Perioperative period hypotension risk prediction and early warning method and system based on artificial intelligence

Abstract

The invention discloses a perioperative period hypotension risk prediction and early warning method and system based on artificial intelligence, belongs to the crossing field of medical artificial intelligence and clinical anesthesia, and is suitable for painless gastrointestinal endoscope perioperative period management of elderly hypertensive patients. Preoperative-intraoperative full-process data is obtained through a multi-source data acquisition module, after feature engineering processing, a double-order AI architecture of a preoperative risk layering sub-model and an intraoperative dynamic prediction sub-model is input, a comprehensive risk value is output through dynamic weight fusion, and a three-level early warning mechanism and an evidence-based intervention closed loop are combined to obtain an early warning result. And hypotension pre-prediction and individualized intervention are realized. According to the method, the problems of monitoring lag, poor generalization and other pain points in the prior art are solved, accurate early warning can be carried out 5 minutes in advance, the haemodynamic management accuracy and anesthesia safety in the perioperative period are remarkably improved, and the occurrence rate of hypotension and related adverse events is reduced.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical artificial intelligence and clinical anesthesiology, specifically relating to an artificial intelligence-based method and system for predicting and warning of perioperative hypotension risk. Background Technology

[0002] With the widespread adoption of comfortable medical care, painless gastroscopy and colonoscopy have become core clinical methods for screening and minimally invasive treatment of digestive tract diseases. Elderly patients with hypertension are the primary beneficiaries of this examination, but they are also a high-risk group for perioperative adverse events. Elderly patients with hypertension have a significantly higher incidence of perioperative hypotension, bradycardia, and respiratory depression due to decreased vascular elasticity, reduced autonomic nervous system regulation, insufficient volume due to preoperative bowel preparation, and the circulatory inhibitory effects of anesthetic drugs. Persistent intraoperative hypotension can lead to hypoperfusion of vital target organs such as the heart, brain, and kidneys, significantly increasing the risk of perioperative acute cardiovascular and cerebrovascular events and acute kidney injury, seriously threatening patient life.

[0003] Currently, the routinely used intermittent noninvasive blood pressure (NIBP) monitoring in clinical practice typically involves measurements every 3-5 minutes. Clinical studies have confirmed that it results in a 20% missed rate of intraoperative hypotension events and a significant recognition lag. It can only provide post-event alerts for hypotension and cannot provide sufficient intervention window for clinical intervention, making it difficult to achieve proactive risk control of hypotension. While continuous noninvasive stroke pressure (CNAP) monitoring can achieve stroke-by-stroke blood pressure waveform acquisition and real-time numerical display, it only has basic monitoring functions and lacks risk prediction and early warning capabilities, failing to address the clinical pain point of delayed hypotension prevention and control at its root.

[0004] In recent years, AI-based intraoperative hypotension prediction models have gradually become a hot topic in clinical research. However, existing technologies suffer from four core flaws, making them unsuitable for practical application in clinical scenarios: First, the sample construction method has inherent defects. Existing models generally use a fixed sliding window method to generate training samples, resulting in serious temporal correlation and label ambiguity issues. This leads to significant sample selection bias and severely inadequate generalization performance. Second, the model input dimension is limited. Existing technologies focus only on the single-dimensional analysis of intraoperative blood pressure waveforms, failing to integrate multi-source clinical data such as preoperative baseline risk, intraoperative medication, and endoscopic events. This makes it impossible to adapt to the dynamic risk throughout the entire process, from pre-anesthesia induction to post-anesthesia induction and during the procedure. The problems include: 1) poor prediction accuracy and scenario adaptability; 2) insufficient model robustness, with existing models generally excluding samples in the hypotension borderline gray zone (MAP 65-75 mmHg) during training, leading to a sharp drop in performance across different centers and surgical cohorts, especially in elderly high-risk hypertension populations; 3) lack of a closed-loop management system, with existing technologies only achieving risk prediction and basic early warning functions, failing to integrate evidence-based medicine to build corresponding individualized intervention plans, and lacking intervention effect feedback and model iteration mechanisms, thus failing to realize the technology's implementation from risk prediction to clinical outcome improvement, making it difficult to truly reduce the incidence of intraoperative hypotension. Summary of the Invention

[0005] To address the aforementioned deficiencies in existing technologies, the core objective of this invention is to provide an artificial intelligence-based perioperative hypotension risk prediction and early warning system. This system systematically solves the core clinical pain points of existing technologies, such as lagging hypotension monitoring, poor model generalization, disconnect between preoperative and intraoperative data, and the lack of a closed-loop intervention system, through multi-source clinical data fusion, a two-stage AI risk prediction model, a tiered differentiated early warning mechanism, and evidence-based intervention closed-loop linkage. This invention aims to achieve a technological breakthrough in perioperative hypotension management, moving from "passive postoperative monitoring" to "active preoperative prediction and early warning," precisely adapting to high-risk perioperative scenarios such as painless gastroscopy and colonoscopy in elderly hypertensive patients, and significantly improving the accuracy of hemodynamic management and anesthesia safety in high-risk populations.

[0006] This invention provides a method for predicting and warning of perioperative hypotension risk based on artificial intelligence, comprising the following steps:

[0007] Step S1: Multi-source data acquisition, standardized synchronous acquisition of multi-dimensional data throughout the perioperative process, including preoperative static baseline clinical data, intraoperative stroke-level hemodynamic dynamic time-series data, and intraoperative anesthesia and operation-related event data;

[0008] Step S2: Feature engineering processing, which involves preprocessing the collected multi-source data, extracting temporal features, and filtering features in sequence to obtain two types of feature results: preoperative core static feature set and intraoperative fusion feature vector.

[0009] Step S3: AI risk prediction model construction, training and fusion. Construct a dual-model architecture of a preoperative hypotension risk stratification sub-model and an intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. Use the preoperative core static feature set as the input of the preoperative hypotension risk stratification sub-model and the intraoperative fused feature vector as the input of the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. After completing the model training and validation respectively, the output results of the two sub-models are weighted and fused through a dynamic weight fusion formula to obtain the final comprehensive risk probability.

[0010] Step S4: Graded early warning. Based on the final comprehensive risk probability and real-time average arterial pressure monitoring value, a three-level early warning mechanism is set up and matched with a graded anti-shake strategy to achieve differentiated early warning of perioperative hypotension.

[0011] Step S5: Clinical intervention linkage. Based on the graded early warning results and evidence-based medicine, individualized intervention recommendations are output. At the same time, hemodynamic data after intervention are collected to evaluate the intervention effect. An intervention effect feedback loop is constructed and the above two sub-models are iteratively optimized through a dual-track mechanism.

[0012] Preferably, step S1 specifically includes:

[0013] S11: Preoperative static data collection, which includes four categories of preoperative static baseline clinical data: demographic characteristics, disease-related characteristics, laboratory test results, and perioperative-related characteristics.

[0014] S12: Intraoperative dynamic time-series data acquisition: Intraoperative stroke-by-stroke dynamic hemodynamic time-series data of heart rate, mean arterial pressure, systolic blood pressure, diastolic blood pressure, cardiac output, peripheral vascular resistance, peripheral blood oxygen saturation, and respiratory rate are acquired through a stroke-by-stroke continuous non-invasive arterial pressure monitoring system at a fixed sampling frequency.

[0015] S13: Intraoperative event and medication data collection, real-time collection of the entire process of anesthesia and medication, data on key nodes of endoscopic operation, and records of intraoperative adverse events and their management.

[0016] Preferably, the data preprocessing in step S2 specifically includes:

[0017] Outlier and missing value handling: For static data, multiple interpolation is used to handle missing values, and outlier samples exceeding the threshold range are removed based on the quartile method; for time-series hemodynamic signals, Gaussian derivative filter and Shannon energy envelope method are used to detect peak arterial blood pressure signals.

[0018] Data standardization and resampling: The original arterial blood pressure signal was resampled to a uniform frequency using discrete Fourier transform, and all numerical features were normalized using the standard deviation standardization method.

[0019] Sample construction and label definition: Using the preset mean arterial pressure index as the gold standard for hypotension, candidate time points are dynamically sampled and sample label assignment is completed by combining a three-level window structure, and the sample set division ensures that the data of the same patient does not cross sets.

[0020] Preferably, the temporal feature extraction in step S2 specifically includes:

[0021] Manual statistical feature extraction: For a 20-second arterial blood pressure input segment, 32-dimensional time-domain, frequency-domain, and nonlinear manual statistical features were extracted. Simultaneously, 8-dimensional time-series statistical features of heart rate, cardiac output, peripheral vascular resistance, and peripheral blood oxygen saturation were extracted, resulting in a total of 40-dimensional manual time-series features.

[0022] Deep waveform feature extraction: The deep morphological features of the 20-second arterial blood pressure waveform are automatically extracted and the deep feature vector is output through the first 6 one-dimensional convolutional layers of the intraoperative stroke-level hypotension dynamic prediction sub-model.

