A postoperative hypothermia early warning method and system

Abstract

The present application relates to the technical field of postoperative body temperature early warning, in particular to a postoperative hypothermia early warning method and system. The method comprises the following steps: acquiring and preprocessing patient multi-source data, extracting features from the multi-source data, adopting a cross-modal attention mechanism, calculating the weight of each extracted feature vector in real time for adaptive feature fusion, performing constraint optimization on the fused features, and outputting a high-dimensional fused feature vector; constructing a hybrid time series prediction model based on the high-dimensional fused feature vector to output the core body temperature prediction value and the hypothermia risk index RI of the patient at a certain future time period; and based on the core body temperature prediction value and the hypothermia risk index RI, formulating a risk standard and a quantitative intervention strategy to realize postoperative hypothermia early warning of the patient. The present application effectively solves the information island problem existing in the traditional single measurement of body surface temperature data, expands the monitoring dimension, and thus improves the accuracy of postoperative risk assessment.

Description

Technical Field

[0001] This invention relates to the field of postoperative body temperature warning technology, and in particular to a method and system for warning of postoperative hypothermia. Background Technology

[0002] Postoperative hypothermia is a common complication in surgical procedures, with an incidence rate ranging from 40% to 65%, and it is directly related to patient prognosis: for every 1°C decrease in core body temperature, thrombin activity decreases by 10% to 15%, intraoperative blood loss increases by 16% to 20%, anesthesia recovery time is prolonged by 30% to 50%, the risk of postoperative infection increases to 2 to 3 times, and the average stay in the ICU increases by 1.5 days. In severe cases, it can even lead to life-threatening complications such as arrhythmia and chills. Therefore, real-time, accurate, and proactive early warning of postoperative hypothermia risk is crucial for reducing the incidence of complications and improving surgical safety.

[0003] Currently, the main methods for monitoring and predicting the risk of postoperative hypothermia include the following categories, but all of them have significant technical limitations: 1. Manual Intermittent Monitoring: This method is still widely used in primary hospitals. It relies on medical staff manually recording body temperature every 30 to 60 minutes using a thermometer or monitor, and combining this with their personal experience to assess risk. The advantage of this method is its simplicity, but it has three major problems: First, there is a time blind spot; intermittent monitoring cannot capture sudden drops in body temperature. Second, the accuracy is insufficient; the error between axillary temperature and core body temperature can reach ±0.8℃, easily leading to misjudgment of risk. Third, the false negative rate is high; due to the limited time and energy of medical staff, the false negative rate during the postoperative recovery period is as high as 29.5%, which cannot meet the needs of continuous monitoring.

[0004] 2. Single-modal static monitoring method: This method deploys a single sensor during surgery to collect body temperature data in real time and triggers an alert when the body temperature drops below 36℃. Although this method achieves continuous monitoring, it suffers from information silos: First, it relies solely on body temperature parameters, ignoring dynamic influencing factors such as intraoperative blood loss, infusion rate (which can increase the rate of temperature drop by 0.2℃ / h), and depth of anesthesia; second, the sensors are susceptible to interference. Infrared sensors on the body surface can have errors of ±0.5℃ due to sweat, and esophageal probes have a false alarm rate as high as 28% due to respiratory airflow, making it impossible to achieve multi-factor collaborative alerts.

[0005] 3. Traditional Static Predictive Model Method: This method, based on algorithms such as Logistic Regression and Random Forest, inputs static parameters such as patient age, BMI, and underlying diseases to predict the probability of hypothermia risk preoperatively. However, this method has drawbacks, including poor dynamic adaptability: First, it does not incorporate intraoperative dynamic parameters, resulting in a postoperative risk prediction accuracy of only 70.3%; second, the model lacks temporal prediction capabilities, failing to provide early warnings with a lag time exceeding 15 minutes, missing the optimal intervention window; and third, it does not consider the coupling relationship between parameters, leading to a false alarm rate as high as 22.4% in scenarios with multiple factors.

[0006] In summary, existing technologies for monitoring and predicting postoperative hypothermia risk generally suffer from four major drawbacks: "single monitoring dimensions, delayed dynamic response, ambiguous risk quantification, and missed intervention window," failing to meet the clinical needs for "advanced early warning and precise intervention." Summary of the Invention

[0007] The present invention aims to at least improve one of the technical problems existing in the prior art. To this end, the present invention proposes a method and system for early warning of postoperative hypothermia.

[0008] The technical solution of the present invention is as follows: A method for early warning of postoperative hypothermia, comprising: S1: Acquire multi-source patient data and preprocess the multi-source data, which includes intraoperative core body temperature data, body surface temperature distribution data, intraoperative dynamic parameter data, and patient static baseline data. S2: Extract features from the multi-source data, employ a cross-modal attention mechanism, calculate the weights of each extracted feature vector in real time to perform adaptive feature fusion, perform constraint optimization on the fused features, and output a high-dimensional fused feature vector. S3: Construct a hybrid time-series prediction model based on the high-dimensional fusion feature vector to output the predicted core body temperature and low body temperature risk index (RI) of the patient for a certain period in the future; S4: Based on the core body temperature prediction value and the hypothermia risk index RI, formulate risk standards and quantitative intervention strategies to achieve early warning of hypothermia after surgery.

