sample set is divided into a training set, a validation set, and a test set according to a preset ratio. The blood pressure measurement deviation learning model is iteratively trained using the training set, the model hyperparameters are adjusted in real time using the validation set, and the deviation estimation accuracy of the model is verified using the test set; S34, based on the trained model, the real-time collected speech feature parameters and original blood pressure data are input, and the initial deviation estimate value corresponding to the measurement scenario is output through nonlinear transformation operations within the model.

[0036] Specifically, step S3 achieves accurate output of the initial deviation estimate through four sub-steps. S31 first decomposes the filtered speech feature parameters from step S2 into multi-dimensional features in the frequency domain, time domain, and Mel-frequency cepstral domain. Frequency domain features include a full-band division from 20Hz to 8kHz; time domain features include key indicators such as signal amplitude and zero-crossing rate; and Mel-frequency cepstral domain features are set to 128 dimensions, forming a structured feature set. Simultaneously, the raw blood pressure data is processed into time-series segments at 50-millisecond intervals to obtain data fragments of equal duration, ensuring data granularity consistency. S32 pairs the structured feature set with the blood pressure data fragments according to their acquisition timestamps, constructing a training sample set including input features and output labels. The input features are a combined vector of multi-dimensional speech features and raw blood pressure data, with a dimension set of 200. The output labels are based on the measurement results from precision medical equipment. A fixed deviation reference value was established, and the total number of samples reached 100,000. In S33, the training sample set was divided into training, validation, and test sets in a 7:2:1 ratio. The blood pressure measurement deviation learning model was iteratively trained using the training set, with a batch size of 64 and an initial learning rate of 0.001, decreasing by 0.1 every 100 iterations. The model's hyperparameters were adjusted in real-time using the validation set to ensure generalization ability. The deviation estimation accuracy of the model was verified using the test set, requiring the mean squared error on the test set to be less than 0.8. In S34, based on the trained model, real-time collected voice feature parameters and raw blood pressure data were input. Through nonlinear transformation operations using a multi-layer neural network within the model, the computation time was controlled within 25 milliseconds, outputting the initial deviation estimate for the corresponding measurement scenario. The absolute error of the estimate was controlled within 2 mmHg, providing reliable basic data support for subsequent dynamic compensation.

[0037] Preferred, such as Figure 3 As shown, S4 includes the following sub-steps: S41, collecting the user's age, weight, vascular elasticity coefficient, and environmental parameters such as temperature, humidity, and electromagnetic interference intensity of the measurement environment; quantifying and encoding these parameters to form a standardized parameter vector; S42, inputting the standardized physiological and environmental parameters into the parameter adaptation layer of the dynamic blood pressure error compensation model; generating a parameter weight matrix that matches the initial deviation estimate through feature mapping operations within the layer; S43, adjusting the initial deviation estimate component by component based on the parameter weight matrix; updating the adjustment range in real time according to the parameter change trend through the model's built-in dynamic adjustment mechanism to eliminate the error effects caused by physiological differences and environmental changes; S44, integrating the adjustment results of each component into a unified dynamic compensation coefficient through the integration operation of the model output layer to ensure the adaptability of the compensation coefficient to the current measurement scenario.

[0038] Specifically, step S4 achieves adaptive generation of dynamic compensation coefficients through a four-step process. S41 first collects eight physiological characteristic parameters of the user: age, weight, height, vascular elasticity coefficient, heart rate, historical average blood pressure, body mass index, and blood viscosity. Simultaneously, it collects six environmental parameters: temperature, humidity, air pressure, electromagnetic interference intensity, light intensity, and wind speed. The temperature acquisition accuracy is ±0.1℃, the humidity acquisition accuracy is ±1%, and the electromagnetic interference intensity measurement range is 0-100dB. All parameters are quantized and encoded, mapping the values ​​to the 0-1 interval to form a standardized parameter vector. S42 inputs the standardized 14-dimensional parameter vector into the parameter adaptation layer of the dynamic blood pressure error compensation model. This adaptation layer uses a fully connected neural network structure with 32 hidden layer nodes. Through feature mapping operations within the layer, a 32-dimensional parameter weight matrix is ​​generated. The weight matrix elements... The value ranges from 0.01 to 0.99, and is dynamically adjusted based on the sensitivity of the parameter to the deviation. S43 adjusts the initial deviation estimate component by component based on the parameter weight matrix. The adjustment range of each component is updated every 5 milliseconds according to the parameter change trend. The adjustment step size is set to a dynamic range of 0.005 to 0.05. When the environmental change rate exceeds 5% / minute or the physiological parameter deviates from the historical mean by more than 10%, the adjustment range is automatically increased by 30%. The dynamic adjustment mechanism built into the model eliminates the error caused by physiological differences and environmental changes. S44 integrates the adjustment results of each component into a unified dynamic compensation coefficient through weighted summation calculation of the model output layer. The compensation coefficient ranges from 0.8 to 1.2. The calculation time of a single round is controlled within 15 milliseconds to ensure that the compensation coefficient is highly adapted to the current measurement scenario and provides an accurate adjustment basis for subsequent correction calculations.

