Operation hypotension early warning system and prediction method based on multi-modal physiological parameters

Abstract

The invention belongs to the technical field of medical clinic and health monitoring, and discloses an operation hypotension early warning system and prediction method based on multi-modal physiological parameters, and the system comprises a data acquisition module, a feature engineering module, a multi-modal integration and prediction module, a decision support module, a result display module and the like. The feature engineering module constructs a multi-dimensional feature space including key indexes, composite indexes and dynamic indexes based on various real-time physiological parameter data of a patient, the multi-modal integration and prediction module carries out feature screening, constructs optimal feature combinations corresponding to different prediction time windows, and carries out prediction on the optimal feature combinations. And designing two-stage model processing to obtain a hypotension risk prediction result of a plurality of prediction time windows. According to the method, the possibility of hypotension events in multiple time periods in the future in the operation process can be predicted, and clinical intervention suggestions are provided, so that medical personnel can intervene and prevent the hypotension events in advance, and postoperative complications related to hypotension are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical clinical and health monitoring technology, and specifically relates to a surgical hypotension early warning system and prediction method based on multimodal physiological parameters. Background Art

[0002] Intraoperative hypotension is a common clinical phenomenon. A large number of studies have shown that intraoperative hypotension (usually defined as mean arterial pressure MAP < 65 mmHg) is closely associated with postoperative acute kidney injury, myocardial injury, postoperative cognitive dysfunction and increased patient mortality. Even a short intraoperative hypotensive event may cause insufficient organ perfusion in patients, seriously affecting postoperative recovery. A study by Walsh et al. (Anesthesiology, 2013, 119: 507-515) pointed out that an intraoperative MAP below 55 mmHg, even if it lasts only 1-5 minutes, significantly increases the risk of acute kidney injury and myocardial injury. Monk et al. (Anesthesiology, 2015, 123: 307–319) further pointed out that when the intraoperative systolic blood pressure is lower than 70 mmHg and lasts for more than 5 minutes, the patient's 30-day mortality rate is significantly increased, reflecting the significant impact of intraoperative hypotension on patient prognosis. The current clinical methods for predicting and preventing intraoperative hypotension mainly include:

[0003] Traditional empirical judgment method: Anesthesiologists judge the risk of hypotension based on preoperative evaluation, medication status, and degree of surgical stimulation, but there is a lack of objective quantitative predictive indicators.

[0004] Single blood pressure parameter monitoring method: monitor the absolute value or relative baseline changes of systolic blood pressure, diastolic blood pressure or MAP, and passively intervene when blood pressure drops significantly, which cannot achieve early prediction and active prevention and control.

[0005] Machine learning prediction technology: In recent years, some hypotension prediction systems based on machine learning have emerged, such as the Hypotension Prediction Index (HPI) developed by Hatib et al. (Anesthesiology, 2018, 129: 663–674), whose predictive ability has been preliminarily verified by clinical studies; Lee et al. (British Journal of Anaesthesia, 2021, 126: 808–817) developed a multi-channel model based on deep learning, which also showed high predictive performance.

[0006] Hemodynamic monitoring technology: including cardiac output monitoring, pulse wave contour analysis, etc. These methods provide some information on cardiovascular function, but usually require additional invasive procedures or special equipment and have certain limitations in practical applications.

[0007] Although the above-mentioned existing technologies have improved the ability to predict intraoperative hypotension to a certain extent, they still have shortcomings such as a limited prediction time window, reliance on a single information source, lack of individualized adaptability, strong dependence on specialized equipment, insufficient real-time support for clinical decision making, etc. Therefore, there is an urgent need for an intraoperative hypotension prediction and decision support system that integrates multimodal physiological data, achieves a longer prediction time window, higher prediction accuracy, and possesses good practicality, individualized adaptability, and clinical interpretability to meet the urgent clinical need for proactive prevention and control of intraoperative hypotension. Summary of the Invention

[0008] Technical purpose: In response to the above technical problems, the present invention proposes a surgical hypotension early warning system and prediction method based on multimodal physiological parameters. By real-time analysis of multi-dimensional physiological monitoring data, it can predict the possibility of hypotension events in multiple future time periods during the operation and provide corresponding clinical intervention recommendations, so that medical personnel can intervene in advance to prevent the occurrence of hypotension events, reduce postoperative complications related to hypotension, and improve perioperative safety management level and patient prognosis.

[0009] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0010] A surgical hypotension early warning system based on multimodal physiological parameters, comprising:

[0011] Data acquisition module, used to obtain real-time physiological parameter data of patients;

[0012] The feature engineering module constructs a multi-dimensional feature space including key indicators, composite indicators, and dynamic indicators based on preprocessed data;

[0013] The multimodal integration and prediction module performs feature screening based on the multidimensional feature space and constructs an optimal feature combination corresponding to different prediction time windows based on the optimal feature subset obtained by screening. The optimal feature combination is input into a pre-built and trained prediction analysis model to output prediction results of the patient's hypotension risk corresponding to different prediction time windows. The prediction analysis model includes a first-stage model and a second-stage model. The first-stage model includes a parallel gradient boosting tree-based ensemble learning model and a time series feature analysis model, and has independent prediction branches for different time prediction windows.

[0014] a decision support module, configured to perform risk stratification on the prediction results and generate clinical intervention recommendations taking into account the risk stratification results and factors leading to an increased risk of hypotension;

[0015] The result display module is used to display the prediction results and clinical intervention suggestions.

[0016] Preferably, the feature engineering module includes:

[0017] Data preprocessing unit, used to preprocess the patient's real-time physiological parameter data, including data validity check, multi-source data completion, and time series standardization;

[0018] The basic feature calculation unit is used to detect outliers and smooth the data after preprocessing, and normalize the key indicators, that is, the key monitored physiological parameters;

[0019] The derived feature calculation unit is used to obtain the output results of the basic feature calculation unit and construct a composite indicator reflecting the patient's real-time health status;

[0020] The dynamic feature calculation unit is used to calculate the short-term change characteristics of multiple key indicators in multiple continuous time windows and construct multiple dynamic indicators that reflect the patient's real-time health status.