[0009] In one possible technical solution, further, in S2, feature extraction of the multi-source data specifically includes: S211: Principal Component Analysis (PCA) is used to reduce the dimensionality of multiple features in the patient's static baseline data. The number of principal components is determined by scree plots to output a three-dimensional static feature vector. ∈ ; S212: A one-dimensional convolutional neural network (1D-CNN) is used to extract local trend features from the time-series variables of multiple dynamic parameters in the intraoperative core body temperature data, outputting a 64-dimensional dynamic local feature vector. ∈ ; S213: A Transformer encoder is used to capture long-term dependencies in the temporal data of intraoperative dynamic parameter data. The association weights of data at different time steps are calculated through a self-attention mechanism, and a 128-dimensional long-term feature vector is output. ∈ ; S214: The U-Net semantic segmentation network is used to extract temperature distribution features from infrared thermal images in body surface temperature distribution data. The enhanced thermal image is input, and the encoder extracts temperature gradient features. The decoder generates pixel-level temperature anomaly region masks and finally outputs three core features: mean body surface temperature, standard deviation of body surface temperature, and percentage of high-temperature area, which constitute a three-dimensional body surface feature vector. ∈ .

[0010] In one possible technical solution, S2 further includes: S221: Define the attention weight matrix , in For static feature weights, For dynamic local feature weights, For long-term dependent feature weights, Weights for body surface features; S222: Weight the 3D static feature vector, dynamic local feature vector, long-term feature vector, and 3D surface feature vector according to their respective weights, and then concatenate them to output a 256-dimensional fused feature vector. ∈ The formula is: Where ⊕ represents the vector concatenation operation.

[0011] In one possible technical solution, further, in S2, the constraint optimization of the fusion features specifically includes body temperature coupling constraints, bleeding volume constraints, and anesthesia depth constraints, which are used to calculate the total constraint loss.

[0012] In one possible technical solution, further, the total loss of the loss function in the hybrid time-series prediction model... Represented as: in, , , All are weighting coefficients. To constrain losses; The loss in body temperature prediction, denoted as RMSE (Reflection on the Deviation Between Predicted and Actual Body Temperature Values), is expressed as: For the predicted core body temperature of the i-th sample, Let N be the actual core body temperature of the i-th sample, and N be the number of samples in the test set. Binary cross-entropy loss (BCE) is used to measure the classification accuracy of the risk index. in, =1 indicates hypothermia. =0 indicates normal body temperature, and RI is the risk index for low body temperature.

[0013] In one possible technical solution, further, in S3, the hypothermia risk index RI is calculated in two steps: weighted summation and sigmoid normalization, specifically as follows: Calculate the unnormalized risk value : in, The predicted core body temperature normalized value for the i-th sample is calculated using a piecewise function; when 35.0℃≤ At ≤36.0℃, ;when At <35.0℃, ;when When the temperature is >36.0℃, ; For the first The normalized cumulative blood loss for each sample was calculated using a piecewise function; when 200ml ≤ When ≤1000ml, ;when When >1000, ;when At 200ml, ; For the first The normalized infusion rate for each sample was calculated using a piecewise function; when 5 ml / (kg·h) ≤ ≤ hour, ;when When >15ml / (kg·h), ;when At 5 ml / (kg·h), ; For the first The normalized BIS values ​​for each sample were calculated using a piecewise function; when 30 ≤ BIS ≤ 60... When BIS < 30, When BIS > 60, ; Sigmoid normalization: Will Mapping to the [0,1] interval yields the low body temperature risk index for the i-th sample. The formula is: The steepness coefficient k=8. Tests showed that when k=8... The highest discrimination was observed in the 0.2-0.8 range, with the threshold value being [missing information]. , That is, the quantitative risk of postoperative hypothermia occurring in the i-th sample within the next 15 minutes. The closer it is to 1, the higher the risk.

[0014] A postoperative hypothermia early warning system, comprising: The acquisition module is used to acquire multi-source patient data and preprocess the multi-source data, which includes intraoperative core body temperature data, body surface temperature distribution data, intraoperative dynamic parameter data, and patient static baseline data. The feature extraction and fusion module is used to extract features from the multi-source data. It adopts a cross-modal attention mechanism to calculate the weights of each extracted feature vector in real time for adaptive feature fusion, performs constraint optimization on the fused features, and outputs a high-dimensional fused feature vector. The calculation module is used to construct a hybrid time-series prediction model based on the high-dimensional fusion feature vector, so as to output the predicted core body temperature value and low body temperature risk index (RI) of the patient for a certain period of time in the future. Risk warning module: Based on the core body temperature prediction value and the hypothermia risk index RI, risk standards and quantitative intervention strategies are formulated to achieve postoperative hypothermia warning for patients.

[0015] The postoperative hypothermia early warning method and system according to the present invention have the following advantages compared with the prior art: Compared with existing technologies, this invention has the following breakthrough advantages: (1) By using a multimodal static and dynamic data fusion strategy, the patient's static baseline data, body surface temperature data, and intraoperative dynamic data are correlated and fused in the spatiotemporal dimensions. This method effectively solves the "information silo" problem that exists in traditional methods that only measure body surface temperature data, expands the monitoring dimensions, and thus improves the accuracy of postoperative risk assessment.