[0039] Preferred, such as Figure 4 As shown, step S5 includes the following sub-steps: S51, inputting the dynamic compensation coefficient, voice feature parameters, raw blood pressure data, and initial deviation estimate into the feature fusion module of the multimodal blood pressure correction analysis engine, and allocating the weight ratio of each input data through an attention mechanism to enhance the representation of calibration information; S52, in the engine's computation layer, performing preliminary correction calculations on the raw blood pressure data based on preset multi-dimensional correction rules and combined with the dynamic compensation coefficient, while constructing correction constraints using voice feature feedback information; S53, extracting the time-series trend of the preliminarily corrected blood pressure data through the engine's trend analysis module, comparing the change patterns of historical correction data, identifying abnormal correction results, and correcting them; S54, inputting the corrected correction results into the engine's output calibration module, and generating the final multi-dimensional error correction result after accuracy verification and format standardization processing.

[0040] Specifically, step S5 achieves multi-dimensional error correction through four collaborative operations. S51 inputs the dynamic compensation coefficient, 184-dimensional speech feature parameters, 16-dimensional blood pressure raw data statistical features, and initial deviation estimates into the feature fusion module of the multimodal blood pressure correction analysis engine. This module uses an attention mechanism to allocate the weight of each input data. The weight allocation range for speech feature feedback is 0.3 to 0.6, for blood pressure data trends it is 0.2 to 0.4, and for dynamic compensation coefficients it is 0.1 to 0.3. This weight optimization strengthens the representation of key information and suppresses invalid interference information. S52, in the engine's core operation layer, incorporates 12 multi-dimensional correction rules, including amplitude correction, temporal correction, and correlation correction. These rules, combined with the dynamic compensation coefficient, perform preliminary correction operations on the raw blood pressure data. Simultaneously, they utilize information such as frequency change trends and amplitude stability from the speech features to construct three correction constraints. The threshold of the constraint condition is dynamically adjusted based on historical calibration data to ensure the rationality of the preliminary calibration results. S53 uses the engine's trend analysis module and a sliding window algorithm with a window size of 20 frames to extract the time-series trend of the blood pressure data after preliminary calibration for 10 consecutive acquisition cycles. It compares the change patterns of the past 100 sets of historical calibration data, calculates the standard deviation of the data, and judges the results that deviate from 3 times the standard deviation as outliers and performs reverse correction. S54 inputs the corrected calibration results into the engine's output calibration module. This module performs accuracy verification by comparing with the reference standard value. The standard for passing the verification is that the deviation between the calibration result and the reference value is less than 1.5 mmHg. After passing the verification, the format is standardized to generate the final calibration result including systolic blood pressure, diastolic blood pressure, and calibration confidence. The entire calibration calculation process takes less than 30 milliseconds to ensure that the timeliness requirements of real-time measurement are met and to give full play to the calibration advantages of multimodal data fusion.