[0021] Preferably, the multimodal integration and prediction module includes:

[0022] The feature selection unit is used to perform feature screening based on the multi-dimensional feature space to obtain the optimal feature subset and construct the optimal feature combination for different prediction time windows;

[0023] A predictive analysis unit, including a first-stage model and a second-stage model constituting a predictive analysis model using a two-stage hybrid model architecture;

[0024] The first-stage model includes a parallel ensemble learning model based on a gradient boosting tree and a time series feature analysis model based on a Transformer. The inputs of the ensemble learning model based on a gradient boosting tree and the time series feature analysis model based on a Transformer are both the optimal feature combination output by the input feature selection unit, and the outputs are weighted and fused to obtain a preliminary prediction result.

[0025] The second-stage model uses a Transformer filter to input the preliminary prediction results output by the first-stage model, conducts refined screening of possible false positive results that may appear in the prediction of the first-stage model, and outputs the final prediction results of hypotension risk in multiple different time windows.

[0026] Preferably, the system further comprises a model fine-tuning unit, which is configured to implement differentiated fine-tuning strategies for different types of model parameters of the multimodal fusion and prediction module, including:

[0027] Fine-tuning of feature importance weights: Based on the distribution characteristics of local data, the importance weight coefficient of each feature in the feature selection unit is adaptively adjusted to make each feature reflect the difference in feature sensitivity of the local patient population;

[0028] Layered fine-tuning of prediction model parameters: For the first-stage model, a freeze-and-thaw progressive fine-tuning strategy is adopted. First, the deep network parameters are kept fixed, and only the decision layer parameters of the last few layers are adjusted. As local data accumulates, the intermediate layer parameters are gradually unfrozen for more comprehensive adaptive adjustments. For the decision tree-based ensemble model, the leaf node weights and decision thresholds are adjusted. For the deep learning-based time series analysis model, the upper feature extraction layer and attention mechanism weights are fine-tuned.

[0029] Full update of calibration layer parameters: A full parameter fine-tuning strategy is adopted for the second-stage model. The fine-tuning process adopts an incremental parameter update strategy, combined with early stopping mechanism, regularization technology and parameter rollback mechanism.

[0030] Preferably, the system further includes an individualized adjustment unit, which is configured to provide individualized adjustment strategies for different types of model parameters of the multimodal fusion and prediction module, including:

[0031] Patient medical record information: A multi-source patient information integration mechanism is used to automatically extract and integrate key information from electronic medical records, including age, gender, ASA classification, baseline blood pressure level, cardiac function assessment results, and comorbidities;

[0032] Patient risk warning threshold: A differentiated sensitivity adjustment algorithm is used to automatically adjust the warning threshold based on patient risk stratification;

[0033] Tailored intervention plans: The intensity, timing, and specific measures of intervention recommendations are tailored based on individual patient characteristics.

[0034] Preferably, the decision support module includes:

[0035] The risk grading unit uses a dynamic threshold division algorithm to map the predicted results, i.e. the predicted risk probability, to the clinical decision space, dividing the risk into multiple levels. It also uses a visual risk heat map and trend curve to present the risk evolution trajectory, in conjunction with an audio-visual warning mechanism;

[0036] The intervention recommendation unit adopts a multi-level reasoning architecture based on the combination of rules and knowledge graphs to generate and display clinical intervention recommendations with clear priority and timeliness markings.

[0037] Preferably, the result display module includes:

[0038] A real-time display unit displays the patient's current risk level and the risk probability value for each predicted time window in digital and color-coded form;

[0039] Trend chart unit, used to display the time trend of key parameters and risk probability, and clearly mark the warning trigger points;

[0040] Alarm control unit, used to trigger different levels of visual and audible alarms based on risk levels, and support medical staff in confirming and pausing operations.

[0041] A method for early warning of surgical hypotension based on multimodal physiological parameters comprises the following steps:

[0042] Used to obtain real-time physiological parameter data of patients;

[0043] Based on the preprocessed data, a multi-dimensional feature space including key indicators, composite indicators and dynamic indicators is constructed;

[0044] Based on the multi-dimensional feature space, feature screening is performed, and based on the optimal feature subset obtained by screening, an optimal feature combination corresponding to different prediction time windows is constructed; the optimal feature combination is input into a pre-built and trained prediction analysis model, and prediction results of the patient's hypotension risk corresponding to different prediction time windows are output;

[0045] Performing risk stratification on the predicted results, and generating clinical intervention recommendations by considering the risk stratification results and factors leading to an increased risk of hypotension;

[0046] The prediction results and clinical intervention recommendations are displayed.

[0047] Preferably, the following steps are performed to construct a multi-dimensional feature space including key indicators, composite indicators, and dynamic indicators:

[0048] Preprocessing of patients' real-time physiological parameter data, including data validity check, multi-source data completion, and time series normalization;

[0049] Perform outlier detection and data smoothing on the pre-processed data, and normalize the key indicators, i.e. the key monitored physiological parameters;

[0050] Construct a composite indicator that reflects the patient's real-time health status;

[0051] Calculate the short-term change characteristics of multiple key indicators in multiple continuous time windows and construct multiple dynamic indicators that reflect the patient's real-time health status.

[0052] Preferably, the first-stage model and the second-stage model of the prediction analysis model, the first-stage model includes a parallel ensemble learning model based on a gradient boosting tree and a time series feature analysis model based on a Transformer, and the optimal feature combination and output of the ensemble learning model based on the gradient boosting tree and the time series feature analysis model based on the Transformer are weightedly fused to obtain a preliminary prediction result;

[0053] The second-stage model uses a Transformer filter to input the preliminary prediction results output by the first-stage model, conducts refined screening of possible false positive results that may appear in the prediction of the first-stage model, and outputs the final prediction results of hypotension risk in multiple different time windows.