[0016] (2) A feature fusion module based on the attention mechanism was designed. This module can calculate the weight of each parameter in real time during the dynamic changes of intraoperative parameters, avoiding the error caused by fixed weights. At the same time, combined with the LSTM-GRU hybrid network to capture the temporal trend of body temperature, the core body temperature of the next 15 minutes can be predicted with a prediction error of less than 0.1℃ and an advance warning time of 15-30 minutes, which is 10-15 minutes earlier than the traditional method, providing a sufficient time window for clinical intervention.

[0017] (3) A hypothermia risk index (RI) integrating "predicted body temperature + dynamic parameters" was constructed, elevating the risk status assessment from "qualitative" to "0-1 continuous quantification" level, and a five-level risk standard was formulated in conjunction with clinical scenarios. The false negative rate of this method was reduced to below 2.1%, and the false positive rate was reduced to below 3.3%, significantly lower than the false negative rate of 29.5% and the false positive rate of 22.4% of manual monitoring, and the false negative rate of 18.7% and the false positive rate of 28% of single-modality monitoring. In addition, each risk level corresponds to a clear quantitative intervention measure, which can be directly applied to clinical practice.

[0018] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the postoperative hypothermia early warning method as described above.

[0019] A computer storage medium storing instructions that, when executed on a computer, cause the computer to perform the postoperative hypothermia early warning method described above.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1This is a flowchart of a postoperative hypothermia early warning method according to an embodiment of the present invention; Figure 2 This is a flowchart of the postoperative hypothermia early warning method according to an embodiment of the present invention; Figure 3 This is a feature fusion module diagram of a postoperative hypothermia early warning method according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a hybrid time-series prediction model for a postoperative hypothermia early warning method according to an embodiment of the present invention. Figure 5 This is a trend chart of postoperative hypothermia risk index changes in Case 1 according to an embodiment of the present invention; Figure 6 This is a five-level risk classification confusion matrix diagram according to Example 2 of the present invention; Figure 7 This is a schematic diagram of a postoperative hypothermia early warning system according to an embodiment of the present invention. Detailed Implementation

[0023] The embodiments of the present invention are described in detail below. The embodiments described with reference to the accompanying drawings are exemplary. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0024] It should be noted that when a component is said to be "fixed to" another component, it can be directly attached to the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects and not to describe a particular order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, it may include a series of steps or units, or optionally, steps or units not listed, or other steps or units inherent to these processes, methods, products, or devices.

[0027] The accompanying drawings show only the portions relevant to this application, not all of them. Before discussing exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process may correspond to a method, function, procedure, subroutine, subprogram, etc.

[0028] The terms “component,” “module,” “system,” “unit,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a unit can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, a thread of execution, a program, and / or distributed between two or more computers. Furthermore, these units can be executed from various computer-readable media on which various data structures are stored. Units can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from a second unit interacting with another unit between a local system, a distributed system, and / or a network; for example, the Internet interacting with other systems via signals).

[0029] Example 1 like Figure 1 As shown, this embodiment provides a method for early warning of postoperative hypothermia, which includes: S1: Acquire multi-source patient data and preprocess the multi-source data, which includes intraoperative core body temperature data, body surface temperature distribution data, intraoperative dynamic parameter data, and patient static baseline data.

[0030] It should be noted that, in this embodiment, the acquisition of multivariate patient data specifically includes the following: S11, Intraoperative Core Temperature Data Acquisition: Core body temperature was collected using a Mindray BeneHeart D6 multi-functional monitor with an esophageal or rectal temperature probe. The sampling frequency was set to 1 time / minute, the measurement range was 32℃ - 42℃, and the accuracy reached ±0.1℃. The collected data was transmitted in real time to the postoperative monitoring cloud platform via the HL7 protocol and stored in the time-series database InfluxDB.

[0031] S12, Body Surface Temperature Distribution Data Acquisition: FLIR A655sc infrared thermal imagers were deployed on both sides of the operating table. These imagers have a resolution of 640×512, a temperature measurement range of -20℃ to 150℃, an accuracy of ±0.05℃, and a frame rate of 1 frame / 30 seconds. This thermal imager was used to acquire images of the patient's chest, abdomen, and limbs' body surface temperature distribution. The average body surface temperature was extracted using OpenCV. with temperature standard deviation The YOLOv12 algorithm was used to identify occlusions. When the confidence level was greater than 0.8, pixels in that area were removed to eliminate the influence of clothing occlusion.

[0032] S13, Intraoperative dynamic parameter acquisition: Blood loss: A combined "volume-based method + weighing method" was used for monitoring. Fluid volume was collected using a suction device and weighed using gauze. The blood loss per episode was recorded in real time via an intelligent weighing RS485 bus. The sampling frequency is 1 time / 5 minutes.

[0033] Infusion rate: Employs B. Braun Infusomat Space intelligent infusion pump, with an infusion rate range of 1–1200 ml / h and an accuracy of ±1 ml / h, recording the hourly infusion volume in real time. Sampling frequency: 1 time / minute; Anesthesia depth: BIS value was monitored using a Philips S / 5 monitor, ranging from 0 to 100, with a sampling frequency of 1 time / minute. BIS < 60 was considered to be a deep anesthesia depth. S14, Patient Static Basic Data Collection: Retrieve preoperative patient data through the hospital's HIS system, specifically including: age (in years), body mass index (BMI) (in kg / m²), underlying diseases (pre-existing = 1, no disease = 0), and preoperative core body temperature. (Unit: °C), Surgical type (General Surgery = 1, Orthopedics = 2, Thoracic Surgery = 3, Obstetrics and Gynecology = 4). After data anonymization, it was matched and associated with dynamic data through patient ID.

