A physiological parameter prediction system and method based on a dual model architecture and intelligent routing

Abstract

The application provides a physiological parameter prediction system and method based on a double model architecture and intelligent routing, relates to the technical field of physiological signal processing and intelligent physiological parameter prediction, and the system comprises: a preprocessing module that acquires synchronous electrocardiogram and pulse wave signals of a to-be-measured object and generates a plurality of time windows; a quality assessment module that performs quality scoring and gating screening on each window, transmits the window to a feature extraction module, and transmits the score to a fusion module; the feature extraction module extracts a high-dimensional feature vector through the window; a baseline management module stores and updates a personal baseline according to the object dimension; a routing prediction module processes the high-dimensional feature vector to obtain a parameter prediction result after identifying the baseline; and the fusion module aggregates and outputs a target physiological parameter from the prediction result and the quality score. The application realizes physiological parameter prediction that takes into account the feasibility of engineering deployment, the robustness of signal quality, the individualized prediction accuracy, and the full-process automatic management.

Description

Technical Field

[0001] This invention relates to the field of physiological signal processing and intelligent physiological parameter prediction technology, and more specifically, to a physiological parameter prediction system and method based on a dual-model architecture and intelligent routing. Background Technology

[0002] Human temporal physiological signals (such as blood pressure) are core temporal data reflecting the human physiological state, and have significant application value in population health screening, long-term vital sign tracking, intraoperative status monitoring, home vital sign recording, and portable terminal data collection scenarios. Physiological parameter fitting and analysis technology based on multi-channel synchronous physiological temporal signals supports long-term uninterrupted data acquisition with a user-friendly experience, and has become an important research direction in mobile data processing and terminal automatic analysis. The core idea of ​​this technology is to extract dual-channel temporal raw data and its derived features to complete the numerical fitting calculation of target physiological parameters. Traditional contact-based segmented acquisition and measurement schemes have mature calculation benchmarks, but they suffer from drawbacks such as intermittent data acquisition, strong wearing constraints, and the inability to achieve long-term continuous temporal acquisition.

[0003] Existing time-series physiological signal parameter fitting data processing solutions use a single global model, which cannot simultaneously take into account the generalization of new samples and the adaptation of individual characteristics of existing objects. Personalized model switching and recalculation both require manual operation, which is difficult to deploy. At the same time, there is a lack of an automatic screening mechanism for pre-processed data quality. Distorted and noisy time-series data will directly participate in the calculation, resulting in large calculation errors and poor process stability. The solution can only output single-window independent results, with large fluctuations in output data and insufficient fault tolerance. Furthermore, the automation level of the entire chain is low, with the entire process of data import, cleaning, modeling, and export relying on manual intervention, which cannot support batch calculations. Summary of the Invention

[0004] The problem solved by this invention is that existing dual-channel time-series physiological signal parameter fitting and processing schemes are difficult to automatically adapt to different acquisition objects, lack pre-data quality screening, lack a multi-window calculation result aggregation mechanism, and have a low degree of automation in the overall data processing chain.

[0005] To address the aforementioned issues, this invention provides a physiological parameter prediction system, system, electronic device, and storage medium based on a dual-model architecture and intelligent routing.

[0006] In a first aspect, the present invention provides a physiological parameter prediction system based on a dual-model architecture and intelligent routing, comprising a data preprocessing module, a signal quality assessment module, a feature extraction module, a calibration baseline management module, an intelligent routing and model prediction module, and a result fusion module.

[0007] The data preprocessing module is used to acquire the synchronous electrocardiogram signal and pulse wave signal of the subject under test, and preprocess the synchronous electrocardiogram signal and the pulse wave signal to obtain multiple time window signals, and transmit the multiple time window signals to the signal quality evaluation module.

[0008] The signal quality assessment module is used to perform quality scoring and gating filtering on multiple time window signals respectively, to obtain multiple time window signals that pass the gating filtering and corresponding quality score results, and to transmit the multiple time window signals to the feature extraction module, while transmitting the corresponding quality score results to the result fusion module.

[0009] The feature extraction module is used to perform multi-dimensional feature extraction on the time window signals that have passed the gating screening to obtain high-dimensional feature vectors, and transmit the high-dimensional feature vectors to the calibration baseline management module and the intelligent routing and model prediction module at the same time.

[0010] The calibration baseline management module is used to store and update the high-dimensional feature vector according to the dimension corresponding to the object under test, to obtain personal baseline data, and transmit the personal baseline data to the intelligent routing and model prediction module.

[0011] The intelligent routing and model prediction module is used to identify the personal baseline data, process the high-dimensional feature vector according to the identification result, obtain the parameter prediction result, and transmit it to the result fusion unit.

[0012] The result fusion module is used to aggregate the parameter prediction results and the corresponding quality score results to output the target physiological parameter prediction results.

[0013] Optionally, the data preprocessing module includes a data access unit and a window segmentation unit;

[0014] The data access unit is used to acquire the synchronous electrocardiogram signal and pulse wave signal of the object under test and transmit them to the window segmentation unit;

[0015] The window segmentation unit is used to verify the validity of the synchronized electrocardiogram signal and the pulse wave signal, and to segment the verified signal into multiple time window signals of fixed length and overlapping each other, and to transmit the time window signals to the signal quality evaluation unit.

[0016] Optionally, the signal quality assessment module includes a signal filtering unit and a quality assessment unit;

[0017] The quality assessment unit is used to determine the quality of multiple time window signals, output the quality score of the synchronized electrocardiogram signal and the quality score of the pulse wave signal, and obtain the quality score result based on the quality score of the synchronized electrocardiogram signal and the quality score of the pulse wave signal.

[0018] The signal filtering unit is used to obtain multiple time window signals that pass the gated filtering by using whether the quality score of the pulse wave signal corresponding to each time window signal is greater than or equal to a preset threshold as a gating filtering condition, and then transmits the multiple time window signals to the feature extraction module, while transmitting the corresponding quality score results to the result fusion module.

[0019] Optionally, the feature extraction module includes a basic feature extraction unit, a core feature filtering unit, and a feature vector construction unit;

[0020] The basic feature extraction unit is used to extract ECG morphological features, ECG heart rate variability features, pulse wave time domain features, pulse wave derivative features, pulse wave frequency domain and entropy features, and ECG and pulse wave interaction features based on multiple time window signals that have been screened by gating, to obtain an initial feature set.

[0021] The core feature filtering unit is used to filter out features from the initial feature set that have a correlation with blood pressure changes that meet a preset threshold, thereby forming a core feature set;

[0022] The feature vector construction unit is used to generate inverse features, logarithmic features and binary interaction features according to the core feature set to obtain an extended feature set, and to concatenate the features in the extended feature set to form the high-dimensional feature vector. The high-dimensional feature vector is then transmitted to the calibration baseline management module and the intelligent routing and model prediction module, respectively.