[0054] Beneficial effects: The surgical hypotension warning system and prediction method based on multimodal physiological parameters proposed in the present invention can overcome many defects and shortcomings in the existing technology. It integrates multiple basic physiological parameters and derived features to construct a comprehensive multi-dimensional feature matrix, and realizes hypotension prediction in multiple time prediction windows through a two-stage hybrid model architecture, maintaining a high prediction accuracy of each time prediction window, and greatly reducing the false alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a system software architecture diagram provided by the first embodiment of the present invention;

[0056] Figure 2 It is a system architecture diagram and flow chart. DETAILED DESCRIPTION

[0057] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0058] Example 1

[0059] 1. System Hardware and Software Architecture

[0060] The surgical hypotension warning system based on multimodal physiological parameters of the present invention can be implemented using the following hardware and software architecture:

[0061] The system hardware mainly includes data acquisition unit, computing processing unit and display interaction unit.

[0062] The data acquisition unit is connected to the physiological parameter monitoring equipment in the operating room through the RS232 serial port, USB interface or Ethernet interface, such as Philips IntelliVue, GE CARESCAPE or Mindray BeneView series monitors.

[0063] The computing processing unit can be a general-purpose computer with at least an Intel i5 or equivalent processor, 8GB of memory and 128GB of storage space, or a dedicated medical-grade touch-screen all-in-one computer. In resource-constrained scenarios, an embedded ARM architecture processor (such as Raspberry Pi 4 or later) can also be used.

[0064] The display interaction unit is a 12-19 inch medical-grade touch screen, with an optional keyboard and mouse as auxiliary input devices.

[0065] System software architecture such as Figure 1 shown.

[0066] The surgical hypotension warning system based on multimodal physiological parameters proposed in the present invention is mainly composed of a data acquisition module, a feature engineering module, a multimodal fusion and prediction module, a decision support module and a result display module. The modules work together to form a complete hypotension prediction and warning closed loop.

[0067] 1. The data acquisition module is the fundamental component of the system, responsible for acquiring real-time patient physiological parameter data from operating room monitoring equipment. This module includes a standard interface unit. The standard interface unit supports data interaction with mainstream monitoring equipment on the market, is compatible with standard medical data protocols such as HL7 and DICOM, and can communicate via serial port, Ethernet, or wirelessly. A key technical feature of this invention is that the data acquisition frequency is set to once per minute. This design greatly reduces the system's hardware requirements, enabling it to run on existing equipment in most medical institutions without the need for additional high-performance hardware support.

[0068] 2. The feature engineering module is one of the key innovations of this invention, which adopts a multi-level, progressive data processing process. First, the system pre-processes the collected data through the data pre-processing unit and establishes a complete data quality management system, including the following steps:

[0069] (1) Data validity check, and abnormal processing of data values that exceed the physiological reasonable range;

[0070] (2) Multi-source data completion: when invasive arterial pressure monitoring data is missing, non-invasive blood pressure monitoring data can be intelligently called to supplement it;

[0071] (3) Time series standardization: data with different sampling frequencies are uniformly processed into one record per minute in order to achieve the possibility of wide application. For multiple records within the same minute, a priority strategy is used to select the most reliable data point.

[0072] On this basis, the system constructs a multi-dimensional feature space through the collaborative work of three feature calculation units.

[0073] The basic feature calculation unit not only performs conventional outlier detection and data smoothing but also incorporates specialized preprocessing algorithms tailored to anesthesia and surgical scenarios, capable of identifying and filtering data anomalies caused by surgically specific interference factors (such as electrosurgical unit use and changes in body position). The system monitors key physiological parameters including heart rate, mean arterial pressure, central venous pressure, end-tidal carbon dioxide partial pressure, blood oxygen saturation, diastolic blood pressure, and systolic blood pressure. These indicators undergo a specially designed normalization process to ensure comparability across parameters of different dimensions.

[0074] Based on the knowledge of anesthesia experts, the derived feature calculation unit has constructed a series of composite indicators that can reflect hemodynamic instability at an early stage, such as the myocardial stress index (MSI = heart rate / mean arterial pressure), perfusion pressure gradient (PPG = mean arterial pressure - central venous pressure), pulse pressure time index (PIT = (1 / pulse) * (diastolic pressure / systolic pressure)), cardiorespiratory index (HRI = (systolic pressure - diastolic pressure) / heart rate * respiratory rate), and shock index correction value (SHOCK = mean arterial pressure / systolic pressure). These indicators have demonstrated early warning capabilities that surpass traditional vital signs in clinical validation.

[0075] The dynamic feature calculation unit uses a segmented continuous calculation method, creating secondary groupings based on cases and continuous time intervals. Within each continuous time window, the short-term variation characteristics of each parameter are calculated. A case refers to a continuous data collection process, such as a surgery or clinical monitoring session. Secondary grouping groups the collected data by case and continuous time. If the minutes are not continuous, the collected data will be grouped. For example, if data from the 1st, 2nd, 3rd, 5th, 6th, and 7th minutes of a surgery are collected, the data from the 1st, 2nd, and 3rd minutes will be grouped together, and the data from the 5th, 6th, and 7th minutes will be grouped together. The system innovatively designs multiple dynamic indicators, including the 5-minute rate of change of ETCO2, heart rate, mean arterial pressure, and pulse (CR5M series of indicators), as well as the 5-minute average rate of change of mean arterial pressure (MAP_ACR5M), 5-minute maximum negative slope (MAP_MNS5M), 5-minute sliding standard deviation (MAP_ST5M), and trend direction (MAP_TD3M). In addition, the dynamic characteristics related to pulse pressure (PP_C5M, PP2MAP, PP_MAP_RC5M) and the changing characteristics of the heart rate to blood pressure ratio (MAP2HR_CR5M, MAP2HR_EWM) were also calculated.