[0034] It should be noted that, in this embodiment, the preprocessing of the multi-source data specifically includes using an STM32F407 microcontroller as an edge gateway to uniformly synchronize the clocks of all devices, ensuring a timestamp error of <100ms, and performing four steps of data processing: 1. Outlier removal: The 3σ principle is used to remove outliers from core body temperature. Average body surface temperature Single bleeding volume Hourly infusion volume BIS value, BMI, patient age, and preoperative core body temperature. For variables, first calculate their mean in historical clinical data. with standard deviation Remove those exceeding [ The values ​​in the interval were statistically verified, and the proportion of outliers for all continuous variables was <0.5%, and the data after removal were consistent with clinical physiological patterns. 2. Missing value imputation: LSTM interpolation was used to imput short-term missing data (less than 5 minutes) based on numerical predictions of 10 time steps before and after the missing data. The imputation error for body temperature variables was <0.1℃, blood loss variables was <5ml, infusion rate was <1ml / h, BIS value was <1, and BMI was <0.1kg / m². 3. Data Standardization: Normalize all variables to the [0,1] interval, using the following formula: Where x represents the original data. , For the corresponding variable, the clinically reasonable range 4. Feature Derivation: Calculate dynamically derived features, including: (1) Core body temperature change rate / (°C / h): Body temperature change over the past 30 minutes / time, i.e.: / =( ) / 0.5 in, Core body temperature at time t The interval is 0.5 hours. (2) Cumulative blood loss (ml): The cumulative value of single-session blood loss from the start of surgery to the current moment, i.e.: n represents the number of samples taken before the current time. (3) Infusion temperature influence coefficient The formula is based on the infusion rate and the patient's weight: weight refers to the patient's weight, in kg.

[0035] In this embodiment, the rationality of the derived logic is verified using 1000 historical data examples: / The consistency with actual body temperature changes reached 98.2%. The error between the recorded intraoperative blood loss and the actual blood loss is less than 3%. The correlation R² with the rate of body temperature decrease was 0.89, ensuring that the derived variables could support subsequent model calculations. The preprocessed data was transmitted to the cloud via a 5G medical private network and stored in a time-series database and a relational database (MySQL) for use by subsequent modules.

[0036] S2: Feature extraction is performed on the multi-source data. A cross-modal attention mechanism is used to calculate the weights of each extracted feature vector in real time for adaptive feature fusion. The fused features are then constrained and optimized to output a high-dimensional fused feature vector. Specifically, feature extraction on the multi-source data includes: S211: Principal component analysis (PCA) is used to reduce the dimensionality of multiple features in the patient's static baseline data. The number of principal components is determined by the Scree plot to output a three-dimensional static feature vector. ∈ ; Specifically, Principal Component Analysis (PCA) was used to perform dimensionality reduction on five static features: patient age, BMI, underlying diseases, preoperative core body temperature, and surgical type. During the process, a screen plot was used to determine the exact number of principal components, ensuring that the retained principal components covered more than 95% of the original information. After processing, a three-dimensional static feature vector was finally output, denoted as... ∈ The variance contribution rates of each principal component were as follows: the first principal component (composed of BMI and preoperative core body temperature) contributed 45.2%; the second principal component (corresponding to age) contributed 32.8%; and the third principal component (covering surgical type and underlying disease) contributed 17.0%.

[0037] S212: A 1D-CNN network is used to extract local trend features from the time-series variables of multiple dynamic parameters in the intraoperative core body temperature data, outputting a 64-dimensional dynamic local feature vector. ∈ ; It should be noted that in this embodiment, a one-dimensional convolutional neural network (1D-CNN) is used to extract local trend features from seven time-series variables in the dynamic parameters: core body temperature, infusion rate, BIS value, single blood loss, core body temperature change rate, cumulative blood loss, and infusion temperature influence coefficient. The specific network structure is as follows: The time step of the input layer is set to 30, and the feature dimension is 7. The next layer is convolutional layer 1, which has 32 filters, a kernel size of 3, uses the ReLU activation function, and has padding set to 1. Next is the max-pooling layer, with a pooling window size of 2 and a stride of 2; Next is convolutional layer 2, with 64 filters and a kernel size of 3. It also uses the ReLU activation function and padding is set to 1. Next is the fully connected layer with 64 neurons, using the ReLU activation function.

[0038] Finally, a 64-dimensional dynamic local feature vector is output. ∈ .

[0039] S213: A Transformer encoder is used to capture long-term dependencies in the temporal data of intraoperative dynamic parameter data. The association weights of data at different time steps are calculated through a self-attention mechanism, and a 128-dimensional long-term feature vector is output. ∈ ; It should be noted that in this embodiment, a Transformer encoder architecture is used to process nearly two hours of time-series data to effectively capture long-term dependencies in the data. This encoder structure specifically includes two multi-head attention mechanism modules, each followed by a LayerNorm (layer normalization) layer and a fully connected layer. During processing, a self-attention mechanism is used to dynamically calculate the association weights between data at different time steps, thereby fully exploring the potential dependency patterns in the time-series data. Finally, a 128-dimensional long-term feature vector is output, denoted as... ∈ .