[0041] The multi-sensor speech collaborative filtering algorithm in this invention is a speech signal optimization processing technology that integrates spatiotemporal information. It integrates raw speech data collected from multiple sensors and combines dynamic weight allocation and interference suppression mechanisms to achieve noise filtering and effective feature extraction. Its implementation is based on a distributed layout of multi-channel speech acquisition modules. Spatiotemporally adaptive weight coefficients adjust the data contribution ratio in real time according to the signal-to-noise ratio of each sensor signal. Sensor confidence factors and speech feature gradient values ​​are used to strengthen blood pressure-related signal components. Simultaneously, interference suppression coefficients are used to specifically suppress environmental noise such as power frequency and high frequency, and a minimum constant is used to avoid computational anomalies. This algorithm accurately separates effective features related to blood pressure changes from mixed-interference speech signals, eliminating the influence of environmental noise, electromagnetic interference, and physiological noise on signal quality. It provides highly reliable speech feature parameters for subsequent bias learning and correction calculations, breaking through the limitations of traditional single-sensor signal processing. Through multi-source data collaboration, it improves the accuracy of speech feature extraction, fully explores the physiological state correlation information contained in the speech signal, provides key data support for blood pressure measurement error self-correction, and lays the foundation for multimodal fusion correction.

[0042] The blood pressure measurement deviation learning model in this invention is an intelligent prediction model that mines the nonlinear correlation between speech features and blood pressure data. Specifically, it constructs a mapping relationship between input and output through a neural network architecture to achieve accurate estimation of the initial deviation of blood pressure measurement. Its implementation process uses a massive amount of labeled samples as training foundation, taking filtered speech feature parameters and raw blood pressure data as input. Through operations such as weight vector optimization, coupled feature matrix construction, and fusion of physiological and environmental influencing factors, combined with nonlinear mapping functions and integral operations, it captures the dynamic correlation of data. After multiple rounds of iterative training and optimization of model parameters, it ensures the accuracy of deviation estimation for different individuals and scenarios. This model is based on the inherent correlation between speech and blood pressure data, outputting an initial deviation estimate. It quantifies the systematic deviation caused by individual physiological differences, measurement posture, and other factors during a single blood pressure sensor measurement, breaking the limitations of traditional deviation estimation relying on empirical formulas. Through a data-driven approach, it achieves adaptive prediction of deviation, providing a precise basis for subsequent dynamic compensation, significantly improving the targeting of error correction, and promoting the transformation of blood pressure measurement from "passive acquisition" to "active deviation prediction."

[0043] The dynamic blood pressure error compensation model in this invention is an adaptive adjustment model that combines physiological and environmental parameters. Specifically, it optimizes the initial deviation estimate in real time through multi-dimensional parameter fusion calculations, generating a dynamic compensation coefficient adapted to the current scenario. The implementation process takes user physiological characteristic parameters (age, vascular elasticity coefficient, etc.) and measurement environment parameters (temperature, humidity, etc.) as inputs. After quantization encoding and feature mapping, a parameter weight matrix is ​​generated. Through mechanisms such as proportional adjustment, secondary compensation, and deviation change rate sensing, the initial deviation is dynamically adjusted component by component, ultimately merging and outputting a unified compensation coefficient. This model eliminates the influence of physiological differences and dynamic environmental changes on blood pressure measurement, enabling deviation correction to adapt to the physiological characteristics of different users and real-time measurement environments. It solves the problem of insufficient adaptability of fixed compensation coefficients, constructs a dynamic adaptation mechanism for deviation correction, and shifts error compensation from a "fixed mode" to a "scenario-based adaptive mode," significantly improving the flexibility and accuracy of error correction. It provides an adjustment basis that fits the actual scenario for multimodal correction analysis, further reducing measurement errors.

[0044] The multimodal blood pressure correction analysis engine in this invention is the core computational unit of the entire error self-correction method. Specifically, it achieves multi-dimensional and accurate correction of blood pressure measurement errors through multi-source data fusion and multi-rule collaborative computation. Its implementation process integrates dynamic compensation coefficients, voice feature parameters, raw blood pressure data, and initial deviation estimates. It allocates weights to each data point through an attention mechanism, performs preliminary correction based on preset correction rules, identifies and corrects abnormal results using time-series trend analysis, and outputs the final correction result after accuracy verification and standardization. This engine integrates voice and blood pressure multimodal data, leveraging the complementary advantages of each data point. Through multi-stage collaborative computation, it achieves comprehensive error correction, transforming initial deviations and dynamic compensation into accurate blood pressure measurement results. It constructs a multi-dimensional, end-to-end correction system, breaking the limitations of traditional single-dimensional correction. It fully utilizes the synergistic value of voice feature feedback and blood pressure data trends to ensure the accuracy and reliability of the correction results, while meeting the timeliness requirements of real-time measurement. This provides core technical support for accurate blood pressure monitoring in non-professional scenarios and promotes the upgrading of home blood pressure measurement devices towards high precision and intelligence.