[0076] 3. The multimodal fusion and prediction module is the core of the present invention. It uses a unique hierarchical and progressive hybrid model architecture to achieve accurate prediction of surgical hypotension. This module first uses a feature selection unit to determine the optimal feature set using a double screening strategy:

[0077] First, a feature pre-screening based on domain knowledge is performed, where anesthesia experts identify potentially relevant features based on their clinical experience. This is followed by a computer-assisted feature importance analysis to achieve final feature optimization. The specific feature screening method is as follows:

[0078] (1) Use all features to train the initial model;

[0079] (2) Calculate the contribution of each feature to the prediction result;

[0080] (3) Eliminate the features with the lowest importance ranking;

[0081] (4) Retrain the model using the remaining features and evaluate the performance

[0082] (5) Repeat the above process until the optimal feature subset is reached.

[0083] For different prediction time windows (e.g., 5-minute, 10-minute, and 15-minute), the system constructs the optimal feature combination, fully considering the specificity of physiological parameter changes at each time scale. The feature value calculation range varies for different prediction time windows. For example, for a 5-minute time window, the "map change rate in the last 5 minutes" is used, while for a 10-minute time window, the "map change rate in the last 10 minutes" is used.

[0084] The predictive analysis unit utilizes an innovative two-stage hybrid model for prediction. The first stage combines the strengths of a decision tree-based ensemble learning model and a deep learning-based time series analysis model. The former focuses on capturing nonlinear relationships and threshold effects between features, and is highly sensitive to potential critical value changes in each parameter. The latter specifically processes time series features, capable of identifying subtle patterns and trends in physiological parameters over time. The system utilizes a weighted voting mechanism and meta-model training to integrate the outputs of the two models, adaptively adjusting the contribution weight of each model in different scenarios.

[0085] The second-stage model specifically screens for potential false-positive results from the first stage. By continuously learning from historical prediction performance, the system can identify patterns of factors that may lead to false positives in specific cases or time periods, significantly improving the system's reliability in complex clinical environments. Furthermore, the system incorporates a contextual calibration mechanism tailored to different anesthesia depths and surgical types, enabling dynamic adjustment of warning thresholds based on specific surgical scenarios.

[0086] This module ultimately outputs hypotension risk probability values for three different time windows: P5 (5-minute window), P10 (10-minute window) and P15 (15-minute window), all with a value range of 0-1. The risk level is intuitively presented through color coding, providing clinicians with sufficient warning time and clear risk assessment.

[0087] 4. In addition, the present invention also designs a model fine-tuning unit, which achieves efficient adaptation and performance optimization of the system across different medical institutions through refined parameter adjustment strategies. This unit adopts a hierarchical selective transfer learning framework and implements differentiated fine-tuning strategies for different types of model parameters:

[0088] (1) Feature Importance Weight Fine-tuning: The system adaptively adjusts the importance weight coefficients of each feature in the feature selection unit based on the distribution characteristics of local data, so that it reflects the differences in feature sensitivity among local patient populations. The fine-tuning algorithm uses Bayesian optimization methods to find the local optimal weight configuration within a relatively small number of iterations.

[0089] (2) Layered fine-tuning of prediction model parameters: For the first-stage hybrid model, the system adopts a freeze-thaw progressive fine-tuning strategy. First, the deep network parameters are kept fixed, and only the decision layer parameters of the last few layers are adjusted; as local data accumulates, the intermediate layer parameters are gradually unfrozen for more comprehensive adaptive adjustments. Specifically, for tree-based ensemble models, the main adjustments are made to the leaf node weights and decision thresholds; for time series deep learning models, the focus is on fine-tuning the upper feature extraction layer and the attention mechanism weights.

[0090] (3) Comprehensive update of calibration layer parameters: The second-stage model responsible for false alarm filtering adopts a more aggressive full-parameter fine-tuning strategy to quickly adapt to the specific alarm preferences and risk tolerance of local medical institutions.

[0091] The fine-tuning process employs an incremental parameter update strategy, combined with an early stopping mechanism and regularization techniques to prevent overfitting to local data. The system incorporates a data quality assessment module to ensure that only high-quality local surgical records are used for model fine-tuning. Initial fine-tuning requires a minimum of 100 diverse local surgical records, and the system automatically expands the fine-tuning depth as data accumulates. Furthermore, the fine-tuning unit includes a parameter rollback mechanism that immediately restores the previously stable configuration upon detecting performance degradation, ensuring reliable system operation under various conditions.

[0092] 5. The decision support module converts prediction results into specific clinical intervention recommendations, enhancing the practical value of the system. This module achieves a complete closed loop from risk assessment to precise intervention through the collaborative work of three functional units:

[0093] First, the risk stratification unit uses a dynamic thresholding algorithm to map predicted risk probabilities to the clinical decision space, with default stratification into three levels: low risk (probability <0.3), medium risk (probability 0.3-0.7), and high risk (probability >0.7). The system uses visual risk heat maps and trend curves to intuitively display the trajectory of risk evolution, and incorporates audio and visual warning mechanisms to ensure efficient clinical perception.

[0094] Secondly, the intervention recommendation unit utilizes a multi-level reasoning architecture that combines rules and knowledge graphs. This architecture not only considers the risk level but also analyzes the primary contributing factors to the increased risk of hypotension. By deeply analyzing the interpretable output of the predictive model, the system can identify factors such as volume depletion, decreased vascular tone, weakened myocardial contractility, and surgical stimulation that contribute to the increased risk, and then generate targeted clinical intervention recommendations. These recommendations are clearly marked with priority and timeliness, assisting anesthesiologists in optimizing their decisions.

[0095] One of the core innovations of the present invention is the individualized adjustment unit, which uses a multi-dimensional patient characteristic adaptive framework to achieve precise individualization of the early warning system. This unit has the following technical features: (1) a multi-source patient information integration mechanism that automatically extracts and integrates key information from electronic medical records, including age, gender, ASA classification, baseline blood pressure level, cardiac function assessment results, comorbidities, etc.; (2) a differentiated sensitivity adjustment algorithm that automatically adjusts the early warning threshold according to patient risk stratification, such as using more sensitive early warning settings for elderly patients or patients with heart failure; (3) a parameterized intervention plan generation system that customizes the intensity, timing, and specific measures of intervention recommendations based on individual patient characteristics, such as emphasizing a micro-fluid rehydration strategy for volume-sensitive patients and recommending moderate positive inotropic support for patients with limited cardiac reserve.