[0040] S214: The U-Net semantic segmentation network is used to extract temperature distribution features from infrared thermal images. Specifically, the enhanced thermal image is used as input, and an encoder extracts temperature gradient features. A decoder then generates pixel-level masks for temperature anomaly regions. Finally, three core features are output: mean body surface temperature, standard deviation of body surface temperature, and percentage of high-temperature areas. These features are used to construct a three-dimensional body surface feature vector. ∈ The test results showed that the vector's characterization error for uneven body surface temperature was less than 0.05℃.

[0041] It is worth noting that a cross-modal attention mechanism is used in stage S2 to achieve adaptive fusion of the extracted feature vectors. This mechanism can calculate the weights corresponding to each feature vector in real time, and then complete the adaptive feature fusion operation based on the weights. Its specific construction process can be divided into the following four key steps: S221: Define the attention weight matrix ,in: The static feature weights are dynamically adjusted based on the patient's preoperative risk level. The preoperative risk level is determined by BMI and underlying diseases; a BMI < 18.5 or the presence of underlying diseases indicates high risk. =0.15; BMI ≥ 18.5 and no underlying diseases are considered low risk. =0.05; the rest are medium risk. =0.1.

[0042] Dynamic local feature weights: determined based on the core body temperature change rate. / When < -0.3℃, =0.35; -0.3℃ < / When < -0.1℃, =0.3; / At -0.1℃, =0.25.

[0043] Long-term dependent feature weights: Based on a joint determination of surgical duration and cumulative blood loss, when the surgical duration is >3 hours and the cumulative blood loss is >500ml, =0.3; when only a single condition is met, =0.25; if none of the above conditions are met, =0.2.

[0044] Weights for body surface features: determined based on the standard deviation of body surface temperature. When >0.5℃, =0.2; 0.3℃ < When <0.5℃, =0.15; When <0.3℃, =0.1; Weights satisfy =1, and is updated every minute based on the latest data to ensure that the weight matches the patient's real-time status.

[0045] S222: Transform the three-dimensional static feature vector Dynamic local feature vectors Long-term eigenvectors and three-dimensional surface feature vectors According to their respective weights After weighting, a vector concatenation operation is performed to finally output a 256-dimensional fused feature vector. ∈ The formula is: Here, the symbol “⊕” indicates a vector concatenation operation, while the weighting process is implemented using Element-Wise multiplication. After this fusion operation, the information entropy of the resulting features is reduced by 18.3%, and the feature redundancy is reduced by 25.6%.

[0046] It should be noted that the constraint optimization work carried out in stage S2 of this embodiment specifically covers three aspects: body temperature coupling constraint, blood loss constraint, and anesthesia depth constraint. The purpose is to calculate the total constraint loss through these constraints, thereby ensuring that the fusion features conform to clinical physiological logic. The specific details are as follows: 1. Body temperature coupling constraint: There is a close physiological relationship between core body temperature and body surface temperature. Under normal physiological conditions, both must meet certain conditions. - The condition is <1.5℃. If the actual measured value exceeds this range, it indicates that the current characteristic state may deviate from the normal physiological state. In this case, the weight of the body surface characteristics needs to be adjusted. A penalty will be imposed, reducing it by 50%. Simultaneously, the constraint loss will be calculated using the following formula: in, Represents core body temperature. Represents body surface temperature. For relevant parameters (specific values ​​can be set according to actual conditions).

[0047] 2. Blood loss control: When the cumulative bleeding When the volume exceeds 1000ml, the risk of hypothermia increases exponentially. In such emergency situations, to better highlight the importance of dynamic local features in reflecting the patient's condition, it is necessary to enforce a higher weighting for dynamic local features. Increase it to 0.4, which means increasing the weight by 0.05 to 0.1 from the original weight. Simultaneously, calculate the constraint loss using the following formula: 3. Anesthesia depth constraint: When the bispectral index (BIS) on an electroencephalogram (EEG) is less than 40, it indicates that the patient's thermoregulatory center is severely suppressed, and their thermoregulatory function is seriously impaired. To more accurately reflect the impact of this physiological change on the feature, it is necessary to forcibly adjust the long-term feature weights. Increase it to 0.35. Simultaneously, calculate the constraint loss using the following formula: In this constrained optimization system, the weights Unconstrained, no constraint loss occurs. Total constraint loss. The calculation method is as follows: This total constraint loss will be incorporated into the total loss function of model training. In this way, it is ensured that the fused features do not deviate from the clinical physiological logic during model training, thereby improving the accuracy and reliability of the model.

[0048] S3: Construct a hybrid time-series prediction model based on the high-dimensional fusion feature vector to output the predicted core body temperature value and low body temperature risk index (RI) of the patient for a certain period in the future.

[0049] In this embodiment, to achieve the goal of accurately predicting the risk index and body temperature, the fused feature vector output in stage S2 is used. As input data, a hybrid time-series model was carefully constructed. This model can effectively predict core body temperature and the hypothermia risk index (RI) for the next 15 minutes based on the fused features of the input data, specifically including the following: Model structure design: In this model, a hybrid network of LSTM and GRU is used to construct the model structure. This design combines the advantages of LSTM in capturing long-term dependencies with the fast convergence advantage of GRU, aiming to improve the model's performance in prediction tasks. The entire network structure can be divided into three parts: input layer, hidden layer, and output layer, as detailed below. Input layer: Input the fused feature vector from the past 30 minutes The input data has a specific format, and its dimensions are represented as [ ] = [32, 30, 256]. Wherein, =32 indicates the batch size, which is the number of data samples input into the model for training or prediction each time; =30 is the time step, representing the time span covered by the input data; =256 is the feature dimension, which is the number of features contained in each time step.