[0023] Preferably, the feature selection in step S2 specifically includes:

[0024] Preoperative static feature screening: First, redundant features with correlation coefficients greater than 0.8 are removed using the Pearson correlation coefficient method. Then, a 28-dimensional core static feature set is selected by ranking the feature importance using a lightweight gradient booster, which serves as the input for the preoperative hypotension risk stratification sub-model.

[0025] Intraoperative temporal feature selection: First, redundant features were removed using the Pearson correlation coefficient method. Then, the top 20 core temporal features were selected by ranking the importance of features using a gradient boosting tree. These 20 core temporal features were then concatenated with deep waveform features, preoperative risk stratification results, and anesthetic drug and operation event coding features to form an intraoperative fusion feature vector, which was used as the input to the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model.

[0026] Preferably, the construction, training, and fusion of the AI ​​risk prediction model in step S3 specifically includes:

[0027] A preoperative hypotension risk stratification sub-model and an intraoperative stroke-level hypotension dynamic prediction sub-model were constructed and trained and fused. The preoperative hypotension risk stratification sub-model adopted a lightweight gradient booster binary classification model. The input of the preoperative core static feature set was used to achieve hypotension risk stratification. The training was completed through 5-fold cross-validation.

[0028] The intraoperative stroke-level hypotension dynamic prediction sub-model is based on a 7-layer one-dimensional convolutional neural network. It integrates arterial blood pressure time-series signals and multiple intraoperative auxiliary features for training, and outputs the probability of hypotension occurring in the next 5 minutes. The constructed AI risk prediction model integrates multi-source datasets. After pre-training on public datasets and fine-tuning on target domain datasets, it completes 5-fold cross-validation and generalization validation in accordance with the principles of patient-level data isolation and single variable control.

[0029] Preferably, the dynamic weight fusion formula in step S3 is as follows:

[0030] ;

[0031] In the formula, For the final overall risk probability, The preoperative risk stratification probability output by the preoperative hypotension risk stratification sub-model. This is the real-time intraoperative prediction probability output by the dynamic prediction sub-model for intraoperative stroke-level hypotension. These are dynamic weighting coefficients.

[0032] Preferably, the three-level early warning mechanism and graded anti-shake strategy in step S4 specifically include:

[0033] Low-risk blue alert: The triggering condition is a final comprehensive risk probability of 0~30% and a real-time average arterial pressure ≥75mmHg. The system continuously monitors and displays the risk value and trend curve. It does not trigger active alerts and has no anti-shake restrictions.

[0034] Medium-risk yellow alert: The triggering condition is that the final comprehensive risk probability is 30%~70% or the real-time average arterial pressure is in the gray range of 65~75 mmHg. The triggering rule is that the medium-risk standard is reached for 3 consecutive prediction cycles. The system will pop up a yellow alert and broadcast a voice reminder at the same time.

[0035] High-risk red alert: The triggering condition is that the final comprehensive risk probability is ≥70% or the real-time mean arterial pressure is <65mmHg. The triggering rule is single-cycle triggering plus double-cycle verification. The alert is triggered immediately when the first prediction cycle reaches the standard. If the risk drops to medium or low risk in the subsequent two cycles, the alert will be automatically lifted. After triggering, the system will highlight the pop-up window and provide continuous voice alarm.

[0036] Preferably, the clinical intervention linkage in step S5 specifically includes:

[0037] Construction of an evidence-based intervention knowledge base: A knowledge base was constructed based on the conclusions of multicenter, prospective, single-blind, randomized controlled clinical trials and guidelines for the management of perioperative hypotension.

[0038] Individualized intervention recommendations are provided: for low-risk patients, recommendations for maintaining and continuously monitoring anesthesia protocols are provided; for medium-risk patients, recommendations for tiered interventions are provided; and for high-risk patients, recommendations for emergency intervention protocols are provided.

[0039] Intervention effect feedback and model iteration: Real-time collection of hemodynamic data after intervention to evaluate the effect. If the risk of hypotension does not decrease within 5 minutes after intervention, the intervention plan is automatically upgraded. The model iteration adopts a dual-track mechanism of offline regular retraining and online incremental fine-tuning. Offline, desensitized clinical data is summarized every quarter to complete full retraining. Online, federated learning is used to complete incremental fine-tuning for single centers / special populations.

[0040] This invention also provides an artificial intelligence-based perioperative hypotension risk prediction and early warning system for performing the above-described method. The system includes:

[0041] The multi-source data acquisition module is used to standardize and synchronously acquire multi-dimensional data throughout the perioperative period. The multi-dimensional data includes preoperative static baseline clinical data, intraoperative stroke-level hemodynamic dynamic time-series data, and intraoperative anesthesia and operation-related event data.

[0042] The feature engineering processing module is used to perform data preprocessing, temporal feature extraction, and feature filtering on the collected multi-source data in sequence to obtain two types of feature results: preoperative core static feature set and intraoperative fusion feature vector.

[0043] The AI ​​risk prediction model construction, training and fusion module is used to build a dual-model architecture of a preoperative hypotension risk stratification sub-model and an intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. The preoperative core static feature set is used as the input of the preoperative hypotension risk stratification sub-model, and the intraoperative fused feature vector is used as the input of the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. After the model training and validation are completed respectively, the output results of the two sub-models are weighted and fused through a dynamic weight fusion formula to obtain the final comprehensive risk probability.

[0044] The graded early warning module is used to set up a three-level early warning mechanism and match a graded anti-shake strategy based on the final comprehensive risk probability and real-time average arterial pressure monitoring value, so as to realize differentiated early warning of perioperative hypotension.

[0045] The clinical intervention linkage module is used to output individualized intervention recommendations based on the results of graded early warning and evidence-based medicine. At the same time, it collects hemodynamic data after intervention to evaluate the intervention effect, constructs a closed loop for intervention effect feedback, and completes the iterative optimization of the above two sub-models through a dual-track mechanism.

[0046] The present invention provides a method and system for predicting and warning the risk of perioperative hypotension based on artificial intelligence, which has the following advantages compared with the prior art:

[0047] (1) This invention constructs a two-stage AI model architecture that combines preoperative risk stratification with intraoperative dynamic prediction. It can accurately predict the occurrence of intraoperative hypotension events 5 minutes in advance, which completely solves the problems of missed diagnosis and delayed identification of hypotension in traditional NIBP monitoring. It reserves sufficient golden treatment window for clinical intervention and reduces the occurrence of hypotension events from the root. At the same time, through the dynamic weight fusion mechanism of the entire anesthesia process, it accurately adapts to the changes of risk-dominant factors in different surgical stages. Compared with the single-model prediction scheme, it significantly improves the accuracy and scenario adaptability of perioperative risk prediction.

[0048] (2) This invention uses a random sliding window framework based on Poisson process to complete sample construction. Through a three-level window structure and strict hard screening rules, it eliminates the time correlation and label ambiguity problems caused by fixed sliding windows from the root, and avoids the sample selection bias of the model. At the same time, it incorporates low blood pressure gray zone samples with MAP 65-75 mmHg into the model training. Combined with the training strategy of "source domain big data pre-training + target domain small sample fine-tuning", the model can achieve an AUC-ROC of 82.37% in the elderly high-risk population with hypertension. It also has excellent cross-center, cross-cohort and cross-subgroup generalization performance, and completely solves the core pain point of the performance drop of existing AI models in different clinical scenarios.

[0049] (3) This invention constructs a closed-loop management system of “prediction-early warning-intervention-feedback”, and outputs tiered and individualized evidence-based intervention suggestions for the three-level early warning results. The intervention knowledge base is constructed based on high-level evidence-based evidence from the multi-center clinical trials of this invention, and outputs an optimized anesthesia regimen with remimazolam as the core for elderly hypertensive patients. Compared with the traditional propofol anesthesia regimen, it can reduce the incidence of intraoperative hypotension in elderly hypertensive patients from 72.7% to 37.3%, while significantly reducing the incidence of adverse events such as bradycardia and respiratory depression, and reducing the clinical use of vasoactive drugs. It truly realizes the implementation of technology from risk prediction to clinical outcome improvement, and greatly improves the perioperative anesthesia safety of elderly hypertensive patients.

[0050] (4) This invention targets the high-risk clinical scenario of painless gastroscopy and colonoscopy for elderly hypertensive patients. It integrates unique scenario-based features such as preoperative bowel preparation data, anesthetic drug data, and endoscopic operation events. Through customized design such as multimodal feature fusion, hierarchical differentiated anti-shake early warning, and federated learning incremental fine-tuning, it achieves synergistic effect of each module. It is not a simple superposition of existing technologies and has outstanding creativity. At the same time, each module of the system is developed based on clinically mature medical equipment and algorithm technology, which perfectly adapts to the clinical workflow of outpatient painless gastroscopy and colonoscopy and other outpatient anesthesia scenarios outside the operating room. It has strong clinical feasibility and promotion value, and provides standardized and intelligent technical support for the optimization of anesthesia programs for clinical painless diagnosis and treatment. Attached Figure Description

[0051] Figure 1 This is a flowchart of a method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to the present invention.