[0023] Optionally, the personal baseline data includes personal characteristic baseline, personal systolic blood pressure baseline, personal diastolic blood pressure baseline, and training sample count;

[0024] The personal feature baseline is generated by averaging all the high-dimensional feature vectors;

[0025] The individual systolic blood pressure baseline and the individual diastolic blood pressure baseline are generated by averaging the historical systolic blood pressure samples and diastolic blood pressure samples of the subject to be tested, respectively.

[0026] When the number of new valid samples corresponding to the test object reaches the preset update threshold, the baseline data of the test object and the training sample count are automatically recalculated and updated.

[0027] Optionally, the intelligent routing and model prediction module includes an intelligent routing unit, a general prediction unit, and a personalized calibration prediction unit;

[0028] The intelligent routing unit is used to query whether the object to be tested belongs to a preset individualized object, and to query whether the object to be tested has the personal baseline data;

[0029] When both conditions are met, the high-dimensional feature vector and personal baseline data corresponding to the test object are routed to the personalized calibration prediction unit: the personalized calibration prediction unit is used to process the high-dimensional feature vector and personal baseline data corresponding to the test object and output the parameter prediction results;

[0030] When any condition is not met, the high-dimensional feature vector corresponding to the test object is routed to the general blood pressure prediction unit: the general prediction unit is used to process the high-dimensional feature vector corresponding to the test object and output the parameter prediction result.

[0031] Optionally, the general prediction unit is used to sequentially perform median imputation, Z-score standardization, and gradient boosting tree regression operations on the high-dimensional feature vector corresponding to the object to be tested, and output the parameter prediction results.

[0032] Optionally, the personalized calibration prediction unit is used to calculate the difference between the high-dimensional feature vector corresponding to the test object in the current time window and the personal baseline data to obtain the relative feature vector;

[0033] The relative feature vector is subjected to median imputation of missing values, Z-score standardization, and gradient boosting tree regression to obtain the bias prediction value.

[0034] The absolute value recovery process is performed on the predicted deviation value and the individual baseline data to obtain the parameter prediction result.

[0035] Optionally, the result fusion module includes an identification and quality aggregation unit, a data fusion unit, and an anomaly handling unit;

[0036] The identification and quality aggregation unit is used to group the received parameter prediction results and the corresponding quality score results according to the identifier corresponding to the test object as the grouping index, select the maximum value of the quality score results of all windows in the same group, generate the overall quality assessment result corresponding to a single collection record, and transmit it to the data fusion subunit and the anomaly handling subunit.

[0037] The data fusion unit is used to perform effective window filtering on the overall quality assessment results, calculate the mean of the parameter prediction results corresponding to the effective window, and generate the target physiological parameter prediction results corresponding to the single collection record.

[0038] The anomaly handling unit is used to generate an invalid identifier and a signal quality anomaly prompt when all windows in a group fail to pass the valid window filtering, and to replace the target physiological parameter prediction result output.

[0039] Secondly, the present invention provides a physiological parameter prediction method based on a dual-model architecture and intelligent routing, comprising:

[0040] Acquire synchronized electrocardiogram (ECG) signals and pulse wave signals of the subject under test, and preprocess the synchronized ECG signals and pulse wave signals to obtain multiple time window signals;

[0041] The multiple time window signals are subjected to quality scoring and gating filtering processes respectively to obtain multiple time window signals that pass the gating filtering and their corresponding quality score results.

[0042] Multi-dimensional feature extraction is performed on the time window signals that have passed the gating screening to obtain high-dimensional feature vectors;

[0043] The high-dimensional feature vector is stored and updated according to the dimension corresponding to the object under test to obtain personal baseline data;

[0044] The personal baseline data is identified, and the high-dimensional feature vector is automatically routed to a general prediction model or a personalized calibration prediction model based on the identification results to obtain parameter prediction results.

[0045] The predicted results of the parameters are aggregated with the corresponding quality score results to output the predicted results of the target physiological parameters.

[0046] The present invention relates to a physiological parameter prediction system, system, electronic device, and storage medium based on a dual-model architecture and intelligent routing. Through a data preprocessing module, synchronous electrocardiogram (ECG) and pulse wave signals are acquired and preprocessed to generate multiple sets of time window signals, which are then transmitted downwards. A signal quality assessment module performs quality scoring and gating on the window signals. The selected effective windows are sent to a feature extraction module, and the quality scores are simultaneously sent to a result fusion module. The feature extraction module extracts multi-dimensional features from the effective windows to generate high-dimensional feature vectors, which are then transmitted to a baseline calibration management module and an intelligent routing model prediction module, respectively. The baseline calibration management module stores and updates the feature vectors according to the object dimension to generate individual baseline data and transmits it to the intelligent routing model prediction module. The model prediction module and the intelligent routing and model prediction module identify baseline data and automatically assign feature vectors to the corresponding prediction units to generate parameter prediction results, which are then sent to the result fusion module. The result fusion module combines the received parameter prediction results with the quality score to complete the aggregation calculation and output the target physiological parameters. The entire architecture relies on the calibration baseline management module in conjunction with the intelligent routing and model prediction module to achieve automatic model switching, adapt to different acquisition objects, and reduce manual operation. It relies on the signal quality assessment module to filter low-quality signals to reduce calculation bias, and uses the result fusion module to aggregate multi-window data to improve the stability of the output results. All modules in the entire link automatically transfer data, and the time series data can be automatically processed without manual intervention. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the physiological parameter prediction system based on dual-model architecture and intelligent routing in an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the automated execution process of the feature extraction module in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the automated decision-making process of the intelligent routing unit in an embodiment of the present invention. Detailed Implementation

[0050] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0051] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0052] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the description below. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or modules, and are not intended to limit the order of functions performed by these devices, modules, or modules, or their interdependencies.

[0053] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0054] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0055] To address the problems existing in the prior art, this embodiment provides a physiological parameter prediction system and method based on a dual-model architecture and intelligent routing.

[0056] Combination Figure 1 As shown in the figure, the physiological parameter prediction system based on dual-model architecture and intelligent routing provided by the present invention includes a data preprocessing module, a signal quality assessment module, a feature extraction module, a calibration baseline management module, an intelligent routing and model prediction module, and a result fusion module.

[0057] The data preprocessing module is used to acquire the synchronous electrocardiogram signal and pulse wave signal of the subject under test, and preprocess the synchronous electrocardiogram signal and the pulse wave signal to obtain multiple time window signals, and transmit the multiple time window signals to the signal quality evaluation module.

[0058] The signal quality assessment module is used to perform quality scoring and gating filtering on multiple time window signals respectively, to obtain multiple time window signals that pass the gating filtering and corresponding quality score results, and to transmit the multiple time window signals to the feature extraction module, while transmitting the corresponding quality score results to the result fusion module.

[0059] The feature extraction module is used to perform multi-dimensional feature extraction on the time window signals that have passed the gating screening to obtain high-dimensional feature vectors, and transmit the high-dimensional feature vectors to the calibration baseline management module and the intelligent routing and model prediction module at the same time.