[0096] The Personalized Adjustment Unit also features a self-learning optimization module that continuously adjusts and optimizes its decision-making logic by recording and analyzing anesthesiologists' adoption of system recommendations and their subsequent effects. This module utilizes a reinforcement learning framework, using clinical intervention effectiveness as feedback to continuously improve system performance and achieve truly personalized and precise prediction and intervention.

[0097] 6. The results display module serves as the interface between the system and medical staff. Through a real-time display unit, trend chart unit, and alarm control unit, it presents complex prediction results in an intuitive and easy-to-understand manner. The real-time display unit displays the current risk level and risk probability values for each time window in numerical and color-coded form. The trend chart unit shows the time-varying trends of key parameters and risk probabilities, and clearly marks the warning trigger points. The alarm control unit triggers different levels of visual and audible alarms based on the risk level and supports medical staff in confirming and pausing operations to avoid alarm fatigue.

[0098] 2. Detailed implementation method of the system

[0099] The data collection frequency of this system is set to be collected every minute, such as once every minute. This design significantly reduces the system's hardware requirements, allowing it to run stably on ordinary computing devices. The specific implementation method of data collection is as follows:

[0100] Algorithm 1: Data collection and preprocessing algorithm

[0101] Input: Raw physiological parameters obtained from monitoring equipment

[0102] Output: Preprocessed and standardized data

[0103] 1. Initialize system parameters and establish a connection with the monitoring device.

[0104] 2. Execute a data collection cycle every 1 minute:

[0105] 2.1. Read the current physiological parameter value through the selected interface;

[0106] 2.2. Check data integrity and mark missing values;

[0107] 2.3. Judge the outliers:

[0108] If the parameter value exceeds the preset physiological range, it is marked as an abnormal value;

[0109] Physiological range setting:

[0110] -Heart rate: 30-180 beats / min

[0111] -MAP:30-150mmHg

[0112] -ETCO2:15-70mmHg

[0113] -Blood oxygen saturation: 70%-100%

[0114] Other parameters were set according to the routine clinical range;

[0115] 2.4. Handling of missing values and outliers:

[0116] If the number of consecutive missing sampling points does not exceed a preset number, for example, no more than 3 sampling points, linear interpolation is used;

[0117] If the outlier is a single-point fluctuation, median filtering is used;

[0118] 2.5. Convert the processed data into a standard format, including timestamp, parameter ID, parameter value, and quality flag.

[0119] 3. The data is stored in a temporary buffer, retaining the most recent data, such as 30 minutes of data for feature calculation:

[0120] The feature engineering module calculates basic features, derived features, and dynamic features based on the collected original physiological parameters:

[0121] Algorithm 2: Feature calculation algorithm

[0122] Input: preprocessed and standardized data

[0123] Output: Feature vector for prediction

[0124] 1. Basic feature calculation:

[0125] 1.1. Normalization: Z-score normalization is applied to all parameters.

[0126] 2. Derived feature calculation:

[0127] 2.1. Calculate the myocardial stress index (MSI) = HEART_RATE / MAP.

[0128] Calculate the perfusion pressure gradient (PPG) = MAP – CVP.

[0129] 2.3. Calculate the pulse pressure time index (PIT) = (1 / PULSE) * (PDBP / PSBP);

[0130] 2.4. Calculate the cardiorespiratory index (HRI) = (PSBP - PDBP) / HEART_RATE * BREATH;

[0131] 2.5. Calculate the corrected shock index: SHOCK = MAP / PSBP.

[0132] 2.6. Calculate the pulse pressure to MAP ratio PP2MAP = (PSBP - PDBP) / MAP.

[0133] 3. Dynamic feature calculation:

[0134] 3.1. Calculate the change rate of each parameter in the last 5 minutes:

[0135] Δ value = (current value - value 5 minutes ago) / value 5 minutes ago

[0136] 3.2. Calculate the sliding standard deviation:

[0137] Select the latest 5 minutes of data to calculate the standard deviation;

[0138] 3.3. Calculate the trend direction:

[0139] MAP_TD3M=-1 (continuous decrease for 3 minutes)

[0140] 1 (3 minutes of continuous ascent)

[0141] 0 (other cases)

[0142] 3.4. Calculate the exponentially weighted average:

[0143] MAP2HR_EWM is calculated using an attenuation factor of 0.8.

[0144] 4. Combine all features to form a feature vector:

[0145] The predictive analysis unit adopts an innovative two-stage hybrid architecture, and its algorithm is implemented as follows:

[0146] Algorithm 3: Two-stage adaptive hybrid prediction model

[0147] Input: multidimensional feature vector X = {X_static, X_temporal}, where X_static is the static feature and X_temporal is the time series feature

[0148] Output: Hypotension risk probability P5, P10, P15 in three time windows

[0149] Phase 1: Dual-model parallel prediction and fusion

[0150] 1. Integrated learning model processing based on gradient boosting tree:

[0151] This model specifically processes structured feature data and has the following characteristics:

[0152] 1.1. Adopt a hierarchical feature processing mechanism and apply differentiated feature transformation and normalization to different types of features (derived features, dynamic features, etc.);

[0153] 1.2. Capture nonlinear relationships between features and automatically discover complex threshold effects and feature interaction patterns through tree structures;

[0154] 1.3. It has a regularization mechanism tailored to the characteristics of medical data, effectively preventing overfitting and enhancing the model's robustness to noise;

[0155] 1.4. Build an independent prediction branch for each prediction window (5 minutes, 10 minutes, 15 minutes) and generate preliminary prediction probabilities P_tree_5, P_tree_10, and P_tree_15.