[0052] Figure 2 This is a flowchart of step S2 of the artificial intelligence-based method for predicting and warning the risk of perioperative hypotension in this invention.

[0053] Figure 3 This is a flowchart of step S3 of the artificial intelligence-based method for predicting and warning the risk of perioperative hypotension in this invention.

[0054] Figure 4 This is a flowchart of step S4 of the artificial intelligence-based method for predicting and warning the risk of perioperative hypotension according to the present invention.

[0055] Figure 5 This is a flowchart of step S5 of the artificial intelligence-based method for predicting and warning the risk of perioperative hypotension in this invention.

[0056] Figure 6 This is a schematic diagram of the structure of a perioperative hypotension risk prediction and early warning system based on artificial intelligence according to the present invention. Detailed Implementation

[0057] The following detailed description of a method and system for predicting and warning the risk of perioperative hypotension based on artificial intelligence, in conjunction with specific embodiments, illustrates the present invention. These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0058] Example 1: Implementation of an artificial intelligence-based method for predicting and warning the risk of perioperative hypotension.

[0059] Combined with appendix Figures 1-5 As shown, this invention provides a method for predicting and warning of perioperative hypotension risk based on artificial intelligence.

[0060] Step S1: Multi-source data acquisition.

[0061] This step is used to achieve standardized and synchronous collection of multi-dimensional data throughout the perioperative period. The data collection process follows medical data privacy protection standards and completes de-identification and encrypted storage.

[0062] Step S1 involves standardized synchronous collection of multi-dimensional data throughout the perioperative period, including preoperative static baseline clinical data, intraoperative stroke-level hemodynamic dynamic time-series data, and intraoperative anesthesia and procedural event data.

[0063] Step S1 specifically includes:

[0064] Preoperative static data collection includes four categories of preoperative static baseline clinical data: demographic characteristics, disease-related characteristics, laboratory test results, and perioperative-related characteristics.

[0065] Intraoperative dynamic time-series data acquisition: The intraoperative stroke-by-stroke dynamic time-series data of heart rate, mean arterial pressure, systolic blood pressure, diastolic blood pressure, cardiac output, peripheral vascular resistance, peripheral blood oxygen saturation, and respiratory rate were collected using a stroke-by-stroke non-invasive arterial pressure monitoring system at a fixed sampling frequency.

[0066] Intraoperative event and medication data acquisition includes real-time recording of the entire anesthesia and medication process, key data points of the endoscopic procedure, and records of intraoperative adverse events and their management. The specific process is as follows:

[0067] Step S11: Preoperative static data collection.

[0068] Collect all preoperative baseline clinical data of the patient, including:

[0069] (1) Demographic characteristics: age, sex, height, weight, body mass index (BMI);

[0070] (2) Disease-related characteristics: American Society of Anesthesiologists (ASA) classification, hypertension classification and course of disease, comorbidities (diabetes, cardiovascular disease, respiratory disease, etc.), and cardiac function classification;

[0071] (3) Laboratory test results: hemoglobin, blood glucose, albumin, creatinine, electrolytes (sodium, potassium), etc.;

[0072] (4) Perioperative characteristics: Preoperative bowel preparation data (type of bowel preparation drugs, frequency of diarrhea, duration of fasting and drinking), preoperative basic vital signs (resting heart rate (HR), mean arterial pressure (MAP), systolic blood pressure (SBP), diastolic blood pressure (DBP)).

[0073] Data sources include the Hospital Information System (HIS), Laboratory Information System (LIS), Preoperative Anesthesia Assessment System, and a complete clinical dataset of 220 elderly hypertensive patients collected from the multicenter clinical trial accompanying this invention.

[0074] Step S12: Intraoperative dynamic time-series data acquisition.

[0075] The continuous non-invasive arterial pressure (CNAP) monitoring system enables real-time synchronous acquisition of stroke-by-stroke hemodynamic time-series data during surgery. The sampling frequency is fixed at 125 Hz. The acquired parameters include HR, MAP, SBP, DBP, cardiac output (CO), and systemic vascular resistance (SVR). Peripheral oxygen saturation (SpO2) and respiratory rate are acquired simultaneously.

[0076] The data acquisition time points cover 2 minutes before anesthesia induction (T0, baseline value), during endoscope placement (T1), and every 5 minutes after the start of the operation (T2-Tn) until the end of the operation, realizing continuous and uninterrupted acquisition of time-series data of the entire anesthesia process.

[0077] Step S13: Intraoperative event and medication data collection.

[0078] Real-time acquisition of event-based data related to intraoperative anesthesia and procedures, including:

[0079] (1) Data on anesthetic drugs: type of induction and maintenance drugs (remazolam / propofol, sufentanil, etc.), dosage, administration time, administration method, and records of vasoactive drug use;

[0080] (2) Endoscopic procedure events: endoscopic insertion, colonoscopy bend passage, mucosal biopsy / treatment, and other operation nodes and total operation time;

[0081] (3) Intraoperative adverse events: timing and management of body movement, coughing, respiratory depression, hypotension / hypertension, bradycardia / tachycardia.

[0082] Step S2, feature engineering processing.

[0083] This step is used to complete the standardization processing of multi-source data, effective feature extraction and sample construction, to provide high-quality input for AI models, and to perform independent processing and screening processes for preoperative static features and intraoperative temporal features to ensure the consistency and adaptability of feature dimensions.

[0084] In step S2, the collected multi-source data are preprocessed, time-series feature extracted, and feature filtered sequentially to obtain two types of feature results: preoperative core static feature set and intraoperative fusion feature vector.

[0085] In step S2, data preprocessing specifically includes: handling outliers and missing values: for static data, multiple interpolation is used to handle missing values, and outlier samples exceeding the threshold range are removed based on the quartile method; for time-series hemodynamic signals, Gaussian derivative filter and Shannon energy envelope method are used to complete the peak detection of arterial blood pressure signals; data standardization and resampling: the original arterial blood pressure signal is resampled to a uniform frequency using discrete Fourier transform, and the standard deviation standardization method is used to normalize all numerical features; sample construction and label definition: the preset mean arterial pressure index is used as the gold standard for hypotension, candidate time points are dynamically sampled and sample labels are assigned using a three-level window structure, and the sample set division ensures that data from the same patient do not cross sets.

[0086] In step S2, the temporal feature extraction specifically includes: manual statistical feature extraction: for the 20-second arterial blood pressure input segment, 32-dimensional time-domain, frequency-domain, and nonlinear manual statistical features are extracted, and 8-dimensional temporal statistical features of heart rate, cardiac output, peripheral vascular resistance, and peripheral blood oxygen saturation are extracted simultaneously, generating a total of 40-dimensional manual temporal features; deep waveform feature extraction: through the first 6 one-dimensional convolutional layers of the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model, the deep morphological features of the 20-second arterial blood pressure waveform are automatically extracted and the deep feature vector is output.

[0087] In step S2, feature selection specifically includes: preoperative static feature selection: First, redundant features with correlation coefficients > 0.8 are removed using the Pearson correlation coefficient method. Then, a lightweight gradient boosting tree is used to rank the features by importance, selecting a 28-dimensional core preoperative static feature set as input to the preoperative hypotension risk stratification sub-model. Intraoperative temporal feature selection: First, redundant features are removed using the Pearson correlation coefficient method. Then, the top 20 core temporal features are selected by ranking the features by importance using a gradient boosting tree. These 20 core temporal features are concatenated with deep waveform features, preoperative risk stratification results, and anesthetic drug and procedural event coding features to form an intraoperative fusion feature vector, which is used as input to the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. The specific process is as follows:

[0088] Step S21, data preprocessing.

[0089] Step S211, handling outliers and missing values.

[0090] (1) Static data processing: missing values ​​are processed by multiple interpolation and outliers are removed based on the interquartile range, that is, outliers that exceed the range of [Q1-1.5IQR, Q3+1.5IQR] are removed, where Q1 is the first quartile, Q3 is the third quartile, and IQR is the interquartile range.

[0091] (2) Time-series signal processing: The peak value of arterial blood pressure (ABP) signal was detected using a Gaussian derivative filter and Shannon energy envelope method. Abnormal signals such as high-frequency noise, square wave artifacts, and motion interference were removed beat by beat. Abnormal pulsations were eliminated based on the Signal Abnormality Index (SAI). The SAI calculation method is as follows:

[0092] ;

[0093] in, The amplitude of the sampling point of the single-beat signal. This represents the signal mean value for that pulsation cycle. This represents the standard deviation of the signal during that beat cycle; when the single beat signal... The proportion of sampling points with a value greater than 1 exceeds If so, it is determined to be an abnormal pulsation and is excluded;

[0094] (3) Missing data completion: Linear interpolation is used to complete time series data with missing duration <10s, and segments with missing duration ≥10s are removed.