[0060] The calibration baseline management module is used to store and update the high-dimensional feature vector according to the dimension corresponding to the object under test, to obtain personal baseline data, and transmit the personal baseline data to the intelligent routing and model prediction module.

[0061] The intelligent routing and model prediction module is used to identify the personal baseline data, process the high-dimensional feature vector according to the identification result, obtain the parameter prediction result, and transmit it to the result fusion unit.

[0062] The result fusion module is used to aggregate the parameter prediction results and the corresponding quality score results to output the target physiological parameter prediction results.

[0063] Specifically, in this embodiment, the synchronized electrocardiogram (ECG) signal and pulse wave (PPG) signal are continuous time-series physiological data acquired synchronously through dual channels. They are automatically acquired through a data preprocessing module. This embodiment supports real-time data access from wearable devices, portable monitoring terminals, and bedside monitoring devices, as well as batch import of offline record files (such as CSV format). The preprocessing steps specifically involve verifying the legality of the synchronized ECG and pulse wave signals of the test subject (e.g., the 30-second 500Hz sampling signal must meet 15,000 sampling points), detecting NaN and infinity values, and checking the synchronization of ECG and PPG signals. The time window signal is segmented time-series data with a fixed duration and overlapping intervals. For example, the original 30-second, 500Hz sampling signal is divided into three overlapping windows of 10 seconds each, i.e., 5,000 sampling points.

[0064] The data preprocessing module in this embodiment uniformly acquires dual-channel raw time-series data and completes standardized preprocessing. It automatically removes invalid data with missing or abnormal data, divides long time-series data into standardized window signals for downward transmission, and unifies the data scale of different acquisition devices. This provides basic time-series data with a unified format and no abnormal interference for subsequent quality assessment and feature extraction, solving the problem that existing solutions have messy raw data formats, lack standardized segmentation processing, and require manual data sorting before subsequent calculations can be carried out.

[0065] In this embodiment, the quality score is a 0-1 interval quantization index calculated based on the waveform and spectral characteristics of the time-series signal. The gating screening is a pre-filtering mechanism that sets a fixed numerical threshold and retains only the signals of the scoring window that meet the standard. By receiving all the time window signals from the preprocessed output, the quantization quality score is completed segment by segment, and the unqualified windows with waveform distortion and severe noise interference are filtered out through the gating mechanism. Only the valid windows are sent to the feature extraction module, while the quality score of each window is separately retained and sent to the result fusion module. Low-quality data is truncated from flowing into the subsequent calculation process, avoiding parameter calculation deviations caused by distorted data. This overcomes the shortcomings of existing solutions that lack pre-data screening and allow poor-quality time-series data to directly participate in the calculation, resulting in poor overall processing stability.

[0066] In one embodiment, automated filtering and cleaning are performed on each time window—the ECG signal is bandpass filtered from 0.5Hz to 48Hz, and the PPG signal is inverted and bandpass filtered from 0.5Hz to 15Hz. The cleaned signals are then uniformly resampled to 250Hz. A signal quality assessment unit performs automated quality determination on the ECG and PPG signals: the ECG quality score (based on the ECG quality classifier) ​​uses ssqi, ksqi, hr, rr_max, rr_ratio, template_corr, and lb_stat as input features; the PPG quality score (based on the ECG quality classifier) ​​uses bassqi and template_corr as input features; both classifiers are trained based on BaggingClassifier. A PPG quality score ≥ 0.7 is used as a hard gate condition to automatically filter low-quality windows.

[0067] The multidimensional features in this embodiment include, but are not limited to, ECG morphological features (such as R wave peak value and QRS width), ECG heart rate variability features (such as RR interval standard deviation and root mean square difference between adjacent RR intervals), PPG time domain features (such as peak value, trough value, and rise time), PPG derivative features (such as first derivative peak value and second derivative extreme point), PPG frequency domain and entropy features (such as low-frequency power, high-frequency power, and sample entropy), and ECG and PPG interaction features (such as pulse conduction time (PTT)).

[0068] Individual baseline data is a unique benchmark formed by averaging multiple historical valid feature vectors and historical parameter values ​​for a single test subject, and is automatically updated with new valid samples. In this embodiment, high-dimensional feature vectors are stored and iteratively updated according to the independent dimension of the test subject, continuously generating corresponding object-specific baseline data and sending it to the intelligent routing and model prediction module; long-term time-series data benchmarks for each test subject are retained, providing individual reference standards for differentiated calculations, supporting the subsequent model to automatically distinguish between newly collected subjects and existing archived subjects, and solving the shortcomings of existing solutions that lack individual baseline data and cannot achieve differentiated calculations.

[0069] The general prediction unit targets new data collection objects without a baseline, directly outputting target parameters based on feature vectors. The personalized calibration prediction unit calculates feature differences based on individual baselines and restores the final parameters by combining feature vectors. Intelligent routing is a scheduling logic that automatically identifies the existence of baselines and autonomously allocates feature vectors to the corresponding calculation units. This embodiment receives baseline data and the current window feature vector, automatically identifies whether the object has a valid personal baseline, and allocates feature vectors to the appropriate calculation units without manual operation. It outputs window-level parameter prediction results and pushes them to the result fusion module. It achieves automatic switching of calculation models by relying on baseline recognition, simultaneously taking into account both generalization calculation of new samples and accurate calculation of differences in existing objects, solving the problems of existing solutions where a single calculation model cannot adapt to different data collection objects and model switching relies on manual operation.

[0070] The aggregation process, based on a single complete data acquisition task, organizes all window measurement results and corresponding quality scores, calculates the mean of effective window measurements, and summarizes the overall quality indicators. The target physiological parameter prediction results are a standardized output file for each single data acquisition task. This embodiment summarizes the prediction results of each window parameter and the synchronously transmitted window quality scores, performs aggregation calculations by grouping by single data acquisition task, integrates all window measurement values ​​and calculates the mean, outputs an anomaly flag if all windows fail the quality gate, and finally outputs the overall measurement results in a unified format. By integrating independently calculated data from multiple windows, numerical fluctuations are smoothed, and complete quality evaluation information is included synchronously, overcoming the shortcomings of existing technologies that can only output independent results from a single window, have insufficient data fault tolerance, and exhibit large fluctuations.

[0071] This embodiment automatically processes data sequentially through each module, forming a complete automated processing chain from raw data access, preprocessing, quality screening, feature extraction, differential calculation to result aggregation and output. No manual intervention is required at any stage, and multiple batches of time-series collected data can be processed directly in batches, making up for the shortcomings of existing solutions, such as low automation level of the entire chain, manual intervention required at each step, and difficulty in batch calculation.