[0045] like Figure 5As shown, a blood pressure measurement error self-correction system based on AI voice interaction is applied to a blood pressure measurement error self-correction method based on AI voice interaction. The system includes: a multi-source data synchronous acquisition unit, used to establish a spatiotemporal alignment link with a distributed blood pressure sensing submodule through a multi-channel voice acquisition submodule, synchronously acquiring user voice signals and raw blood pressure data, and transmitting the data to a signal preprocessing unit; a signal preprocessing unit, communicatively connected to the multi-source data synchronous acquisition unit, with a built-in multi-sensor voice collaborative filtering algorithm module, performing noise suppression processing on the voice signal and extracting voice feature parameters, transmitting the processed voice feature parameters and raw blood pressure data to a deviation estimation unit; and a deviation estimation unit, communicatively connected to both the signal preprocessing unit and the compensation adjustment unit, equipped with a blood pressure measurement deviation learning model, and mining... The system explores the nonlinear mapping relationship between voice features and blood pressure data, outputting an initial deviation estimate to the compensation and adjustment unit. The compensation and adjustment unit, connected to the deviation estimation unit and the correction analysis unit, generates dynamic compensation coefficients based on physiological features and environmental parameters using a dynamic blood pressure error compensation model, and transmits the compensation coefficients and initial deviation values ​​to the correction analysis unit. The correction analysis unit connects the multimodal blood pressure correction analysis engine to the compensation and adjustment unit and the feedback correction unit, integrates multi-dimensional data to perform error correction calculations, and outputs the correction results to the feedback correction unit. The feedback correction unit forms a closed-loop connection with the correction analysis unit and the multi-source data synchronous acquisition unit, receiving the correction results through an AI-based voice interaction-based blood pressure measurement error self-correction module, providing real-time feedback on the data acquisition and correction process, and performing dynamic self-correction of the error.

[0046] A self-correction method and system for blood pressure measurement errors based on AI voice interaction has been developed. Through the synchronous linkage of multi-channel voice acquisition and distributed blood pressure sensors, combined with a specialized filtering algorithm, it accurately extracts voice features related to blood pressure. This represents the first deep fusion of voice information and blood pressure data, completely overcoming the limitations of traditional methods that rely solely on single-sensor data. It fully explores the inherent correlation between the two types of data, providing richer support for error correction from a data perspective and significantly improving the adaptability to individual physiological differences. Simultaneously, relying on the synergistic operation of a bias learning model, a dynamic compensation model, and a multi-modal correction engine, a full-process processing mechanism from bias estimation and dynamic adjustment to multi-dimensional correction has been formed, replacing the crude mode of traditional fixed-coefficient or simple linear correction, and greatly improving the accuracy and adaptability of error correction.

[0047] This method and system address the shortcomings of traditional methods, such as the lack of integration of voice and blood pressure information and insufficient adaptation to physiological differences. By simultaneously collecting two types of data and mining nonlinear mapping relationships, the correction process is made to fully fit the individual physiological characteristics of users, effectively eliminating systematic biases caused by differences in vascular elasticity, circulatory status, etc. To address the deficiencies of existing technologies in terms of dynamic adaptive capabilities and closed-loop feedback logic, this method integrates physiological characteristics and environmental parameters in real time through a dynamic compensation model to generate compensation coefficients adapted to the current scenario, responding to dynamic changes in environmental factors such as temperature and humidity. At the same time, it uses AI voice interaction to build a closed-loop feedback mechanism to optimize correction parameters in real time during the measurement process, solving the problems of limited timeliness and accuracy of traditional correction methods, and providing reliable technical support for accurate blood pressure monitoring in non-professional scenarios.