[0095] Step S212, data standardization and resampling.

[0096] (1) Resampling: For the original ABP signal acquired by CNAP, the Discrete Fourier Transform is used to resample to a uniform 125Hz frequency to ensure the consistency of the time data length;

[0097] (2) Standardization: For all numerical features, the Z-score standardization method (i.e. standard deviation standardization method) is used to complete the normalization process.

[0098] Step S213: Sample construction and label definition.

[0099] Referring to the clinical gold standard definition of a hypotensive event as MAP < 65 mmHg lasting at least 1 minute, a Poisson process-based random sliding window framework was used to construct the sample, resolving the temporal correlation and label ambiguity issues of fixed sliding windows. The specific rules are as follows:

[0100] (1) Candidate time point sampling: candidate time points Dynamically sample from the exponential distribution, distribution rate =1 / 3min (average sampling interval 3 minutes), set the maximum number of sampling points per patient to no more than 20 to avoid oversampling;

[0101] (2) Three-level window structure: each sampling point Anchored three-level window structure: ① 1 minute 20 second observation window: starting from The first minute is the baseline period without hypotension, and the next 20 seconds of ABP time sequence is the model input; ② 5-minute prediction window: immediately following the observation window, if a hypotension event begins in this window, the sample is marked as a positive sample; ③ 1-minute relaxation window: immediately following the prediction window, if a hypotension event begins in this window, the sample is removed to avoid edge labeling error.

[0102] (3) Hard screening rules: All samples with any hypotension event within the observation window are removed; only when there is no hypotension event within the prediction window and relaxation window is the sample marked as a negative sample; when dividing the sample set, ensure that the samples of the same patient are not split into the training set and the test set to avoid data leakage.

[0103] Based on the above rules, 400-500 valid samples can be generated from a single surgical patient. The 2,603 ​​patients used in the training of this invention generated more than 1.2 million valid samples in total, which meets the data scale requirements for model training.

[0104] Step S22, temporal feature extraction.

[0105] By integrating manual statistical features and deep waveform features, a multi-dimensional feature set is constructed, and the extraction and fusion logic of multi-modal features is clarified, specifically including:

[0106] (1) Manual statistical feature extraction: For a 20-second ABP (arterial blood pressure) input segment, a total of 32-dimensional time-domain features (MAP mean / maximum / minimum / standard deviation / slope change rate, HR mean and variability, CO and SVR mean and trend, blood pressure waveform pulse pressure / slope of ascending branch, etc.), frequency-domain features (low-frequency / high-frequency power ratio obtained by fast Fourier transform), and nonlinear features (sample entropy, approximate entropy) are extracted; 8-dimensional time-series statistical features of HR, CO, SVR and SpO2 in the same period are extracted simultaneously, for a total of 40-dimensional manual time-series features;

[0107] (2) Deep feature extraction: The first 6 one-dimensional convolutional layers of the intraoperative pulse-level hypotension dynamic prediction sub-model automatically extract the deep morphological features of the 20-second ABP waveform, capture the subtle waveform changes that cannot be identified manually before the occurrence of hypotension, and output the deep feature vector to supplement the information blind spot of manual features; the convolutional layer shares weights with the intraoperative prediction model, and there is no need to train the feature extraction network separately, which reduces the system complexity.

[0108] Step S23, Feature Filtering.

[0109] Independent screening processes are performed for preoperative static features and intraoperative temporal features to ensure consistency of model input dimensions and feature effectiveness. Specifically, this includes:

[0110] (1) Screening of preoperative static features: For the 35-dimensional preoperative static features initially collected, redundant features were first removed by the Pearson correlation coefficient method (i.e., Pearson correlation coefficient method) (features with more significant clinical significance were retained in feature pairs with correlation coefficient > 0.8), and then the 28-dimensional core preoperative static feature set was selected by ranking the importance of features through the Lightweight Gradient Boosting Machine (i.e. LightGBM) as the input of the preoperative hypotension risk stratification sub-model;

[0111] (2) Screening of intraoperative time series features: For the 40-dimensional manual time series features, redundant features were removed by using the Pearson correlation coefficient method. The top 20 core time series features were selected by ranking the importance of features through gradient boosting tree. These 20 core manual features were then concatenated with the deep features, the preoperative risk stratification results (1-dimensional), and the anesthetic drug and operation event coding features (4-dimensional) to form the final intraoperative fusion feature vector, which was used as the input of the fully connected layer of the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model.

[0112] Step S3: Construction and training of AI risk prediction model.

[0113] This step is the core of the method, which constructs a dual-model architecture of "preoperative hypotension risk stratification sub-model + intraoperative stroke-level hypotension dynamic prediction sub-model". The two sub-models are operated in parallel, and the results are weighted and fused to output the final comprehensive risk probability.

[0114] Step S3 constructs a dual-model architecture: a preoperative hypotension risk stratification sub-model and an intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. The preoperative core static feature set is used as the input to the preoperative hypotension risk stratification sub-model, and the intraoperative fused feature vector is used as the input to the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. After completing the training and validation of the models respectively, the output results of the two sub-models are weighted and fused using a dynamic weight fusion formula to obtain the final comprehensive risk probability.

[0115] In step S3, the construction, training, and fusion of the AI ​​risk prediction model specifically includes:

[0116] A preoperative hypotension risk stratification sub-model and an intraoperative stroke-level hypotension dynamic prediction sub-model were constructed and trained and fused. The preoperative hypotension risk stratification sub-model adopted a lightweight gradient booster binary classification model. The input of the preoperative core static feature set was used to achieve hypotension risk stratification. The training was completed through 5-fold cross-validation.

[0117] The intraoperative stroke-by-stroke hypotension dynamic prediction sub-model is based on a 7-layer one-dimensional convolutional neural network, which integrates arterial blood pressure time-series signals and multiple intraoperative auxiliary features for training, outputting the probability of hypotension occurring in the next 5 minutes. The completed AI risk prediction model integrates multi-source datasets, is pre-trained on a public dataset, and fine-tuned on a target domain dataset. Following the principles of patient-level data isolation and single-variable control, 5-fold cross-validation and generalization validation are performed. The specific process is as follows:

[0118] Step S31: Construction and training of a preoperative hypotension risk stratification sub-model.

[0119] Step S311, Model structure design.

[0120] The LightGBM binary classification model was adopted, with the following hyperparameters: maximum tree depth of 6, learning rate of 0.05, number of iterations of 200, number of leaf nodes of 31, L2 regularization coefficient of 1e-3, and outputting the probability value of preoperative hypotension (range 0-1).

[0121] Step S312, define the model input and output.

[0122] (1) Input: The 28-dimensional preoperative core static feature set after filtering in step S23;

[0123] (2) Output: Preoperative hypotension risk stratification results, divided into three levels: low risk (risk probability < 0.3), medium risk (0.3 ≤ risk probability < 0.7), and high risk (risk probability ≥ 0.7).

[0124] Step S313, Model training and validation.

[0125] (1) Dataset construction: The training data comes from the clinical dataset of 220 elderly hypertensive patients in the multicenter of this invention, and is supplemented by the preoperative baseline data of 1964 patients in the VitalDB (i.e., vital signs database) public dataset (https: / / vitaldb.net);

[0126] (2) Data partitioning: Five-fold cross-validation was used, and the data was divided into training set, validation set and test set in a 7:1:2 ratio in the patient dimension to ensure that the data of the same patient are not distributed across sets;

[0127] (3) Training process: The model training was completed based on the LightGBM framework, and the AUC-ROC (area under the receiver operating characteristic curve) of the cross-validation set was used as the model performance evaluation index.

[0128] Step S32: Construction and training of the intraoperative stroke-level hypotension dynamic prediction sub-model.

[0129] Step S321, Model structure design.