[0072] Optionally, the data preprocessing module includes a data access unit and a window segmentation unit;

[0073] The data access unit is used to acquire the synchronous electrocardiogram signal and pulse wave signal of the object under test and transmit them to the window segmentation unit;

[0074] The window segmentation unit is used to verify the validity of the synchronized electrocardiogram signal and the pulse wave signal, and to segment the verified signal into multiple time window signals of fixed length and overlapping each other, and to transmit the time window signals to the signal quality evaluation unit.

[0075] Specifically, the decoupled operation of data input and segmented preprocessing can be achieved through the data access unit and the window segmentation unit. The data access unit is responsible for the unified acquisition and transmission of dual-channel synchronous physiological signals, forming a standardized data access point that is compatible with synchronous ECG and pulse wave signals output from different acquisition terminals. This eliminates the need for manual data format alignment and synchronization verification, reducing the manual operation cost of the data import process. The window segmentation unit first performs validity verification on the original signal, automatically removing invalid data that is too short, contains outliers, or is out of sync with the dual signals, preventing abnormal data from flowing downwards and interfering with subsequent calculations. Then, the verified long-term signal is segmented into multiple fixed-length, overlapping time windows. This provides a standardized processing unit of uniform length for subsequent quality assessment and feature extraction, ensuring the consistency of the operational logic of each window. Furthermore, the overlapping window design preserves the continuous variation characteristics of the time-series data, reducing the impact of hard segmentation on the integrity of the time-series information and improving the reliability of subsequent parameter calculations.

[0076] In one embodiment, the automated processing flow of the data preprocessing module is as follows:

[0077] In real-time scenarios, automated connections can be established with wearable devices (such as smartwatches and finger-clip pulse oximeters), portable monitoring terminals, and bedside monitoring devices via wireless communication protocols such as Bluetooth and Wi-Fi, receiving synchronized ECG and PPG signals in real time. In offline scenarios, the system can automatically scan CSV format record files in a specified directory, batch reading ECG and PPG sequences without requiring manual file-by-file import. The data access unit automatically records the object identifier and acquisition record identifier for each data entry, providing an index for subsequent model routing and result fusion.

[0078] (a) Validity check: Automatically detects whether the signal length reaches the preset number of sampling points (e.g., 30 seconds × 500Hz = 15000 sampling points), whether there is NaN or infinite value, and whether the ECG and PPG signals are synchronized. Records that do not meet any of the conditions will be automatically marked and skipped.

[0079] (b) Time window segmentation: After verification, the long-term signal is automatically segmented into multiple fixed-length overlapping time windows. Preferably, the original signal sampled for 30 seconds at 500Hz is segmented into three overlapping windows of 10 seconds each (5000 sampling points), with an overlap of about 5 seconds between adjacent windows.

[0080] (c) Filtering and cleaning: Automatically perform 0.5Hz to 48Hz bandpass filtering on ECG signals to eliminate baseline drift and power frequency interference; automatically perform inversion operation (to make the peak face upward to conform to the subsequent feature extraction convention) and 0.5Hz to 15Hz bandpass filtering on PPG signals to eliminate high frequency noise and motion artifacts.

[0081] (d) Resampling: The cleaned ECG and PPG signals are uniformly resampled to 250Hz to eliminate the sampling rate differences between different acquisition devices and ensure the consistency of subsequent feature extraction scale.

[0082] Optionally, the signal quality assessment module includes a signal filtering unit and a quality assessment unit;

[0083] The quality assessment unit is used to determine the quality of multiple time window signals, output the quality score of the synchronized electrocardiogram signal and the quality score of the pulse wave signal, and obtain the quality score result based on the quality score of the synchronized electrocardiogram signal and the quality score of the pulse wave signal.

[0084] The signal filtering unit is used to obtain multiple time window signals that pass the gated filtering by using whether the quality score of the pulse wave signal corresponding to each time window signal is greater than or equal to a preset threshold as a gating filtering condition, and then transmits the multiple time window signals to the feature extraction module, while transmitting the corresponding quality score results to the result fusion module.

[0085] Specifically, the quality assessment unit and signal filtering unit ensure that quality quantification and gating filtering are independent, facilitating separate adjustments to assessment rules and filtering thresholds, and improving the flexibility and maintainability of system deployment. The quality assessment unit performs quality assessments on ECG and pulse wave signals for each time window, outputting quantitative quality scores for the two signals to form a complete quality score result. This accurately reflects the quality status of the two signals within each time window, providing comprehensive quantitative basis for subsequent filtering operations and result quality labeling. The signal filtering unit uses a hard-gating condition—whether the pulse wave signal quality score is greater than or equal to a preset threshold—to automatically filter out qualified windows and transmit them to the feature extraction module. This intercepts low-quality signals from the early stages of data processing, preventing distorted and noisy data from entering subsequent calculations, reducing interference from invalid data on parameter calculation results, and improving the overall stability of the calculation results. Simultaneously, this unit transmits the quality score results corresponding to each window to the result fusion module, ensuring that the final output physiological parameter results match the corresponding signal quality information, facilitating subsequent result verification and quality traceability.

[0086] In one embodiment, the signal quality assessment module uses an automated classification model to determine the quality, as follows:

[0087] (a) ECG quality classifier: It is trained based on BaggingClassifier with seven quality features as input: ssqi (spectral quality index), ksqi (kurtosis quality index), hr (heart rate), rr_max (maximum RR interval), rr_ratio (RR interval ratio), template_corr (template correlation coefficient) and lb_stat (Ljung-Box statistic). It automatically outputs the ECG quality score in the 0-1 interval for each time window.

[0088] (b) PPG quality classifier: It takes two quality features, baseqi (baseline drift quality index) and template_corr (template correlation coefficient), as input and is also trained based on BaggingClassifier. It automatically outputs the PPG quality score in the 0-1 interval for each time window.

[0089] (c) Hard gating mechanism: A PPG quality score ≥ 0.7 (this threshold can be dynamically configured) is used as a hard gating condition, allowing entry into the subsequent blood pressure prediction process only through the gating time window.

[0090] ECG quality results are provided as supplementary information with the final results and are for clinical data reference only. The data output in this embodiment should not be used as the basis for diagnosis.

[0091] Optionally, the feature extraction module includes a basic feature extraction unit, a core feature filtering unit, and a feature vector construction unit;

[0092] The basic feature extraction unit is used to extract ECG morphological features, ECG heart rate variability features, pulse wave time domain features, pulse wave derivative features, pulse wave frequency domain and entropy features, and ECG and pulse wave interaction features based on multiple time window signals that have been screened by gating, to obtain an initial feature set.

[0093] The core feature filtering unit is used to filter out features from the initial feature set that have a correlation with blood pressure changes that meet a preset threshold, thereby forming a core feature set;

[0094] The feature vector construction unit is used to generate inverse features, logarithmic features and binary interaction features according to the core feature set to obtain an extended feature set, and to concatenate the features in the extended feature set to form the high-dimensional feature vector. The high-dimensional feature vector is then transmitted to the calibration baseline management module and the intelligent routing and model prediction module, respectively.