[0050] Hidden layer: The hidden layer consists of one LSTM layer and one GRU layer, with the following specific configuration: LSTM layer: This layer has 256 hidden units to better capture complex patterns and long-term dependencies in the input data. Activation function selection... The activation function effectively alleviates the vanishing gradient problem and accelerates model convergence. The dropout probability is set to 0.2 to prevent overfitting and improve the model's generalization ability. For weight initialization, the weights of the input gate, forget gate, and output gate are initialized using a Xavier uniform distribution. This initialization method helps maintain stable variance of activation values ​​in each layer during the early stages of training, thereby accelerating model convergence. After processing by the LSTM layer, the output dimension is [32, 30, 256]. Layer: To further optimize the model training process and avoid the vanishing gradient problem, a layer was added after the LSTM layer. This layer normalizes the output of the LSTM layer, with parameter ε=1. By adjusting the numerical stability during the normalization process, we can ensure that the model can stably perform gradient propagation and parameter updates during training.

[0051] GRU layer: In After the first layer, a GRU layer is connected. The GRU layer has 128 hidden units, and also uses... As the activation function, the dropout probability is also set to 0.2. The weights of the update and reset gates are initialized using a He normal distribution; this initialization method is suitable for... The activation function enables the neuron's output to have an appropriate variance in the early stages of training, which helps the model converge faster. After processing by the GRU layer, the output dimension is [32, 30, 128].

[0052] Output layer: The output layer is divided into two parallel branches, which share the features of the hidden layer. The specific design is as follows: Body Temperature Prediction Branch: This branch employs a linear activation function. It takes the output of the last time step of the GRU layer as input, performs feature mapping and transformation through a fully connected layer, and ultimately outputs the core body temperature prediction value for the next 15 minutes. This design enables the model to accurately predict core body temperature over a future period based directly on the temporal characteristics of the input.

[0053] Risk Index Branch: This branch first compresses the feature dimension to 32 using a fully connected layer to reduce feature complexity and highlight key information. Then, it employs... The activation function maps the output to the 0-1 interval to obtain the hypothermia risk index RI. The formula for the function is ,in This is the output of the fully connected layer. This design gives the risk index a clear range of values ​​and intuitive interpretability, facilitating subsequent assessment and analysis of hypothermia risk.

[0054] Model training strategy: To construct a dataset suitable for this model, data from 1200 surgical patients at a tertiary hospital between January 2024 and January 2025 were collected. These patients covered multiple departments, including general surgery (400 cases), orthopedics (300 cases), thoracic surgery (300 cases), and obstetrics and gynecology (200 cases), and were broadly representative. Regarding patient characteristics, the age range was 18–85 years, and the BMI range was 15.2–38.5 kg / m², fully considering the impact of different ages and body mass indexes on surgery and body temperature changes.

[0055] Based on the collected data, a total of 60,000 time-series samples were generated. Each sample contains 30 minutes of input features and 15 minutes of prediction targets. This design enables the model to predict core body temperature and hypothermia risk index for the future based on feature information from the past period. To ensure the effectiveness of model training and evaluation, the dataset was randomly divided into a training set (48,000 samples) and a test set (12,000 samples) in an 8:2 ratio. Furthermore, the training set employed 10-fold cross-validation to optimize hyperparameters. Through multiple partitions and validations, the model's performance under different parameter settings can be more comprehensively evaluated, thereby selecting the optimal hyperparameter combination and improving the model's generalization ability.

[0056] Loss function design Total loss function Loss Prediction by Body Temperature Risk classification loss With constraint loss It is obtained by weighted summation of three parts, and the specific formula is as follows: in, =0.4、 =0.3、 =0.1 represent the weights of each loss component. By setting the weights appropriately, the impact of different losses on model training can be balanced, enabling the model to achieve good performance in both body temperature prediction and risk classification tasks.

[0057] Body temperature prediction loss The root mean square error (RMSE) is used to measure the deviation between the predicted and actual body temperature values. The formula is as follows: In the formula, Let i be the predicted core body temperature value for the i-th sample. Let be the actual core body temperature of the i-th sample, and N=12000 be the number of samples in the test set. RMSE can intuitively reflect the degree of difference between the predicted and actual values; a smaller RMSE value indicates that the model's body temperature prediction is more accurate.

[0058] Risk Classification Loss The binary cross-entropy loss (BCE) is used to measure the classification accuracy of the risk index, and its formula is as follows: in, =1 indicates hypothermia. =0 indicates normal body temperature. Let be the low body temperature risk index for the i-th sample. Binary cross-entropy loss (BCE) is an effective measure of a model's performance on classification tasks; a smaller BCE value indicates that the model classifies low body temperature risk more accurately.

[0059] Constraint loss The constraint loss defined in S2 to S3 is adopted to ensure that the model output conforms to clinical logic, so that the model can follow medical common sense and rules in the prediction process, thereby improving the credibility and practicality of the prediction results.