[0130] An optimized 7-layer one-dimensional convolutional neural network (CNN) architecture is adopted to adapt to feature extraction and multimodal feature fusion of one-dimensional ABP time-series signals. Same padding (padding=4) ensures that the output length of the convolutional layer is consistent with the input. Feature dimensionality reduction is only achieved through pooling layers. The specific structure and size changes are as follows:

[0131] (1) Input layer: The input dimension is [2500, 1], corresponding to the ABP timing signal sampled for 20 seconds and 125Hz (20×125=2500 sampling points);

[0132] (2) Layer 1: One-dimensional convolutional layer (kernel size 10, stride 1, padding=4, number of kernels 32) → output dimension [2500, 32]; batch normalization layer + double ReLU activation layer + dropout layer (dropout rate 0.01); max pooling layer (pooling window 2, stride 2) → output dimension [1250, 32]; where ReLU is the corrected linear unit, one of the most commonly used activation functions in deep learning. Its core function is to introduce non-linear features into the neural network (so that the model can fit the complex clinical data patterns), while solving the gradient vanishing problem of traditional activation functions (such as Sigmoid);

[0133] (3) Layer 2: One-dimensional convolutional layer (kernel size 10, stride 1, padding=4, number of kernels 64) → output dimension [1250, 64]; batch normalization layer + double ReLU activation layer + dropout layer (dropout rate 0.01); max pooling layer (pooling window 2, stride 2) → output dimension [625, 64];

[0134] (4) Layer 3: One-dimensional convolutional layer (kernel size 10, stride 1, padding=4, number of kernels 64) → output dimension [625, 64]; batch normalization layer + double ReLU activation layer + dropout layer (dropout rate 0.01); max pooling layer (pooling window 2, stride 2) → output dimension [312, 64] (rounded down);

[0135] (5) Layer 4: One-dimensional convolutional layer (kernel size 10, stride 1, padding=4, number of kernels 128) → output dimension [312, 128]; batch normalization layer + double ReLU activation layer + dropout layer (dropout rate 0.01); max pooling layer (pooling window 2, stride 2) → output dimension [156, 128];

[0136] (6) Layer 5: One-dimensional convolutional layer (kernel size 10, stride 1, padding=4, number of kernels 128) → output dimension [156, 128]; batch normalization layer + double ReLU activation layer + dropout layer (dropout rate 0.01); max pooling layer (pooling window 2, stride 2) → output dimension [78, 128];

[0137] (7) 6th layer: one-dimensional convolutional layer (kernel size 10, stride 1, padding=4, number of kernels 64) → output dimension [78, 64]; batch normalization layer + double ReLU activation layer + dropout layer (dropout rate 0.01); max pooling layer (pooling window 2, stride 2) → output dimension [39, 64] (rounded down);

[0138] (8) Layer 7: One-dimensional convolutional layer (kernel size 10, stride 1, padding=4, number of kernels 32) → output dimension [39, 32]; batch normalization layer + double ReLU activation layer + dropout layer (dropout rate 0.01); no pooling layer;

[0139] (9) Feature fusion layer: The [39, 32] depth features output from the 7th layer are flattened into a 1248-dimensional vector, and then concatenated with the 20-dimensional core manual temporal features, the 1-dimensional preoperative risk stratification results, and the 4-dimensional anesthesia / operation event coding features to finally form a 1273-dimensional fusion feature vector;

[0140] (10) Fully connected layer: The fused features are connected to two fully connected layers with 128 and 64 neurons respectively, and the activation function is ReLU;

[0141] (11) Output layer: The Sigmoid activation function is used to output the probability value (range 0-1) of a hypotension event occurring within the next 5 minutes.

[0142] Step S322, loss function and optimizer configuration.

[0143] (1) Loss function: The weighted binary cross-entropy loss function Loss is used to solve the problem of imbalance between positive and negative samples during surgery. The formula is as follows:

[0144] ;

[0145] in, For the true labels of the samples, To predict probabilities for the model, Set the weights for positive samples (to 20). Negative sample weights (set to 1) are used to match positive and negative samples during surgery. The actual distribution ratio;

[0146] (2) Optimizer: The Adam optimizer (i.e., adaptive moment estimator) is used, with an initial learning rate of 0.001 and a weight decay coefficient. .

[0147] Step S323, Model Input and Output.

[0148] (1) Core input: ABP timing signal sampled at 125Hz for 20 seconds;

[0149] (2) Auxiliary inputs: time-series characteristics of HR, CO, and SVR during the same period, preoperative risk stratification results, and coding characteristics of anesthetic drugs and procedural events;

[0150] (3) Output: The predicted probability value of a hypotension event occurring within the next 5 minutes. The output frequency is synchronized with CNAP data acquisition and the prediction result is updated every 20 seconds to achieve dynamic risk monitoring at the beat-by-beat level.

[0151] Step S324, Model training and validation.

[0152] Step S3241, Construction of training dataset.

[0153] The training data consists of three parts: ① the complete intraoperative CNAP dataset of 220 elderly hypertensive patients from the multicenter clinical trial of this invention (China); ② intraoperative ABP data of 419 patients from the Karolinska public dataset (i.e., Karolinska University Hospital, Sweden); ③ intraoperative ABP data of 438 patients from the VDB Matched cohort and 1526 patients from the VitalDB public dataset (i.e., Vital Signs Database public dataset, South Korea); the total dataset contains 2603 patients, and more than 1.2 million valid samples are generated according to the sample construction rules in step S213, which are divided into training set (1822 cases), validation set (260 cases) and test set (521 cases) according to the patient dimension in a 7:1:2 ratio.

[0154] Step S3242, data preprocessing.

[0155] Following the process of steps S21-S23, complete signal denoising, standardization, sample construction and label assignment for all samples. Use time-shifting and amplitude scaling to augment the positive samples in the training set to alleviate the sample imbalance problem.

[0156] Step S3243, model pre-training.

[0157] Pre-training was performed using the Karolinska and VitalDB datasets, with a batch size of 20 and a maximum training epoch of 150. An early stopping mechanism was adopted, where training was stopped and the pre-trained weights were saved when the validation set AUC-ROC showed no improvement for 10 consecutive epochs.

[0158] Step S3244: Model fine-tuning.

[0159] Using the target domain dataset of 220 elderly hypertensive patients from multiple centers in this invention, the pre-trained model was fine-tuned. The weights of the first four convolutional layers were frozen, and only the last three convolutional layers, feature fusion layer, and fully connected layer were trained. The learning rate was adjusted to 1e-4, and the fine-tuning rounds were 50, so that the model could be adapted to the hemodynamic characteristics of painless gastrointestinal endoscopy in elderly hypertensive patients.

[0160] Step S3245, Model Validation.

[0161] This validation process strictly follows the TRIPOD-AI guidelines (i.e., the "Guideline for Transparent Reporting of Individual Prognoses or Diagnoses by Predictive Models (Artificial Intelligence Extended Version)"), with the following core principles: ① Patient-level data isolation: When dividing samples, ensure that all data from the same patient are not split into training and test sets, thus avoiding data leakage caused by intra-patient sample correlation at the source; ② Single variable control principle: All comparative experiments adjust only one target validation variable, and the remaining training data, pre-training weights, hyperparameters, and preprocessing procedures are completely consistent, ensuring that performance differences can be uniquely attributed; ③ Unified definition of the gray area: In all training and validation stages, the "gray area" for hypotension is uniformly defined as MAP 65~75mmHg, with no definitional bias.

[0162] Subgroup definition basis:

[0163] 1. Age subgroups: Referring to the definition of ≥65 years old as the elderly population in the "Guidelines for the Management of Hypertension in the Elderly in my country (2023 Edition)", and combined with the age distribution characteristics of the population enrolled in this study (median age 67 years), 67 years and older were defined as the elderly cohort, and 18-55 years old were defined as the non-elderly young cohort. The transitional age group of 56-66 years old was excluded to avoid confounding factors of age.

[0164] 2. ASA classification subgroup: Based on the American Society of Anesthesiologists (ASA) physical condition classification criteria, only patients with ASA grade II (mild systemic disease, no functional limitation) were included. This subgroup accounted for 76.4% (168 / 220) of the total enrolled population in this study. The sample size of each cross-validation test set was 33-34 cases, which met the sample size requirements for statistical testing.

[0165] The validation was divided into two parts: 50% cross-validation of core performance and validation of multi-dimensional generalization mechanisms. The former validated the model's overall performance in the target scenario, while the latter clarified the independent gains of "including gray zone samples" and "pre-training + fine-tuning strategy" by separating variables. The specific results are as follows:

[0166] Step S32451: Cross-validation of core performance at 50% discount.

[0167] Using the perioperative dataset of 220 elderly hypertensive patients undergoing painless gastrointestinal endoscopy in the multicenter clinical trial of this invention as the validation object, a patient-level stratified 5-fold cross-validation was adopted: the 220 patients were stratified according to the 1:1 ratio of hypotension event occurrence and randomly divided into 5 mutually exclusive subsets. In each fold, 4 subsets were selected as the training set and 1 subset was selected as the independent test set, ensuring that the data of the same patient only appeared in a single subset and there was no cross-set leakage.

[0168] Threshold selection process: In each fold cross-validation, only the training and validation sets of that fold are used to optimize the threshold; the independent test set is not involved in the threshold determination at all. The specific process is as follows:

[0169] 1. During each training iteration, with the goal of achieving 80% specificity on the validation set, the probability thresholds from 0.01 to 0.99 are iterated over on the validation set to lock in the optimal threshold that satisfies specificity ≥ 80%.

[0170] 2. Apply this threshold directly to the independent test set of this fold and calculate all performance metrics;

[0171] The mean of the 3.5-fold results was taken, and the 95% confidence interval was calculated based on the normal distribution assumption of the 5-fold test results. At the same time, the stability of the interval was verified by the bootstrap method (1000 resamplings), and the results showed no significant difference.