[0095] Specifically, this embodiment achieves hierarchical decoupling of the entire feature extraction process through a basic feature extraction unit, a core feature selection unit, and a feature vector construction unit. Each unit has clearly defined responsibilities, facilitating independent adjustment of the feature extraction range, selection rules, and derived construction logic to adapt to the feature dimension requirements of different deployment scenarios. Specifically, the basic feature extraction unit simultaneously extracts six categories of features—ECG morphology, ECG heart rate variability, pulse wave time domain, pulse wave derivative, pulse wave frequency domain and entropy, and dual-signal interaction—within an effective time window, forming an initial feature set. This initial feature set fully mines effective information from time-series signals from multiple dimensions such as waveform morphology, temporal changes, frequency distribution, and channel correlation, comprehensively covering signal features related to the target physiological parameters and providing a comprehensive feature foundation for subsequent parameter calculations. The core feature selection unit selects features from the initial feature set whose correlation with the target parameters meets a preset threshold to form a core feature set. This eliminates features with high redundancy and poor extraction stability, retaining high-value features. Simultaneously, the feature dimension is compressed, reducing the amount of data for subsequent calculations and improving overall processing efficiency, while avoiding interference from invalid features in the prediction results. The feature vector construction unit generates reciprocal features, logarithmic features, and binary interaction features based on the core feature set, forming an extended feature set, which can enhance the nonlinear expressive power of features and more fully characterize the nonlinear relationship between features and target parameters. The extended features are then concatenated into standardized high-dimensional feature vectors, which are transmitted to the calibration baseline management module and the intelligent routing and model prediction module, respectively. This provides data support for the iterative update of personal baselines and provides input data in a unified format for subsequent parameter predictions, ensuring data consistency in subsequent calculation stages.

[0096] In one embodiment, such as Figure 2 As shown, the feature extraction module adopts a modular automated extraction architecture, extracting the following six categories of features for each time window that passes the quality gate:

[0097] (a) ECG morphological features: including R wave peak value, QRS width, QT interval, corrected QT interval, PQ interval, TQ interval, ST segment duration, etc., which reflect the morphological information of cardiac electrical activity.

[0098] (b) ECG heart rate variability characteristics: including mean heart rate, standard deviation of RR interval (SDNN), root mean square difference of adjacent RR intervals (RMSSD), low frequency power (LF), high frequency power (HF), low frequency to high frequency ratio (LF / HF), sample entropy, etc., reflecting the state of regulation of heart rhythm by the autonomic nervous system.

[0099] (c) PPG time-domain characteristics: including pulse wave peak value, trough value, rise time, fall time, half width at half maximum, relative amplitude of notch, relative amplitude of diastolic peak, reflection index, area ratio and geometric power characteristics, reflecting the pulse wave propagation morphology and vascular elasticity.

[0100] (d) PPG derivative characteristics: including the maximum slope of the first derivative, the minimum slope of the first derivative, the b / a ratio, c / a ratio, d / a ratio of the second derivative, and the aging index, which reflect changes in the compliance of the blood vessel wall.

[0101] (e) PPG frequency domain and entropy characteristics: including dominant frequency band power, spectral harmonic characteristics, sample entropy and permutation entropy, reflecting the frequency distribution and complexity of the pulse wave signal.

[0102] (f) ECG and PPG interaction features: The R wave in ECG is used as the cardiac ejection reference point, and the wave foot, peak and maximum slope point in PPG are used as the peripheral pulse arrival reference points. The pulse conduction time from R wave to PPG wave foot (PTT_foot), from R wave to PPG wave peak (PTT_peak) and from R wave to PPG maximum slope point (PTT_slope) are extracted.

[0103] In addition, this embodiment automatically constructs reciprocal features (such as 1 / PTT), logarithmic features (such as log(PTT)) and binary interaction features (such as feature product and feature ratio) for the above core features to enhance the model's ability to express features nonlinearly and form the final high-dimensional fusion feature vector.

[0104] Optionally, the personal baseline data includes personal characteristic baseline, personal systolic blood pressure baseline, personal diastolic blood pressure baseline, and training sample count;

[0105] The personal feature baseline is generated by averaging all the high-dimensional feature vectors;

[0106] The individual systolic blood pressure baseline and the individual diastolic blood pressure baseline are generated by averaging the historical systolic blood pressure samples and diastolic blood pressure samples of the subject to be tested, respectively.

[0107] When the number of new valid samples corresponding to the test object reaches the preset update threshold, the baseline data of the test object and the training sample count are automatically recalculated and updated.

[0108] Specifically, this embodiment stores and automatically updates personal baseline data according to the dimension of the object being tested, providing a stable individual benchmark for individualized parameter prediction and achieving automated maintenance of baseline data without manual intervention. The personal baseline data simultaneously includes personal characteristic baselines, personal systolic blood pressure baselines, personal diastolic blood pressure baselines, and training sample counts. It can characterize the baseline features of an individual's time-series signals from a feature perspective and reflect the baseline levels of an individual's physiological parameters from a parameter perspective, providing a comprehensive reference benchmark for subsequent personalized calibration and prediction. This allows for more accurate adaptation to signal differences among different individuals and improves the accuracy of individualized predictions. Using the mean as the baseline generation rule ensures stable and reliable calculation logic, smoothing out fluctuations in single samples, and truly reflecting the long-term baseline state of an individual, avoiding baseline shifts caused by single sample anomalies. The mechanism of automatically recalculating and updating when the number of newly added valid samples reaches a preset threshold allows for automatic iterative optimization of baseline data as the number of collected samples accumulates, continuously improving the individual representativeness of the baseline without requiring manual triggering of updates. The accompanying training sample counts visually reflect the sample support level of the baseline data, facilitating the assessment of baseline reliability.

[0109] In one embodiment, the calibration baseline management module performs automated data management by object dimension, as follows:

[0110] (a) Baseline generation: For subjects that have completed multiple valid data collections, the system automatically calculates the mean of the feature vectors of multiple historical samples to generate a personal feature baseline; it also automatically calculates the mean of historical systolic and diastolic blood pressure samples to generate a personal systolic blood pressure baseline and a personal diastolic blood pressure baseline.

[0111] (b) Sample count: The number of training samples on which the baseline calculation depends is recorded as a criterion for the validity of the baseline.

[0112] (c) Automatic update: When the number of new valid samples reaches a preset threshold (e.g., 5), the system automatically triggers baseline recalculation, incorporating the new data into the mean calculation to ensure the timeliness and accuracy of the baseline.

[0113] (d) Storage and indexing: Baseline data is stored with object identifiers and supports fast querying and retrieval of model routing units.