[0060] Training parameters: During model training, a series of optimization strategies and parameter settings were employed to ensure that the model could converge to the optimal solution efficiently and stably. Among these, Optimizer: The Adam optimizer is selected, which can adaptively adjust the learning rate and combines momentum gradient descent with... Its advantages include good performance when handling complex optimization problems. The initial learning rate is set to 1. By gradually adjusting the learning rate, it is possible to achieve rapid convergence in the early stages of training, while avoiding getting stuck in local optima in the later stages.

[0061] Learning rate decay: The learning rate is decayed to 0.5 of its original value every 20 rounds. This learning rate decay strategy can gradually reduce the learning rate as training progresses, allowing the model to finely adjust parameters in the later stages of training and further improve the model's performance.

[0062] Number of iterations: Set to 100 rounds. Through multiple iterations, the model can fully learn the features and patterns in the data, thereby improving the accuracy of predictions.

[0063] Batch size: Set to 32. A smaller batch size can increase the randomness of model training, which helps to improve the model's generalization ability, and can also reduce memory usage to some extent.

[0064] Early stopping strategy: When the validation set... <1 Furthermore, training is stopped if there is no decrease in performance for five consecutive rounds. The early stopping strategy can effectively prevent model overfitting by stopping training in time when the model performance reaches its optimal state, thus avoiding performance degradation on the test set due to overtraining.

[0065] Through the setting and optimization of the above training parameters, the final model achieved excellent performance on the test set, with a mean absolute error (MAE) of body temperature prediction of <0.1℃, a risk classification accuracy of >96%, and a constraint satisfaction rate of >98%, which can provide strong support for body temperature monitoring and hypothermia risk assessment in clinical surgery.

[0066] It should be noted that in S3, the hypothermia risk index RI is calculated in two steps: weighted summation and sigmoid normalization. Specifically, it includes: Step 1: Calculate the unnormalized risk value The calculation formula is as follows: in, The weights for each dynamic parameter were obtained by evaluating the feature importance of 5000 sets of historical risk data using the random forest algorithm. The feature importance ranking was as follows: predicted core body temperature (35.2%) > cumulative blood loss (25.1%) > infusion rate (20.3%) > BIS value (19.4%). Predicted normalized core body temperature The calculation is performed using a piecewise function, and the specific rules are as follows: when hour, In this segment, the closer the predicted body temperature is to the predicted temperature... The closer the normalized value is to 1, the higher the risk of hypothermia; the closer it is to 1, the higher the risk of hypothermia. The closer the normalized value is to 0, the lower the risk.

[0067] when At <35.0℃, This indicates that when the predicted body temperature is below 35.0℃, the individual is at higher risk of hypothermia.

[0068] when When the temperature is >36.0℃, This indicates that when the predicted body temperature is above 36.0℃, the risk of hypothermia is low.

[0069] Normalized value of cumulative blood loss Calculated using a piecewise function: When 200ml≤ When ≤1000ml, Within this range, the greater the cumulative bleeding volume and the closer the normalized value is to 1, the higher the risk of hypothermia; conversely, the closer it is to 0, the lower the risk.

[0070] when When >1000ml, This indicates a higher risk of hypothermia when the cumulative blood loss exceeds 1000ml.

[0071] when At 200ml, This indicates that the risk of hypothermia is low when the cumulative blood loss is less than 200ml.

[0072] Normalized value of infusion rate Calculated using a piecewise function: When 5ml / (kg·h)≤ ≤ hour, Within this range, the faster the infusion rate (the greater the infusion volume per unit body weight), the closer the normalized value is to 1, and the higher the risk of hypothermia; conversely, the closer the normalized value is to 0, the lower the risk.

[0073] when When >15ml / (kg·h), This indicates that when the infusion rate exceeds a certain range, the patient is at higher risk of hypothermia.

[0074] when At 5 ml / (kg·h), This indicates that the risk of hypothermia is lower when the infusion rate is below a certain range.

[0075] BIS value normalized value Calculated using a piecewise function: Calculation using piecewise functions; when 30 ≤ When ≤60, Within this range, the lower the BIS value, the deeper the anesthesia; the closer the normalized value is to 1, the higher the risk of hypothermia; conversely, the closer the value is to 0, the lower the risk.

[0076] when <30 hours, This indicates a high risk of hypothermia when the depth of anesthesia is too high.

[0077] when >60, This indicates that the depth of anesthesia is appropriate or that the risk of hypothermia is low during the patient's recovery process. Sigmoid normalization: In this embodiment, Mapping to the [0,1] interval yields the low body temperature risk index for the i-th sample. The formula is: The steepness coefficient k=8. Tests showed that when k=8... The highest discrimination was observed in the 0.2-0.8 range, with the threshold value being [missing information]. , That is, the quantitative risk of postoperative hypothermia occurring in the i-th sample within the next 15 minutes. The closer it is to 1, the higher the risk.

[0078] S4: Based on the core body temperature prediction value and the hypothermia risk index RI, formulate risk standards and quantitative intervention strategies to achieve early warning of postoperative hypothermia in patients. Specifically, this includes: Level 1 Risk (Safety Status): Risk Index (RI) less than 0.2, predicted core body temperature not lower than 36.5℃, cumulative blood loss less than 200ml, infusion rate per kilogram of body weight less than 5ml / (kg·h), bispectral index (BIS) value greater than 80, and operation duration less than 60 minutes. Monitoring frequency is every 15 minutes. Intervention measures mainly include routine warming measures to maintain the ambient temperature between 24-26℃, without the need for additional equipment intervention.