[0048] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" 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 communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent 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 self-correction method for blood pressure measurement error based on AI voice interaction, characterized in that, Includes the following steps: S1. Simultaneously acquire user voice signals and raw blood pressure data through a multi-channel voice acquisition module and a distributed blood pressure sensor array to establish a spatiotemporally aligned data acquisition link. S2. Use a multi-sensor voice collaborative filtering algorithm to suppress noise in the mixed-interference voice signal and extract voice feature parameters related to blood pressure changes. S3. Input the filtered voice feature parameters and raw blood pressure data into a blood pressure measurement deviation learning model to explore the nonlinear mapping relationship between the two and output an initial deviation estimate. S4. Based on user physiological characteristic parameters and measurement environment parameters, adaptively adjust the initial deviation estimate using a dynamic blood pressure error compensation model to generate dynamic compensation coefficients. S5. Input the dynamic compensation coefficients into a multimodal blood pressure correction analysis engine, integrate voice feature feedback and blood pressure data trend analysis, and perform multi-dimensional error correction calculations. S6. Receive the correction results through an AI-based voice interaction-based blood pressure measurement error self-correction module, construct a closed-loop feedback mechanism, and perform real-time self-correction of errors during the measurement process.

2. The blood pressure measurement error self-correction method based on AI voice interaction according to claim 1, characterized in that, The expression for the multi-sensor speech collaborative filtering algorithm is: ,in: This is the filtered speech signal. The original voice signal collected by the i-th sensor. For spatiotemporal adaptive weighting coefficients, As the sensor trust factor, The gradient value of the speech feature. The interference suppression coefficient is... It is the minimum constant. For the number of sensors, For frequency parameters, This is a time parameter.

3. The blood pressure measurement error self-correction method based on AI voice interaction according to claim 1, characterized in that, The expression for the blood pressure measurement bias learning model is: ,in, This is the initial deviation estimate. For the model weight vector, This is the coupled feature matrix of speech features and blood pressure data. For the set of speech feature parameters, This is a collection of raw blood pressure data. This is a matrix of physiological and environmental influencing factors. This is the original blood pressure measurement value. A set of environmental parameters This is the deviation adjustment coefficient. It is a nonlinear mapping function. For measuring time windows, This is for the Hadamard product operation.

4. The method for self-correcting blood pressure measurement errors based on AI voice interaction according to claim 1, characterized in that, The expression for the dynamic blood pressure error compensation model is: ,in, For dynamic compensation coefficients, As a proportional adjustment factor, The function is influenced by physiological characteristics. A set of user physiological characteristic parameters For heart rate-related parameters, It is a secondary compensation factor. As a dynamic environmental factor, For ambient temperature parameters, For ambient humidity parameters, For constraint coefficients, For the k-th state influence factor, The number of state factors. The derivative adjustment coefficient is... This represents the rate of change of deviation.

5. The method for self-correcting blood pressure measurement errors based on AI voice interaction according to claim 1, characterized in that, The correction calculation expression of the multimodal blood pressure correction analysis engine is as follows: ,in, This is the corrected blood pressure value. As a voice-blood pressure co-correction factor, Let q be the q-th multimodal influence parameter. To affect the number of parameters, To correct the threshold, The activation coefficient, It is a non-linear activation function.

6. The method for self-correcting blood pressure measurement errors based on AI voice interaction according to claim 1, characterized in that, The feedback adjustment expression of the blood pressure measurement error self-correction module based on AI voice interaction is as follows: This is the estimated value for the next round of deviation. For feedback coefficients, For reference blood pressure values, These are newly acquired speech feature parameters. For feedback constraint constants, For the iterative decay factor, This is the estimated deviation value from the previous round. This represents the number of iterations.

7. The method for self-correcting blood pressure measurement errors based on AI voice interaction according to claim 1, characterized in that, S3 includes the following steps: S31, the filtered speech feature parameters are split into multi-dimensional features according to the frequency domain, time domain, and Mel-frequency cepstral domain to form a structured feature set. At the same time, the original blood pressure data is processed by time-series segmentation to obtain data segments of equal duration; S32, the structured feature set and the blood pressure data segments are paired one-to-one to construct a training sample set including input features and output labels. The input features are a combination vector of multi-dimensional speech features and original blood pressure data, and the output labels are preset deviation reference values; S33, the training sample set is divided into a training set, a validation set, and a test set according to a preset ratio. The blood pressure measurement deviation learning model is iteratively trained using the training set, the model hyperparameters are adjusted in real time using the validation set, and the deviation estimation accuracy of the model is verified using the test set; S34, based on the trained model, the real-time collected speech feature parameters and original blood pressure data are input, and the initial deviation estimate value corresponding to the measurement scenario is output through nonlinear transformation operations within the model.