[0114] Optionally, the intelligent routing and model prediction module includes an intelligent routing unit, a general prediction unit, and a personalized calibration prediction unit;

[0115] The intelligent routing unit is used to query whether the object to be tested belongs to a preset individualized object, and to query whether the object to be tested has the personal baseline data;

[0116] When both conditions are met, the high-dimensional feature vector and personal baseline data corresponding to the test object are routed to the personalized calibration prediction unit: the personalized calibration prediction unit is used to process the high-dimensional feature vector and personal baseline data corresponding to the test object and output the parameter prediction results;

[0117] When any condition is not met, the high-dimensional feature vector corresponding to the test object is routed to the general blood pressure prediction unit: the general prediction unit is used to process the high-dimensional feature vector corresponding to the test object and output the parameter prediction result.

[0118] Specifically, this embodiment, by setting up an intelligent routing unit with dual-condition judgment logic, can automatically complete the prediction path allocation for the test object, realizing automatic switching between general prediction and personalized calibration prediction, without manual intervention to adjust the model throughout the process, thus improving the system's automated operation capability. It simultaneously verifies two conditions: "whether it belongs to a preset individualized object" and "whether there is valid personal baseline data." This dual judgment ensures routing accuracy, avoiding misallocation based on a single condition and ensuring that the personalized prediction unit is only activated when the object meets the configuration requirements and has sufficient baseline data support, guaranteeing the reliability of personalized prediction results. When both conditions are met, the high-dimensional feature vector and personal baseline data are routed together to the personalized calibration prediction unit. Adaptability calculations are performed based on the individual baseline data to fully match the object's signal characteristics and parameter benchmarks, improving the accuracy of personalized calculations. When either condition is not met, the feature vector is directly routed to the general prediction unit, ensuring that new objects and objects with insufficient baseline data can complete parameter calculations normally, guaranteeing the system's universal adaptability. This routing logic can automatically adapt to different types of objects under test without interrupting the processing flow, while taking into account both the coverage of generalized measurement and the accuracy of individualized measurement. It solves the problems of existing solutions where measurement model switching depends on manual operation and a single model cannot adapt to multiple types of collection objects at the same time.

[0119] In one embodiment, such as Figure 3 As shown, the automated decision-making process of the intelligent routing unit is as follows:

[0120] (a) After the system obtains the identifier of the object to be tested, it automatically queries whether the identifier belongs to the preset individualized object set (such as objects 01-30, which can be dynamically adjusted through the configuration file).

[0121] (b) At the same time, automatically check whether there is a valid baseline for the object in the calibration baseline management unit (the number of training samples corresponding to the baseline is ≥3).

[0122] (c) When both of the above conditions are met, the system will automatically route to the personalized blood pressure prediction unit; otherwise, it will automatically route to the general blood pressure prediction unit.

[0123] (d) This routing mechanism can adapt to individualized range adjustments for different deployment scenarios without modifying the core code.

[0124] Optionally, the general prediction unit is used to sequentially perform median imputation, Z-score standardization, and gradient boosting tree regression operations on the high-dimensional feature vector corresponding to the object to be tested, and output the parameter prediction results.

[0125] Specifically, the general prediction unit adopts an automated processing logic that sequentially connects missing value median imputation, Z-score standardization, and gradient boosting tree regression. It can independently complete parameter prediction for test objects that do not meet the personalized calculation conditions, ensuring the coverage and stability of the full-scenario calculation. Among these features, median imputation automatically fills in missing data in high-dimensional feature vectors. Compared to mean imputation, median imputation is less affected by extreme outliers, resulting in more robust imputation. It eliminates the need for manual checks to address missing data, ensuring the integrity of input data and preventing subsequent regression anomalies caused by data loss. Z-score standardization normalizes the numerical scale of features across different dimensions, eliminating dimensional differences and magnitude gaps between features. This ensures all feature dimensions are on a unified evaluation benchmark in the regression model, preventing features with large numerical ranges from excessively dominating the calculation results, optimizing the rationality of model weight allocation, and improving the stability of regression calculations. Gradient boosting tree regression effectively fits the nonlinear mapping relationship between high-dimensional features and target physiological parameters. It exhibits strong adaptability to multi-dimensional physiological time-series features and stable generalization performance. Even in general scenarios without individual baseline data support, it can still output reliable parameter prediction results, ensuring the accuracy of measurements for newly added test subjects and those with insufficient baseline data. The entire process in this embodiment is executed automatically and sequentially without manual intervention to adjust calculation parameters. It complements the personalized calibration prediction unit, jointly covering the measurement needs of different types of test subjects.

[0126] In one embodiment, the general prediction unit is designed for new users without a baseline and employs an automated regression pipeline: first, missing values ​​are imputed using the median; then, Z-score standardization is performed; and finally, absolute systolic and absolute diastolic blood pressure are directly regressed using a gradient boosting tree regressor (XGBoost). The general prediction unit is trained on a widely collected historical sample of test subjects and employs a cross-individual partitioning strategy based on test subject identification groups to ensure that data from the same test subject do not appear simultaneously in the training and validation sets, thus achieving cross-test subject generalization.

[0127] Optionally, the personalized calibration prediction unit is used to calculate the difference between the high-dimensional feature vector corresponding to the test object in the current time window and the personal baseline data to obtain the relative feature vector;

[0128] The relative feature vector is subjected to median imputation of missing values, Z-score standardization, and gradient boosting tree regression to obtain the bias prediction value.

[0129] The absolute value recovery process is performed on the predicted deviation value and the individual baseline data to obtain the parameter prediction result.

[0130] Specifically, by calculating the difference between the high-dimensional feature vector of the current window and the individual's baseline data to generate a relative feature vector, the absolute feature value can be converted into a fluctuation relative to the individual benchmark. This eliminates interference caused by differences in the basic signal levels between different individuals, allowing subsequent regression calculations to focus on the correspondence between relative signal changes and parameter fluctuations, avoiding result deviations caused by different individual baselines in the general calculation logic. The relative feature vector is then subjected to median imputation of missing values, Z-score standardization, and gradient boosting tree regression to obtain the deviation prediction value. Combining the deviation prediction value with the individual's baseline data completes the absolute value recovery process, restoring the deviation to an absolute parameter prediction result. This preserves the individual's long-term parameter benchmark level while incorporating the signal fluctuation characteristics of the current window, ensuring that the final prediction result both fits the individual's baseline state and accurately reflects parameter changes in the current period, resulting in stronger individual adaptability. It complements the general prediction unit, automatically activating when individualized calculation conditions are met, without requiring manual switching of calculation logic. While ensuring full-scenario coverage, it further improves the calculation accuracy of existing data.

[0131] In one embodiment, the personalized calibration prediction unit is designed for existing users with established records, and its core idea is "feature differentiation + baseline restoration":

[0132] (a) Difference vector calculation: The difference between the feature vector of the current time window and the baseline of the individual's features is automatically calculated to obtain the relative feature vector (difference vector), which reflects the deviation of the current physiological state relative to the individual's historical mean.