[0172] Table 1 shows the 5-fold cross-validation performance of the model of this invention on the target dataset. The results are presented as "mean ± 95% confidence interval", in %.

[0173]

[0174] The incidence of hypotension events at the test set level (positive sample proportion) was 28.7% (based on the statistics of segments generated by the sliding window; due to the different number of samples generated per patient, this differs from the patient-level prevalence of 37.3%, which is in line with expectations). Based on the sample-level prevalence of 28.7%, substituting the sensitivity (76.52%) and specificity (80.18%), the theoretical PPV calculated using the standard formula was 62.9% and the theoretical NPV was 89.8%, which is highly consistent with the actual values ​​in Table 1 (PPV 62.74%, NPV 90.26%).

[0175] Step S32452, multi-dimensional generalization verification.

[0176] To clarify the independent gains of different technical modules, this section uses a separation of variables design to split the model into two independent validation objectives. All validation scenarios adopt a strict generalization rule of pure source domain training and zero data testing in the target domain: the model is trained only on the source domain dataset, without ever touching any patient data in the target domain, and its performance is evaluated directly on the independent test set in the target domain.

[0177] All AUC-ROC 95% confidence intervals were calculated using the bootstrap method (1000 resampling); differences between groups were calculated using the DeLong test to calculate the original p-value, and multiple comparisons were corrected using Bonferroni. The corrected significance level was α=0.0083 (0.05 / 6 independent validation scenarios), and only differences with p<0.0083 were considered statistically significant.

[0178] (1) Validation objective 1: Generalization gain of samples included in the gray area.

[0179] 1. Variable control: Only "whether the training set contains gray area samples" is adjusted, while all other variables (pre-training weights, fine-tuning process, hyperparameters, network structure) remain completely unchanged;

[0180] 2. Model A (Complete Strategy of This Invention): The training set includes all samples in the MAP65~75mmHg gray range, and adopts the complete training process of "source domain pre-training + source domain fine-tuning";

[0181] 3. Baseline Model A1: The training set strictly excludes all samples in the MAP 65~75mmHg gray range, and the rest of the training process (pre-training weights, fine-tuning steps, hyperparameters) is completely consistent with Model A.

[0182] Table 2. Validation results of generalization gain for samples included in the gray area (AUC-ROC, %)

[0183]

[0184] Note: * indicates statistically significant differences after Bonferroni correction (α = 0.0083, p < 0.0083); unlabeled results did not reach the significance level after correction. The performance improvement in the young → old transfer scenario was significant, confirming that including samples from the gray zone effectively improves the model's generalization ability to the elderly high-risk population for hypertension.

[0185] (2) Validation of Objective 2: Performance gain of pre-training + fine-tuning strategy.

[0186] 1. Variable control: Only "whether to use pre-training + fine-tuning strategy" is adjusted, while all other variables (training set includes gray zone samples, hyperparameters, network structure) remain completely unchanged;

[0187] 2. Model A (Complete Strategy of This Invention): Same as Model A for Validation Objective 1;

[0188] 3. Baseline Model A2: It does not use source domain big data for pre-training, but is trained from scratch on the source domain dataset (with weights randomly initialized). The training set contains gray zone samples, and the other hyperparameters are completely consistent with Model A.

[0189] Table 3 Performance gain validation results of the pre-training + fine-tuning strategy (AUC-ROC, %)

[0190]

[0191] Note: * indicates statistically significant differences after Bonferroni correction (α = 0.025, 0.05 / 2 independent validation scenarios); for brevity, only core representative scenarios are shown. The results confirm that the "pre-training + fine-tuning" strategy can significantly improve the training efficiency and generalization performance of the model on small sample source domains.

[0192] Step S33: Fusion of results from multiple models.

[0193] The outputs of the preoperative hypotension risk stratification sub-model and the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model are weighted and fused to obtain the final comprehensive risk prediction value. The dynamic weight fusion formula is as follows:

[0194] ;

[0195] In the formula, For the final overall risk probability, The preoperative risk stratification probability output by the preoperative hypotension risk stratification sub-model. This is the real-time intraoperative prediction probability output by the dynamic prediction sub-model for intraoperative stroke-level hypotension. The dynamic weighting coefficients are adjusted according to the following rules:

[0196] (1) Before anesthesia induction (T0): The value was fixed at 0.8, with preoperative baseline risk as the core assessment criterion;

[0197] (2) From the start of anesthesia induction (T1) to the end of the endoscopic procedure: It decreases linearly with the duration of the procedure, dropping to exactly 0.2 at the end of the procedure, thus reflecting the dominant influence of real-time changes in intraoperative hemodynamics on risk.

[0198] (3) Event-triggered dynamic adjustment: When key events such as endoscopic placement, additional anesthetic drugs, or use of vasoactive drugs occur, The value was temporarily lowered by 0.1 and maintained for 3 prediction cycles to strengthen the weight of real-time intraoperative data; when the MAP entered the 65-75 mmHg gray zone, The threshold was temporarily increased by 0.1 to improve the specificity of the early warning system when combined with preoperative baseline risk.

[0199] Step S4: Tiered early warning.

[0200] This step sets up a three-level early warning mechanism and a graded anti-shake strategy based on the final comprehensive risk probability and real-time MAP monitoring value, balancing the proactiveness and accuracy of early warnings and avoiding early warning delays in emergency scenarios.

[0201] Step S4 sets up a three-level early warning mechanism and matches it with a graded anti-shake strategy based on the final comprehensive risk probability and real-time average arterial pressure monitoring value to achieve differentiated early warning of perioperative hypotension.

[0202] In step S4, the three-level warning mechanism and graded image stabilization strategy specifically include:

[0203] Low-risk blue alert: The triggering condition is a final comprehensive risk probability of 0~30% and a real-time average arterial pressure ≥75mmHg. The system continuously monitors and displays the risk value and trend curve. It does not trigger active alerts and has no anti-shake restrictions.

[0204] Medium-risk yellow alert: The triggering condition is that the final comprehensive risk probability is 30%~70% or the real-time average arterial pressure is in the gray range of 65~75 mmHg. The triggering rule is that the medium-risk standard is reached for 3 consecutive prediction cycles. The system will pop up a yellow alert and broadcast a voice reminder at the same time.

[0205] High-risk red alert: The trigger condition is a final comprehensive risk probability ≥70% or a real-time mean arterial pressure <65mmHg. The triggering rule is a single-cycle trigger plus a two-cycle review. The alert is triggered immediately when the standard is met in the first prediction cycle. If the risk falls back to low or medium risk in the subsequent two cycles, the alert is automatically lifted. After triggering, the system will display a highlighted pop-up window and provide a continuous voice alarm. The specific process is as follows:

[0206] Step S41, Low-risk warning (blue warning).

[0207] (1) Triggering conditions: The overall risk probability is 0~30%, and the real-time MAP is ≥75mmHg;

[0208] (2) Warning method: The system continuously monitors and updates the risk, and displays the current risk value and trend curve on the anesthesia monitoring interface. It does not trigger active warnings and has no anti-shake restrictions.

[0209] Step S42, medium risk warning (yellow warning).

[0210] (1) Triggering conditions: The overall risk probability is 30%~70%, or the real-time MAP is in the "grey range" of 65~75 mmHg;

[0211] (2) Triggering rule: The medium risk standard is reached for three consecutive prediction periods (60 seconds in total);

[0212] (3) Warning method: The system will pop up a yellow warning on the monitoring interface and broadcast a voice reminder to remind the anesthesiologist to pay attention to the trend of hemodynamic changes and prepare for intervention.

[0213] Step S43, High-risk warning (red warning).

[0214] (1) Triggering conditions: The overall risk probability is ≥70%, or the real-time MAP is <65mmHg;

[0215] (2) Triggering rules: The 60-second anti-shake rule is not applicable. The "single-cycle triggering + double-cycle review" mechanism is adopted. When the first prediction cycle reaches the high-risk standard, an early warning is triggered immediately, and a double-cycle review is started at the same time. If the risk falls back to medium-low risk in the subsequent two cycles, the early warning is automatically lifted.

[0216] (3) Warning method: The system immediately executes a red strong warning, highlights the monitoring interface pop-up window, continuously alarms the voice, locks the hemodynamic waveform and data of the current risk period, and pushes it to the clinical intervention linkage steps simultaneously.

[0217] Step S5: Clinical intervention linkage.

[0218] This step, based on the results of tiered early warning, the patient's real-time status, and evidence-based medicine, outputs individualized intervention recommendations and constructs a closed-loop feedback mechanism for intervention effects, achieving closed-loop management of the entire process of "prediction-early warning-intervention".

[0219] Step S5 outputs individualized intervention recommendations based on the graded early warning results and evidence-based medicine. At the same time, it collects hemodynamic data after intervention to evaluate the intervention effect, constructs a closed loop for intervention effect feedback, and completes the iterative optimization of the above two sub-models through a dual-track mechanism.