[0156] 2. Transformer-based time series feature analysis model processing:

[0157] The model specifically captures the time series patterns of physiological parameters and has the following characteristics:

[0158] 2.1. Using a multi-head self-attention mechanism to capture patterns at different time scales in parallel;

[0159] 2.2. Strengthen the sequential information of time series through position encoding while preserving long-term dependencies;

[0160] 2.3. Introducing skip connections and residual learning mechanisms to ensure stable gradient propagation and long-sequence information preservation;

[0161] 2.4. Design a dedicated time masking mechanism to address irregular sampling and missing issues in medical data;

[0162] 2.5. Similarly generate independent prediction probabilities P_trans_5, P_trans_10, P_trans_15 for the three prediction windows

[0163] 3. Multi-model fusion mechanism:

[0164] 3.1. Dynamic Weighted Ensemble: Unlike simple weighted averaging, the system dynamically adjusts the weights of each model based on the characteristics of the current feature vector;

[0165] P_stage1_i=α_i·P_tree_i+β_i·P_trans_i(i=5,10,15)

[0166] The weight coefficients α_i and β_i are determined by the quality indicators of the feature vector (such as temporal integrity and parameter stability);

[0167] 3.2. Confidence correction: Make conservative adjustments to low-confidence prediction results to improve system reliability.

[0168] [Phase 2: Refined false positive filtering]

[0169] 4. Advanced Transformer Filters:

[0170] Different from the first-stage Transformer model, the second-stage model has the following characteristics:

[0171] 4.1. The input data has a higher dimensionality: it contains not only the original features but also the prediction results and confidence indicators from the first stage.

[0172] 4.2. More focused attention mechanism: Adopting a guided attention mechanism, it pays special attention to potential false positive patterns;

[0173] 4.3. More refined feature extraction: Including multi-scale convolutional network layers, it can capture more fine-grained waveform features and trend changes;

[0174] 4.4. Reinforcement Learning Component: Optimizes filtering decision strategies through continuous learning from past false positive cases;

[0175] 5. False positive risk assessment:

[0176] 5.1. Perform a detailed analysis of all risk predictions exceeding the threshold in the first stage (P_stage1_i>θ_i);

[0177] 5.2. Calculate a correction factor C_i for each medium- and high-risk prediction, which represents the confidence that the prediction is a true event;

[0178] 5.3. Correction factor calculations take into account patient-specific factors, surgical stage characteristics, and historical prediction performance;

[0179] 6. Final prediction results are generated:

[0180] Apply corrections to medium and high risk predictions: P_final_i = P_stage1_i · C_i;

[0181] 6.2. For low-risk predictions, remain unchanged: P_final_i = P_stage1_i (when P_stage1_i ≤ θ_i);

[0182] 6.3. Enhanced visualization through color coding and risk level marking;

[0183] 6.4. Output the final hypotension risk probabilities P5, P10, and P15 for the three time windows;

[0184] 7. Context-adaptive calibration:

[0185] 7.1. Dynamically adjust the final risk assessment based on clinical contextual factors such as surgical stage and depth of anesthesia;

[0186] 7.2. Implement personalized early warning strategies for different clinical scenarios.

[0187] 3. Incremental Learning and Local Adaptation Implementation Methods

[0188] To validate the system's incremental learning capabilities and local adaptability, this example was implemented on the VitalDB open dataset. The system was first deployed using a general pre-trained model. Then, local model fine-tuning was performed using 100 surgical cases collected from the VitalDB dataset. Results showed that local data fine-tuning significantly improved system performance across various medical environments.

[0189] The overall architecture of the system is as follows Figure 2 shown.

[0190] The surgical hypotension early warning system and prediction method based on multimodal physiological parameters proposed in this invention addresses several defects and deficiencies in the existing technology and achieves the following significant advantages through innovative technical design:

[0191] (1) Extended prediction time window

[0192] In response to the defect of limited prediction time window in the existing technology, the present invention adopts a two-stage hybrid model architecture to realize multiple prediction time windows, such as hypotension prediction in three time windows of 5 minutes, 10 minutes and 15 minutes. Through the collaborative work of XGBoost and Transformer models, the system can capture short-term fluctuations and long-term trend characteristics at the same time, maintaining high prediction accuracy for each time window. Experimental data show that the accuracy of this system in the 15-minute prediction window can still be maintained at above 85%, which is a significant improvement over the 70% of the existing technology, thus gaining more preparation time for clinical intervention and fundamentally reducing the risk of hypotension-related complications.

[0193] (2) Multimodal information fusion reduces false alarm rate

[0194] To address the high false alarm rate caused by existing technologies' reliance on a single information source, this invention integrates 12 basic physiological parameters and over 20 derived features to construct a comprehensive multidimensional feature matrix. In particular, the second-stage Transformer model meticulously screens for false positives, significantly reducing the false alarm rate. Practical verification has shown that the positive predictive value of this system is significantly higher than the approximately 13%-15% level of existing systems, effectively reducing alarm fatigue among medical staff and improving the system's clinical acceptability.

[0195] (3) Lowering the threshold for technology application

[0196] Addressing the limitations of existing technologies that require additional specialized equipment, this invention utilizes minute-by-minute data acquisition, significantly reducing the performance requirements of monitoring equipment. It seamlessly connects to existing monitoring equipment through standard medical data interfaces (such as HL7 and DICOM), enabling deployment without additional hardware investment. This design enables widespread adoption in resource-limited medical institutions, effectively lowering the economic and technical barriers to technology adoption.

[0197] (4) Individualized prediction and adaptability

[0198] To address the lack of individualized adaptability in existing technologies, this paper designs an individualized adjustment unit and a model fine-tuning unit. These units automatically adjust warning thresholds based on preoperative patient assessment information. Using transfer learning technology, the system adapts to the specific characteristics of different medical institutions. Experiments have shown that through individualized parameter adjustment, this system can improve prediction accuracy by 3%-5%, making it a true precision medicine system suitable for different populations and surgical procedures.