[0133] (b) Deviation prediction: Input the difference vector into the personalized calibration model (using the same missing value imputation → standardization → XGBoost regression pipeline) to output the systolic blood pressure deviation prediction value and the diastolic blood pressure deviation prediction value.

[0134] (c) Absolute value recovery: The absolute blood pressure value is automatically recovered through the formula "deviation + baseline", that is: systolic blood pressure prediction value = individual systolic blood pressure baseline + systolic blood pressure deviation prediction value; diastolic blood pressure prediction value = individual diastolic blood pressure baseline + diastolic blood pressure deviation prediction value.

[0135] The mechanism in this embodiment allows the model to focus on learning the physiological changes of an individual relative to their baseline, rather than regressing to absolute values ​​from zero, thereby significantly improving the accuracy of individualized predictions.

[0136] Optionally, the result fusion module includes an identification and quality summarization unit, a data fusion unit, and an anomaly handling unit;

[0137] The identification and quality aggregation unit is used to group the received parameter prediction results and the corresponding quality score results according to the identifier corresponding to the test object as the grouping index, select the maximum value of the quality score results of all windows in the same group, generate the overall quality assessment result corresponding to a single collection record, and transmit it to the data fusion subunit and the anomaly handling subunit.

[0138] The data fusion unit is used to perform effective window filtering on the overall quality assessment results, calculate the mean of the parameter prediction results corresponding to the effective window, and generate the target physiological parameter prediction results corresponding to the single collection record.

[0139] The anomaly handling unit is used to generate an invalid identifier and a signal quality anomaly prompt when all windows in a group fail to pass the valid window filtering, and to replace the target physiological parameter prediction result output.

[0140] Specifically, this embodiment enables the decoupled operation of three types of logic: grouping and aggregation, result fusion, and anomaly handling. Each unit has clearly defined responsibilities, facilitating independent adjustment of grouping rules, fusion strategies, and anomaly judgment criteria, thus improving system maintainability and adaptability. Specifically, the identification and quality aggregation unit uses the target object identifier as the grouping index, automatically aggregating all window parameter prediction results and quality scores corresponding to a single acquisition. This eliminates the need for manual classification and organization of multi-window data, reducing manual processing costs. Simultaneously, it selects the maximum quality score from all windows within the same group to generate the overall quality assessment result for a single acquisition record. This directly reflects the optimal signal quality level of the acquisition, providing a unified quantitative basis for subsequent effective window selection and result quality labeling. It also ensures that the final output matches the overall quality information, facilitating subsequent data verification and quality traceability. The data fusion unit selects effective windows based on the overall quality assessment result, using only the parameter prediction results of compliant windows to calculate the arithmetic mean, generating the target physiological parameter prediction result for a single acquisition record. By merging the results from multiple windows and taking the average, numerical fluctuations caused by independent calculations in a single window can be smoothed out, reducing the impact of random noise on the final result and improving the stability and fault tolerance of the output data. Simultaneously, only valid windows are included in the calculation, preventing low-quality windows from lowering the overall output accuracy and ensuring the reliability of the final output result. When all windows within a group fail the valid window screening, the anomaly handling unit automatically generates an invalid identifier and a signal quality anomaly warning message, replacing the normal target physiological parameter prediction result output. This mechanism avoids outputting invalid calculation results with large deviations when the overall quality of the collected data is unqualified, ensuring the rigor of the output data and preventing erroneous data from being misused in subsequent stages. Furthermore, the replacement output maintains the uniformity of the output format, ensuring that the batch processing flow is not interrupted due to a single data quality anomaly. In this embodiment, each unit is sequentially connected to form a complete result output link, automatically transferring data throughout the process. This ensures both the stability and accuracy of the output results under normal acquisition scenarios and covers abnormal scenarios where the quality of all windows is unqualified, adapting to the needs of automated batch data processing.

[0141] In one embodiment, the automated fusion process of the result fusion module is as follows:

[0142] (a) Grouping: Multiple time windows are automatically grouped according to "Identifier of the object to be tested + Identifier of the original collection record".

[0143] (b) Quality summary: The maximum value of the ECG quality score and PPG quality score of the same window is taken as the overall quality assessment result of the acquisition record.

[0144] (c) Predictive value fusion: For the effective window that passes through PPG quality gate, the mean predicted systolic blood pressure, the mean predicted diastolic blood pressure, and the mean heart rate are automatically calculated.

[0145] (d) Anomaly handling: If none of the windows in the same group pass the PPG quality gate, an invalid flag and a prompt message "Signal quality does not meet prediction requirements" will be automatically output.

[0146] (e) Output format: The final automated output is a file-level result in a uniform format, including object identifier, acquisition time, predicted systolic blood pressure, predicted diastolic blood pressure, heart rate, ECG quality score, and PPG quality score.

[0147] In summary, this embodiment can achieve fully automated deployment through batch prediction scripts: automatically scan all collected files in a specified directory, group them by object identifier, and then sequentially perform automated preprocessing, quality assessment, feature extraction, model routing, prediction inference, and result fusion. Finally, the prediction results of all objects are automatically summarized into an Excel format report, which significantly improves batch processing efficiency and meets the needs of large-scale application scenarios such as clinical monitoring and population health monitoring.

[0148] This invention also provides a physiological parameter prediction method based on a dual-model architecture and intelligent routing, comprising:

[0149] Acquire synchronized electrocardiogram (ECG) signals and pulse wave signals of the subject under test, and preprocess the synchronized ECG signals and pulse wave signals to obtain multiple time window signals;

[0150] The multiple time window signals are subjected to quality scoring and gating filtering processes respectively to obtain multiple time window signals that pass the gating filtering and their corresponding quality score results.

[0151] Multi-dimensional feature extraction is performed on the time window signals that have passed the gating screening to obtain high-dimensional feature vectors;

[0152] The high-dimensional feature vector is stored and updated according to the dimension corresponding to the object under test to obtain personal baseline data;

[0153] The personal baseline data is identified, and the high-dimensional feature vector is automatically routed to a general prediction model or a personalized calibration prediction model based on the identification results to obtain parameter prediction results.

[0154] The predicted results of the parameters are aggregated with the corresponding quality score results to output the predicted results of the target physiological parameters.

[0155] The physiological parameter prediction method based on dual-model architecture and intelligent routing of the present invention has the same advantages over the prior art as the physiological parameter prediction system based on dual-model architecture and intelligent routing described above, and will not be repeated here.