[0220] Step S5 specifically includes:

[0221] Construction of an evidence-based intervention knowledge base: A knowledge base was constructed based on the conclusions of multicenter, prospective, single-blind, randomized controlled clinical trials and guidelines for the management of perioperative hypotension.

[0222] Individualized intervention recommendations are provided: for low-risk patients, recommendations for maintaining and continuously monitoring anesthesia protocols are provided; for medium-risk patients, recommendations for tiered interventions are provided; and for high-risk patients, recommendations for emergency intervention protocols are provided.

[0223] Intervention effect feedback and model iteration: Real-time collection of hemodynamic data after intervention to assess the effect; if the risk of hypotension does not decrease within 5 minutes after intervention, the intervention plan is automatically upgraded; model iteration adopts a dual-track mechanism of offline periodic retraining and online incremental fine-tuning. Offline, desensitized clinical data is summarized quarterly to complete full retraining; online, federated learning is used to complete incremental fine-tuning for single centers / special populations. The specific process is as follows:

[0224] Step S51: Intervene in the construction of the knowledge base.

[0225] Based on the findings of the multicenter clinical trial accompanying this invention, perioperative hypotension management guidelines, and top-tier clinical research evidence, a standardized intervention knowledge base was constructed. The core evidence-based basis comes from the multicenter, prospective, single-blind, randomized controlled clinical trial of this invention: 220 elderly hypertensive patients scheduled for painless gastroscopy and colonoscopy from three hospitals were included and randomly assigned 1:1 to either the remimazolam group or the propofol group. The results demonstrated that remimazolam, compared to propofol, reduced the incidence of intraoperative hypotension in elderly hypertensive patients from 72.7% to 37.3%, while significantly reducing the incidence of bradycardia and respiratory depression, and exhibiting superior hemodynamic stability. The knowledge base includes anesthesia protocol adjustment strategies for different risk levels, guidelines for the use of vasoactive drugs, volume management protocols, respiratory and circulatory support measures, as well as the dosage adjustment principles for remimazolam / propofol and the safe drug thresholds for elderly hypertensive patients.

[0226] Step S52: Output individualized intervention recommendations.

[0227] Step S521, Low-risk intervention recommendations.

[0228] Maintain the current anesthesia protocol, and maintain a sedation depth of 1-2 points on the modified observer's assessment of alertness / sedation scale (MOAA / S), while continuously monitoring hemodynamic changes.

[0229] Step S522, Intervention recommendations for medium-risk cases.

[0230] Output tiered intervention recommendations, including:

[0231] (1) Adjustment of anesthesia regimen: If propofol is currently used to maintain anesthesia, it is recommended to switch to remimazolam infusion, or reduce the propofol infusion rate by 10% to 20% and add remimazolam 0.1 mg / kg for rescue sedation;

[0232] (2) Capacity management: Replenish capacity by uniformly infusing crystalloid solution;

[0233] (3) Respiratory management: Confirm airway patency and maintain oxygen flow rate of 6L / min;

[0234] (4) Drug preparation: Prepare vasoactive drugs such as ephedrine and metaraminol in advance.

[0235] Step S523, High-risk intervention recommendations.

[0236] Output emergency intervention plan, including:

[0237] (1) Operation adjustment: Immediately stop the endoscopic operation to reduce operational stimulation;

[0238] (2) Use of vasoactive drugs: If HR < 60 bpm, administer 6 mg of ephedrine intravenously; if HR is normal, administer 0.3 mg of metaraminol intravenously.

[0239] (3) Adjustment of anesthesia plan: Immediately stop the infusion of sedative drugs, and after the blood pressure rises, switch to low-dose infusion of remimazolam to maintain sedation;

[0240] (4) Respiratory and circulatory support: Lift the chin to keep the airway open, increase the oxygen flow rate to 10L / min, and provide mask-assisted ventilation if necessary;

[0241] (5) Enhanced monitoring: Continuously monitor vital signs until hemodynamics stabilizes.

[0242] Step S53: Closed-loop feedback of intervention effect and model iteration.

[0243] Step S531, evaluation of intervention effect.

[0244] Real-time data collection of hemodynamic changes after intervention measures are implemented to assess the intervention effect; if the risk of hypotension does not decrease or the mean arterial pressure (MAP) does not rise back to the target range within 5 minutes of a certain prognosis, the intervention plan is automatically upgraded and more effective intervention measures are pushed.

[0245] Step S532, model iteration mechanism.

[0246] A dual-track mechanism is adopted, consisting of offline periodic retraining and online incremental fine-tuning.

[0247] (1) Offline retraining: The clinical process data after desensitization is summarized every quarter. After ethical approval, the model is fully retrained, the model weights are updated, and the generalization is optimized.

[0248] (2) Online incremental fine-tuning: For datasets of single centers and special populations, federated learning is used to fine-tune the model without leaking the original patient data, while adapting to the clinical operating habits of different centers;

[0249] (3) Version control: All model iterations retain historical versions to achieve traceable version management.

[0250] Example 2: Implementation of an AI-based perioperative hypotension risk prediction and early warning system.

[0251] Combined with appendix Figure 6 As shown, the present invention provides an artificial intelligence-based perioperative hypotension risk prediction and early warning system.

[0252] An artificial intelligence-based perioperative hypotension risk prediction and early warning system is used to execute the method of the above embodiments. The system includes a multi-source data acquisition module, a feature engineering processing module, an AI risk prediction model construction, training and fusion module, a graded early warning module, and a clinical intervention linkage module that are connected in sequence.

[0253] The multi-source data acquisition module includes a preoperative static data acquisition unit, an intraoperative dynamic time-series data acquisition unit, and an intraoperative event and medication data acquisition unit, which respectively realize the standardized acquisition of preoperative static baseline clinical data, intraoperative stroke-level hemodynamic dynamic time-series data, and intraoperative anesthesia and procedure-related event data; it is used to standardize and synchronously acquire multi-dimensional data of the entire perioperative process, including preoperative static baseline clinical data, intraoperative stroke-level hemodynamic dynamic time-series data, and intraoperative anesthesia and procedure-related event data.

[0254] The feature engineering processing module includes a data preprocessing unit, a temporal feature extraction unit, and a feature filtering unit, which respectively realize outlier and missing value handling, sample construction, feature extraction, and independent filtering of multi-source data; it is used to perform data preprocessing, temporal feature extraction, and feature filtering on the collected multi-source data in sequence to obtain two types of feature results: preoperative core static feature set and intraoperative fusion feature vector.

[0255] The AI ​​risk prediction model construction, training, and fusion module is the core module of this system. It includes a preoperative hypotension risk stratification sub-model, an intraoperative stroke-by-stroke hypotension dynamic prediction sub-model, and a multi-model fusion unit, which respectively realizes preoperative risk stratification, intraoperative real-time prediction, and dynamic weighted fusion of the results of the two models. It is used to construct a dual-model architecture for the preoperative hypotension risk stratification sub-model and the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. The preoperative core static feature set is used as the input of the preoperative hypotension risk stratification sub-model, and the intraoperative fused feature vector is used as the input of the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. After the model training and validation are completed, the output results of the two sub-models are weighted and fused through a dynamic weighted fusion formula to obtain the final comprehensive risk probability.

[0256] The graded early warning module is configured with three-level early warning trigger rules and graded anti-shake strategy to realize differentiated early warning for low, medium and high risks; it is used to set up a three-level early warning mechanism and match graded anti-shake strategy according to the final comprehensive risk probability and real-time average arterial pressure monitoring value to realize differentiated early warning for perioperative hypotension.

[0257] The clinical intervention linkage module includes an evidence-based intervention knowledge base unit, an intervention suggestion output unit, and an intervention effect feedback and model iteration unit, which respectively realize the storage of evidence-based intervention basis, the output of individualized intervention suggestions, the evaluation of intervention effect, and the dual-track iterative optimization of the model; it is used to output individualized intervention suggestions based on the results of graded early warning combined with evidence-based medicine evidence, while collecting hemodynamic data after intervention to evaluate the intervention effect, constructing an intervention effect feedback closed loop, and completing the iterative optimization of the above two sub-models through a dual-track mechanism.