[0199] (5) Complete clinical decision support

[0200] To address the inadequacy of existing clinical decision support technologies, this invention not only provides early warnings of hypotension risk but also generates specific clinical intervention recommendations based on risk level and key risk factors. Through a specialized decision support process, the system provides medical staff with specific recommendations for volume status assessment, vascular tone regulation, and myocardial contractility support, forming a complete "prediction-diagnosis-intervention" closed-loop management system, significantly enhancing the system's clinical utility.

[0201] (6) Enhanced interpretability

[0202] To address the lack of interpretability in existing technologies, the present invention employs a feature importance analysis method guided by medical knowledge within a feature selection unit, providing clear clinical interpretation for prediction results. The system not only provides risk probability values but also identifies the primary factors contributing to increased risk, enabling healthcare professionals to understand and trust the system's predictions and improving clinical adoption. Practice has demonstrated that this enhanced interpretability has increased healthcare professionals' response rate to system warnings by approximately 30%, leading to more effective prevention of hypotensive events.

[0203] (7) Higher real-time performance and stability

[0204] To address the non-real-time nature of existing technologies, this invention optimizes the algorithm structure and data processing flow, ensuring the system can analyze and deliver predictions within three seconds of receiving new data, meeting the demands of immediate predictions in operating room environments. Furthermore, the system's two-stage architecture enhances prediction stability, reduces misjudgments caused by transient fluctuations, and makes clinical warnings more reliable.

[0205] (8) Wide clinical applicability

[0206] This system, based on a model trained using large-scale multicenter data, is applicable to all major surgical procedures, including but not limited to general surgery, orthopedics, neurosurgery, and obstetrics and gynecology. The system's design takes into account the characteristics of patients of different ages and underlying conditions. Combined with personalized adjustment capabilities, it can provide accurate hypotension warnings for a wide range of patient groups, demonstrating its high clinical potential.

[0207] Example 2

[0208] Based on the above system architecture, the surgical hypotension prediction method proposed in the present invention specifically includes the following process:

[0209] First, the system collects the patient's real-time physiological parameters at a frequency of 1 minute through data acquisition and preprocessing steps, and performs data smoothing. It uses sliding median filtering to remove outliers and forward fills or linearly interpolates missing values to ensure the continuity and reliability of the data.

[0210] Then, in the feature calculation and conversion step, the system calculates the statistical characteristics of the basic physiological parameters (mean, variance, extreme values, etc.), calculates the derived features based on medical knowledge, and extracts time series features and trend features to construct a complete feature vector.

[0211] In the multi-stage prediction step, the system uses a two-stage hybrid model architecture for prediction. In the first stage, the XGBoost model and the Transformer model are used simultaneously: the XGBoost model analyzes static and short-term dynamic features, and the Transformer model analyzes long-term time series features. The outputs of the two are weighted fusion to generate preliminary prediction results. In the second stage, the Transformer model is used specifically to identify and screen the possible false positive results in the first stage, significantly reducing the false alarm rate. The system ultimately generates multiple prediction time windows, such as the risk probability of hypotension in 5-minute, 10-minute, and 15-minute windows.

[0212] In the risk assessment and decision support steps, the system determines the risk level based on the risk probability value, analyzes the main factors leading to high risk, and generates individualized intervention recommendations to provide clear clinical decision support for medical staff.

[0213] Finally, in the early warning result display step, the system displays the prediction results in a graphical interface, issues an alarm when the risk exceeds the preset threshold, and displays corresponding clinical intervention recommendations to ensure that medical staff can obtain necessary information in a timely manner and take appropriate intervention measures.

[0214] The prediction model of the present invention ensures its performance and reliability through a rigorous training and validation process. First, a large-scale multicenter dataset containing data from patients with different types of surgery, different age groups, and different underlying diseases is constructed. Then, the data is annotated according to the definition of hypotensive events (MAP<65mmHg for at least 1 minute), and all the features involved in the present invention are extracted. In the model training phase, the XGBoost model and the Transformer model are trained separately, and the model performance is evaluated by cross-validation. Afterwards, the model fusion mechanism is optimized through the validation set to maximize the comprehensive prediction performance. Finally, the model performance is evaluated on an independent test set to ensure that the prediction accuracy, sensitivity, and specificity of each time window meet the expected goals.

[0215] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, apparatus, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0216] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (apparatus), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as a combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0217] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0218] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0219] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A surgical hypotension warning system based on multimodal physiological parameters, characterized in that: include: Data acquisition module, used to obtain real-time physiological parameter data of patients; The feature engineering module constructs a multi-dimensional feature space including key indicators, composite indicators, and dynamic indicators based on preprocessed data; The multimodal integration and prediction module performs feature screening based on the multidimensional feature space and constructs an optimal feature combination corresponding to different prediction time windows based on the optimal feature subset obtained by screening. The optimal feature combination is input into a pre-built and trained prediction analysis model to output prediction results of the patient's hypotension risk corresponding to different prediction time windows. The prediction analysis model includes a first-stage model and a second-stage model. The first-stage model includes a parallel gradient boosting tree-based ensemble learning model and a time series feature analysis model, and has independent prediction branches for different time prediction windows. a decision support module, configured to perform risk stratification on the prediction results and generate clinical intervention recommendations taking into account the risk stratification results and factors leading to an increased risk of hypotension; The result display module is used to display the prediction results and clinical intervention suggestions.

2. The surgical hypotension early warning system based on multimodal physiological parameters according to claim 1, characterized in that: The feature engineering module includes: Data preprocessing unit, used to preprocess the patient's real-time physiological parameter data, including data validity check, multi-source data completion, and time series standardization; The basic feature calculation unit is used to detect outliers and smooth the data after preprocessing, and normalize the key indicators, that is, the key monitored physiological parameters; The derived feature calculation unit is used to obtain the output results of the basic feature calculation unit and construct a composite indicator reflecting the patient's real-time health status; The dynamic feature calculation unit is used to calculate the short-term change characteristics of multiple key indicators in multiple continuous time windows and construct multiple dynamic indicators that reflect the patient's real-time health status.