[0156] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A physiological parameter prediction system based on a dual-model architecture and intelligent routing, characterized in that, The system includes a data preprocessing module, a signal quality assessment module, a feature extraction module, a calibration baseline management module, an intelligent routing and model prediction module, and a result fusion module. The data preprocessing module is used to acquire the synchronous electrocardiogram signal and pulse wave signal of the subject under test, and preprocess the synchronous electrocardiogram signal and the pulse wave signal to obtain multiple time window signals, and transmit the multiple time window signals to the signal quality evaluation module. The signal quality assessment module is used to perform quality scoring and gating filtering on multiple time window signals respectively, to obtain multiple time window signals that pass the gating filtering and corresponding quality score results, and to transmit the multiple time window signals to the feature extraction module, while transmitting the corresponding quality score results to the result fusion module. The feature extraction module is used to perform multi-dimensional feature extraction on the time window signals that have passed the gating screening to obtain high-dimensional feature vectors, and transmit the high-dimensional feature vectors to the calibration baseline management module and the intelligent routing and model prediction module at the same time. The calibration baseline management module is used to store and update the high-dimensional feature vector according to the dimension corresponding to the object under test, to obtain personal baseline data, and transmit the personal baseline data to the intelligent routing and model prediction module. The intelligent routing and model prediction module is used to identify the personal baseline data, process the high-dimensional feature vector according to the identification result, obtain the parameter prediction result, and transmit it to the result fusion unit. The result fusion module is used to aggregate the parameter prediction results and the corresponding quality score results to output the target physiological parameter prediction results.

2. The physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 1, characterized in that, The data preprocessing module includes a data access unit and a window segmentation unit; The data access unit is used to acquire the synchronous electrocardiogram signal and pulse wave signal of the object under test and transmit them to the window segmentation unit; The window segmentation unit is used to verify the validity of the synchronized electrocardiogram signal and the pulse wave signal, and to segment the verified signal into multiple time window signals of fixed length and overlapping each other, and to transmit the time window signals to the signal quality evaluation unit.

3. The physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 1, characterized in that, The signal quality assessment module includes a signal filtering unit and a quality assessment unit; The quality assessment unit is used to determine the quality of multiple time window signals, output the quality score of the synchronized electrocardiogram signal and the quality score of the pulse wave signal, and obtain the quality score result based on the quality score of the synchronized electrocardiogram signal and the quality score of the pulse wave signal. The signal filtering unit is used to obtain multiple time window signals that pass the gated filtering by using whether the quality score of the pulse wave signal corresponding to each time window signal is greater than or equal to a preset threshold as a gating filtering condition, and then transmits the multiple time window signals to the feature extraction module, while transmitting the corresponding quality score results to the result fusion module.

4. The physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 1, characterized in that, The feature extraction module includes a basic feature extraction unit, a core feature filtering unit, and a feature vector construction unit; The basic feature extraction unit is used to extract ECG morphological features, ECG heart rate variability features, pulse wave time domain features, pulse wave derivative features, pulse wave frequency domain and entropy features, and ECG and pulse wave interaction features based on multiple time window signals that have been screened by gating, to obtain an initial feature set. The core feature filtering unit is used to filter out features from the initial feature set that have a correlation with blood pressure changes that meet a preset threshold, thereby forming a core feature set; The feature vector construction unit is used to generate inverse features, logarithmic features and binary interaction features according to the core feature set to obtain an extended feature set, and to concatenate the features in the extended feature set to form the high-dimensional feature vector. The high-dimensional feature vector is then transmitted to the calibration baseline management module and the intelligent routing and model prediction module, respectively.

5. A physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 1, characterized in that, The personal baseline data includes personal characteristic baseline, personal systolic blood pressure baseline, personal diastolic blood pressure baseline, and training sample count; The personal feature baseline is generated by averaging all the high-dimensional feature vectors; The individual systolic blood pressure baseline and the individual diastolic blood pressure baseline are generated by averaging the historical systolic blood pressure samples and diastolic blood pressure samples of the subject to be tested, respectively. When the number of new valid samples corresponding to the test object reaches the preset update threshold, the baseline data of the test object and the training sample count are automatically recalculated and updated.

6. The physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 1, characterized in that, The intelligent routing and model prediction module includes an intelligent routing unit, a general prediction unit, and a personalized calibration prediction unit. The intelligent routing unit is used to query whether the object to be tested belongs to a preset individualized object, and to query whether the object to be tested has the personal baseline data; When both conditions are met, the high-dimensional feature vector and personal baseline data corresponding to the test object are routed to the personalized calibration prediction unit: the personalized calibration prediction unit is used to process the high-dimensional feature vector and personal baseline data corresponding to the test object and output the parameter prediction results; When any condition is not met, the high-dimensional feature vector corresponding to the object to be tested is routed to the general prediction unit: the general prediction unit is used to process the high-dimensional feature vector corresponding to the object to be tested and output the parameter prediction result.

7. A physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 6, characterized in that, The general prediction unit is used to sequentially perform median imputation, Z-score standardization, and gradient boosting tree regression operations on the high-dimensional feature vector corresponding to the object to be tested, and output the parameter prediction results.

8. A physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 6, characterized in that, The personalized calibration and prediction unit is used to calculate the difference between the high-dimensional feature vector corresponding to the test object in the current time window and the personal baseline data to obtain the relative feature vector; The relative feature vector is subjected to median imputation of missing values, Z-score standardization, and gradient boosting tree regression to obtain the bias prediction value. The absolute value recovery process is performed on the predicted deviation value and the individual baseline data to obtain the parameter prediction result.

9. A physiological parameter prediction system based on a dual-model architecture and intelligent routing according to claim 1, characterized in that, The result fusion module includes an identification and quality aggregation unit, a data fusion unit, and an anomaly handling unit. The identification and quality aggregation unit is used to group the received parameter prediction results and the corresponding quality score results according to the identifier corresponding to the test object as the grouping index, select the maximum value of the quality score results of all windows in the same group, generate the overall quality assessment result corresponding to a single collection record, and transmit it to the data fusion subunit and the anomaly handling subunit. The data fusion unit is used to perform effective window filtering on the overall quality assessment results, calculate the mean of the parameter prediction results corresponding to the effective window, and generate the target physiological parameter prediction results corresponding to the single collection record. The anomaly handling unit is used to generate an invalid identifier and a signal quality anomaly prompt when all windows in a group fail to pass the valid window filtering, and to replace the target physiological parameter prediction result output.

10. A physiological parameter prediction method based on a dual-model architecture and intelligent routing, characterized in that, The method includes: Acquire synchronized electrocardiogram (ECG) signals and pulse wave signals of the subject under test, and preprocess the synchronized ECG signals and pulse wave signals to obtain multiple time window signals; The multiple time window signals are subjected to quality scoring and gating filtering processes respectively to obtain multiple time window signals that pass the gating filtering and their corresponding quality score results. Multi-dimensional feature extraction is performed on the time window signals that have passed the gating screening to obtain high-dimensional feature vectors; The high-dimensional feature vector is stored and updated according to the dimension corresponding to the object under test to obtain personal baseline data; The personal baseline data is identified, and the high-dimensional feature vector is automatically routed to a general prediction model or a personalized calibration prediction model based on the identification results to obtain parameter prediction results. The predicted results of the parameters are aggregated with the corresponding quality score results to output the predicted results of the target physiological parameters.

Images