8. The method for self-correcting blood pressure measurement errors based on AI voice interaction according to claim 1, characterized in that, S4 includes the following sub-steps: S41, collecting physiological characteristic parameters of the user such as age, weight, and vascular elasticity coefficient, as well as environmental parameters of the measurement environment such as temperature, humidity, and electromagnetic interference intensity; quantifying and encoding various parameters to form a standardized parameter vector; S42, inputting the standardized physiological characteristic parameters and environmental parameters into the parameter adaptation layer of the dynamic blood pressure error compensation model; generating a parameter weight matrix that matches the initial deviation estimate through feature mapping operations within the layer; S43, adjusting the initial deviation estimate component by component based on the parameter weight matrix; updating the adjustment range in real time according to the parameter change trend through the model's built-in dynamic adjustment mechanism to eliminate the error influence caused by physiological differences and environmental changes; S44, merging the adjustment results of each component into a unified dynamic compensation coefficient through the integration operation of the model output layer to ensure the adaptability of the compensation coefficient to the current measurement scenario.

9. A method for self-correcting blood pressure measurement errors based on AI voice interaction according to claim 1, characterized in that, S5 includes the following sub-steps: S51, inputting the dynamic compensation coefficient, voice feature parameters, raw blood pressure data, and initial deviation estimate into the feature fusion module of the multimodal blood pressure correction analysis engine, and allocating the weight ratio of each input data through an attention mechanism to strengthen the representation of calibration information; S52, at the engine's computation layer, performing preliminary correction calculations on the raw blood pressure data based on preset multi-dimensional correction rules and combined with the dynamic compensation coefficient, while constructing correction constraints using voice feature feedback information; S53, extracting the time-series trend of the preliminarily corrected blood pressure data through the engine's trend analysis module, comparing the change patterns of historical correction data, identifying abnormal correction results, and correcting them; S54, inputting the corrected correction results into the engine's output calibration module, and generating the final multi-dimensional error correction result after accuracy verification and format standardization processing.

10. A blood pressure measurement error self-correction system based on AI voice interaction, characterized in that, This system is applied to a blood pressure measurement error self-correction method based on AI voice interaction as described in claim 1, comprising: a multi-source data synchronous acquisition unit, used to establish a spatiotemporal alignment link with a distributed blood pressure sensing submodule through a multi-channel voice acquisition submodule, synchronously acquire user voice signals and raw blood pressure data, and transmit the data to a signal preprocessing unit; a signal preprocessing unit, communicatively connected to the multi-source data synchronous acquisition unit, having a built-in multi-sensor voice collaborative filtering algorithm module, performing noise suppression processing on the voice signal and extracting voice feature parameters, and transmitting the processed voice feature parameters and raw blood pressure data to a deviation estimation unit; and a deviation estimation unit, communicatively connected to both the signal preprocessing unit and the compensation adjustment unit, equipped with a blood pressure measurement deviation learning model, mining voice features and blood pressure data... Based on the nonlinear mapping relationship, the initial deviation estimate is output to the compensation and adjustment unit. The compensation and adjustment unit communicates with the deviation estimation unit and the correction analysis unit. Based on physiological characteristics and environmental parameters, it generates dynamic compensation coefficients through the dynamic blood pressure error compensation model and transmits the compensation coefficients and the initial deviation value to the correction analysis unit. The correction analysis unit communicates with the multimodal blood pressure correction analysis engine, the compensation and adjustment unit, and the feedback correction unit. It integrates multi-dimensional data to perform error correction calculations and outputs the correction results to the feedback correction unit. The feedback correction unit forms a closed-loop connection with the correction analysis unit and the multi-source data synchronous acquisition unit. It receives the correction results through the blood pressure measurement error self-correction module based on AI voice interaction, and provides real-time feedback on the data acquisition and correction process to perform dynamic self-correction of the error.

Images