[0258] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0259] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting and warning the risk of perioperative hypotension based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Multi-source data acquisition, standardized synchronous acquisition of multi-dimensional data throughout the perioperative process, including preoperative static baseline clinical data, intraoperative stroke-level hemodynamic dynamic time-series data, and intraoperative anesthesia and operation-related event data; Step S2: Feature engineering processing, which involves preprocessing the collected multi-source data, extracting temporal features, and filtering features in sequence to obtain two types of feature results: preoperative core static feature set and intraoperative fusion feature vector. Step S3: AI risk prediction model construction, training and fusion. Construct a dual-model architecture of a preoperative hypotension risk stratification sub-model and an intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. Use the preoperative core static feature set as the input of the preoperative hypotension risk stratification sub-model and the intraoperative fused feature vector as the input of the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. After completing the model training and validation respectively, the output results of the two sub-models are weighted and fused through a dynamic weight fusion formula to obtain the final comprehensive risk probability. Step S4: Graded early warning. Based on the final comprehensive risk probability and real-time average arterial pressure monitoring value, a three-level early warning mechanism is set up and matched with a graded anti-shake strategy to achieve differentiated early warning of perioperative hypotension. Step S5: Clinical intervention linkage. Based on the graded early warning results and evidence-based medicine, individualized intervention recommendations are output. At the same time, hemodynamic data after intervention are collected to evaluate the intervention effect. An intervention effect feedback loop is constructed and the above two sub-models are iteratively optimized through a dual-track mechanism.

2. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, Step S1 specifically includes: S11: Preoperative static data collection, which includes four categories of preoperative static baseline clinical data: demographic characteristics, disease-related characteristics, laboratory test results, and perioperative-related characteristics. S12: Intraoperative dynamic time-series data acquisition: Intraoperative stroke-by-stroke dynamic hemodynamic time-series data of heart rate, mean arterial pressure, systolic blood pressure, diastolic blood pressure, cardiac output, peripheral vascular resistance, peripheral blood oxygen saturation, and respiratory rate are acquired through a stroke-by-stroke continuous non-invasive arterial pressure monitoring system at a fixed sampling frequency. S13: Intraoperative event and medication data collection, real-time collection of the entire process of anesthesia and medication, data on key nodes of endoscopic operation, and records of intraoperative adverse events and their management.

3. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, The data preprocessing in step S2 specifically includes: Outlier and missing value handling: For static data, multiple interpolation is used to handle missing values, and outlier samples exceeding the threshold range are removed based on the quartile method; for time-series hemodynamic signals, Gaussian derivative filter and Shannon energy envelope method are used to detect peak arterial blood pressure signals. Data standardization and resampling: The original arterial blood pressure signal was resampled to a uniform frequency using discrete Fourier transform, and all numerical features were normalized using the standard deviation standardization method. Sample construction and label definition: Using the preset mean arterial pressure index as the gold standard for hypotension, candidate time points are dynamically sampled and sample label assignment is completed by combining a three-level window structure, and the sample set division ensures that the data of the same patient does not cross sets.

4. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, The temporal feature extraction in step S2 specifically includes: Manual statistical feature extraction: For a 20-second arterial blood pressure input segment, 32-dimensional time-domain, frequency-domain, and nonlinear manual statistical features were extracted. Simultaneously, 8-dimensional time-series statistical features of heart rate, cardiac output, peripheral vascular resistance, and peripheral blood oxygen saturation were extracted, resulting in a total of 40-dimensional manual time-series features. Deep waveform feature extraction: The deep morphological features of the 20-second arterial blood pressure waveform are automatically extracted and the deep feature vector is output through the first 6 one-dimensional convolutional layers of the intraoperative stroke-level hypotension dynamic prediction sub-model.

5. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, The feature selection in step S2 specifically includes: Preoperative static feature screening: First, redundant features with correlation coefficients greater than 0.8 are removed using the Pearson correlation coefficient method. Then, a 28-dimensional core static feature set is selected by ranking the feature importance using a lightweight gradient booster, which serves as the input for the preoperative hypotension risk stratification sub-model. Intraoperative temporal feature selection: First, redundant features were removed using the Pearson correlation coefficient method. Then, the top 20 core temporal features were selected by ranking the importance of features using a gradient boosting tree. These 20 core temporal features were then concatenated with deep waveform features, preoperative risk stratification results, and anesthetic drug and operation event coding features to form an intraoperative fusion feature vector, which was used as the input to the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model.

6. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, The AI ​​risk prediction model construction, training, and fusion in step S3 specifically includes: A preoperative hypotension risk stratification sub-model and an intraoperative stroke-level hypotension dynamic prediction sub-model were constructed and trained and fused. The preoperative hypotension risk stratification sub-model adopted a lightweight gradient booster binary classification model. The input of the preoperative core static feature set was used to achieve hypotension risk stratification. The training was completed through 5-fold cross-validation. The intraoperative stroke-level hypotension dynamic prediction sub-model is based on a 7-layer one-dimensional convolutional neural network. It integrates arterial blood pressure time-series signals and multiple intraoperative auxiliary features for training, and outputs the probability of hypotension occurring in the next 5 minutes. The constructed AI risk prediction model integrates multi-source datasets. After pre-training on public datasets and fine-tuning on target domain datasets, it completes 5-fold cross-validation and generalization validation in accordance with the principles of patient-level data isolation and single variable control.

7. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, The dynamic weight fusion formula in step S3 is as follows: ； In the formula, For the final overall risk probability, The preoperative risk stratification probability output by the preoperative hypotension risk stratification sub-model. This is the real-time intraoperative prediction probability output by the dynamic prediction sub-model for intraoperative stroke-level hypotension. These are dynamic weighting coefficients.

8. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, The three-level early warning mechanism and graded anti-shake strategy in step S4 specifically include: Low-risk blue alert: The triggering condition is a final comprehensive risk probability of 0~30% and a real-time average arterial pressure ≥75mmHg. The system continuously monitors and displays the risk value and trend curve. It does not trigger active alerts and has no anti-shake restrictions. Medium-risk yellow alert: The triggering condition is that the final comprehensive risk probability is 30%~70% or the real-time average arterial pressure is in the gray range of 65~75 mmHg. The triggering rule is that the medium-risk standard is reached for 3 consecutive prediction cycles. The system will pop up a yellow alert and broadcast a voice reminder at the same time. High-risk red alert: The triggering condition is that the final comprehensive risk probability is ≥70% or the real-time mean arterial pressure is <65mmHg. The triggering rule is single-cycle triggering plus double-cycle verification. The alert is triggered immediately when the first prediction cycle reaches the standard. If the risk drops to medium or low risk in the subsequent two cycles, the alert will be automatically lifted. After triggering, the system will highlight the pop-up window and provide continuous voice alarm.

9. The method for predicting and warning the risk of perioperative hypotension based on artificial intelligence according to claim 1, characterized in that, The clinical intervention linkage in step S5 specifically includes: Construction of an evidence-based intervention knowledge base: A knowledge base was constructed based on the conclusions of multicenter, prospective, single-blind, randomized controlled clinical trials and guidelines for the management of perioperative hypotension. Individualized intervention recommendations are provided: for low-risk patients, recommendations for maintaining and continuously monitoring anesthesia protocols are provided; for medium-risk patients, recommendations for tiered interventions are provided; and for high-risk patients, recommendations for emergency intervention protocols are provided. Intervention effect feedback and model iteration: Real-time collection of hemodynamic data after intervention to evaluate the effect. If the risk of hypotension does not decrease within 5 minutes after intervention, the intervention plan is automatically upgraded. The model iteration adopts a dual-track mechanism of offline regular retraining and online incremental fine-tuning. Offline, desensitized clinical data is summarized every quarter to complete full retraining. Online, federated learning is used to complete incremental fine-tuning for single centers / special populations.

10. A perioperative hypotension risk prediction and early warning system based on artificial intelligence, characterized in that, The system for executing the artificial intelligence-based perioperative hypotension risk prediction and early warning method according to any one of claims 1-9 includes: The multi-source data acquisition module is used to standardize and synchronously acquire multi-dimensional data throughout the perioperative period. The multi-dimensional data includes preoperative static baseline clinical data, intraoperative stroke-level hemodynamic dynamic time-series data, and intraoperative anesthesia and operation-related event data. The feature engineering processing module is used to perform data preprocessing, temporal feature extraction, and feature filtering on the collected multi-source data in sequence to obtain two types of feature results: preoperative core static feature set and intraoperative fusion feature vector. The AI ​​risk prediction model construction, training and fusion module is used to build a dual-model architecture of a preoperative hypotension risk stratification sub-model and an intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. The preoperative core static feature set is used as the input of the preoperative hypotension risk stratification sub-model, and the intraoperative fused feature vector is used as the input of the intraoperative stroke-by-stroke hypotension dynamic prediction sub-model. After the model training and validation are completed respectively, the output results of the two sub-models are weighted and fused through a dynamic weight fusion formula to obtain the final comprehensive risk probability. The graded early warning module is used to set up a three-level early warning mechanism and match a graded anti-shake strategy based on the final comprehensive risk probability and real-time average arterial pressure monitoring value, so as to realize differentiated early warning of perioperative hypotension. The clinical intervention linkage module is used to output individualized intervention recommendations based on the results of graded early warning and evidence-based medicine. At the same time, it collects hemodynamic data after intervention to evaluate the intervention effect, constructs a closed loop for intervention effect feedback, and completes the iterative optimization of the above two sub-models through a dual-track mechanism.

Images