3. The surgical hypotension early warning system based on multimodal physiological parameters according to claim 1, characterized in that: The multimodal integration and prediction module includes: The feature selection unit is used to perform feature screening based on the multi-dimensional feature space to obtain the optimal feature subset and construct the optimal feature combination for different prediction time windows; A predictive analysis unit, including a first-stage model and a second-stage model constituting a predictive analysis model using a two-stage hybrid model architecture; The first-stage model includes a parallel ensemble learning model based on a gradient boosting tree and a time series feature analysis model based on a Transformer. The inputs of the ensemble learning model based on a gradient boosting tree and the time series feature analysis model based on a Transformer are both the optimal feature combination output by the input feature selection unit, and the outputs are weighted and fused to obtain a preliminary prediction result. The second-stage model uses a Transformer filter to input the preliminary prediction results output by the first-stage model, conducts refined screening of possible false positive results that may appear in the prediction of the first-stage model, and outputs the final prediction results of hypotension risk in multiple different time windows.

4. The surgical hypotension early warning system based on multimodal physiological parameters according to claim 3, characterized in that: The system further includes a model fine-tuning unit, which is used to implement differentiated fine-tuning strategies for different types of model parameters of the multimodal fusion and prediction module, including: Fine-tuning of feature importance weights: Based on the distribution characteristics of local data, the importance weight coefficient of each feature in the feature selection unit is adaptively adjusted to make each feature reflect the difference in feature sensitivity of the local patient population; Layered fine-tuning of prediction model parameters: For the first-stage model, a freeze-and-thaw progressive fine-tuning strategy is adopted. First, the deep network parameters are kept fixed, and only the decision layer parameters of the last few layers are adjusted. As local data accumulates, the intermediate layer parameters are gradually unfrozen for more comprehensive adaptive adjustments. For the decision tree-based ensemble model, the leaf node weights and decision thresholds are adjusted. For the deep learning-based time series analysis model, the upper feature extraction layer and attention mechanism weights are fine-tuned. Full update of calibration layer parameters: A full parameter fine-tuning strategy is adopted for the second-stage model. The fine-tuning process adopts an incremental parameter update strategy, combined with early stopping mechanism, regularization technology and parameter rollback mechanism.

5. The surgical hypotension early warning system based on multimodal physiological parameters according to claim 3, characterized in that: The system further includes an individualized adjustment unit configured to provide individualized adjustment strategies for different types of model parameters of the multimodal fusion and prediction module, including: Patient medical record information: A multi-source patient information integration mechanism is used to automatically extract and integrate key information from electronic medical records, including age, gender, ASA classification, baseline blood pressure level, cardiac function assessment results, and comorbidities; Patient risk warning threshold: A differentiated sensitivity adjustment algorithm is used to automatically adjust the warning threshold based on patient risk stratification; Tailored intervention plans: The intensity, timing, and specific measures of intervention recommendations are tailored based on individual patient characteristics.

6. The surgical hypotension early warning system based on multimodal physiological parameters according to claim 1, characterized in that: The decision support module includes: The risk grading unit uses a dynamic threshold division algorithm to map the predicted results, i.e. the predicted risk probability, to the clinical decision space, dividing the risk into multiple levels. It also uses a visual risk heat map and trend curve to present the risk evolution trajectory, in conjunction with an audio-visual warning mechanism; The intervention recommendation unit adopts a multi-level reasoning architecture based on the combination of rules and knowledge graphs to generate and display clinical intervention recommendations with clear priority and timeliness markings.

7. The surgical hypotension early warning system based on multimodal physiological parameters according to claim 1, characterized in that: The result display module includes: A real-time display unit displays the patient's current risk level and the risk probability value for each predicted time window in digital and color-coded form; Trend chart unit, used to display the time trend of key parameters and risk probability, and clearly mark the warning trigger points; Alarm control unit, used to trigger different levels of visual and audible alarms based on risk levels, and support medical staff in confirming and pausing operations.

8. A surgical hypotension warning method based on multimodal physiological parameters, characterized in that: The following steps are involved: Used to obtain real-time physiological parameter data of patients; Based on the preprocessed data, a multi-dimensional feature space including key indicators, composite indicators and dynamic indicators is constructed; Based on the multi-dimensional feature space, feature screening is performed, and based on the optimal feature subset obtained by screening, an optimal feature combination corresponding to different prediction time windows is constructed; the optimal feature combination is input into a pre-built and trained prediction analysis model, and prediction results of the patient's hypotension risk corresponding to different prediction time windows are output; Performing risk stratification on the predicted results, and generating clinical intervention recommendations by considering the risk stratification results and factors leading to an increased risk of hypotension; The prediction results and clinical intervention recommendations are displayed.

9. The method for early warning of surgical hypotension based on multimodal physiological parameters according to claim 8, characterized in that: Perform the following steps to construct a multi-dimensional feature space including key indicators, composite indicators, and dynamic indicators: Preprocessing of patients' real-time physiological parameter data, including data validity check, multi-source data completion, and time series normalization; Perform outlier detection and data smoothing on the pre-processed data, and normalize the key indicators, i.e. the key monitored physiological parameters; Construct a composite indicator that reflects the patient's real-time health status; Calculate the short-term change characteristics of multiple key indicators in multiple continuous time windows and construct multiple dynamic indicators that reflect the patient's real-time health status.

10. The method for early warning of surgical hypotension based on multimodal physiological parameters according to claim 9, characterized in that: The first-stage model and the second-stage model of the prediction analysis model, the first-stage model includes a parallel ensemble learning model based on a gradient boosting tree and a time series feature analysis model based on a Transformer, and the optimal feature combination and output of the ensemble learning model based on the gradient boosting tree and the time series feature analysis model based on the Transformer are weightedly fused to obtain a preliminary prediction result; The second-stage model uses a Transformer filter to input the preliminary prediction results output by the first-stage model, conducts refined screening of possible false positive results that may appear in the prediction of the first-stage model, and outputs the final prediction results of hypotension risk in multiple different time windows.

Images