[0079] Level 2 Risk (Status Requiring Monitoring): Risk index between 0.2 and 0.4, predicted core body temperature between 36.0℃ and 36.5℃, cumulative blood loss between 200ml and 500ml, infusion rate per unit weight between 5ml / (kg·h) and 10ml / (kg·h), BIS value between 60 and 80, and surgical duration between 60 and 120 minutes. Monitoring frequency is every 10 minutes. Interventions include activating a medical heating blanket and setting the temperature to 38℃, with a heating area of ​​0.8m²; ensuring the infusion temperature is maintained between 25-28℃; and recording core body temperature every 10 minutes.

[0080] Level 3 Risk (Mild Warning State): Risk index between 0.4 and 0.6, predicted core body temperature between 35.5℃ and 36.0℃, cumulative blood loss between 500ml and 1000ml, infusion rate per unit weight between 10ml / (kg·h) and 15ml / (kg·h), BIS value between 40 and 60, and surgical duration between 120 and 240 minutes. Monitoring frequency is every 5 minutes. Interventions include raising the heating blanket temperature to 40℃ and adding a warm air blower, ensuring an airflow speed of 3m / s, a temperature of 38℃, and a distance of 1.5m from the patient; starting the infusion warmer and setting the temperature to 40℃ with a warming rate of 5℃ / min; reducing patient limb exposure; and nurses conducting on-site checks every 5 minutes.

[0081] Level 4 Risk (Moderate Warning): Risk index between 0.6 and 0.8, predicted core body temperature between 35.0℃ and 35.5℃, cumulative blood loss between 1000ml and 1500ml, infusion rate per kilogram between 15ml / (kg·h) and 20ml / (kg·h), BIS value between 30 and 40, and surgical duration between 240 and 360 minutes. Monitoring frequency is every 2 minutes. Interventions include activating external warming equipment; reducing the infusion of hypothermic fluids and raising the infusion warmer temperature to 42℃; adjusting the operating room temperature to 28℃, and notifying the anesthesiologist to remotely check data every 2 minutes.

[0082] Level 5 Risk (Severe Warning Status): Risk index not lower than 0.8, predicted core body temperature below 35.0℃, cumulative blood loss exceeding 1500ml, infusion rate exceeding 20ml / (kg·h) based on body weight, BIS value below 30, and operation duration exceeding 360 minutes. Monitoring frequency is once every 1 minute. Intervention measures include simultaneously activating external warming equipment and dual-pathway warming infusion; urgently notifying the anesthesiology and surgery departments to arrive and initiating a multidisciplinary consultation; performing blood gas analysis every 30 minutes; providing full-body warmth, exposing only the surgical incision area not exceeding 0.1m², and recording core body temperature and vital signs every minute.

[0083] Early warning response process: The postoperative monitoring cloud platform is deployed in the hospital's private cloud environment, possessing efficient real-time computing capabilities with a response latency of less than 100ms. This platform can calculate the risk index in real time (…). ), predicting core body temperature ( ) and sample-level predicted core body temperature ( When a patient falls within a specific risk level for three consecutive time steps, the system will automatically trigger a closed-loop early warning response mechanism. The specific process is as follows: 1. Signal Triggering and Push The platform leverages The protocol pushes early warning signals to nurse station displays and healthcare workers' mobile apps in real time. Warning messages are presented in bright red flashing font to ensure healthcare workers can quickly detect them. The warning information is comprehensive, covering key information such as patient ID, current risk level, core body temperature indicators, and recommended interventions, providing clear and concise guidance for healthcare workers.

[0084] 2. Medical equipment linkage The platform sends precise control commands to relevant equipment during and after surgery via RS485 bus or HL7 FHIR protocol, enabling intelligent linkage of medical devices. Level 2 warning response: When a Level 2 warning is triggered, the heating blanket controller will automatically switch the temperature to 38℃, and the response time will be strictly controlled within 2 seconds to ensure that the patient can receive appropriate warming measures in a timely manner.

[0085] Level 3 Early Warning Response: If the Level 3 early warning standard is reached, the infusion warmer will automatically start and set the temperature to 40℃. At the same time, the system will automatically turn off the air conditioning in the operating room to avoid further impact on the patient from the low temperature environment, and comprehensively ensure the stability of the patient's body temperature.

[0086] Level 5 Early Warning Response: In an emergency situation triggering a Level 5 early warning, the system will automatically dial the anesthesiology department's on-call phone to promptly notify relevant experts. Simultaneously, it will rapidly upload the patient's real-time data (approximately 30 minutes) to the anesthesiology department's workstation, providing anesthesiologists with comprehensive and timely patient information to enable them to make swift and accurate decisions.

[0087] 3. Recording and Tracing All warning events are automatically stored in the hospital's Electronic Medical Record (EMR) system, forming a complete "warning-intervention-feedback" closed-loop record. The data storage format strictly adheres to the HL7 FHIR standard, ensuring data standardization, consistency, and shareability, achieving 100% traceability. Furthermore, the system automatically generates a weekly "Hypothermia Warning Statistical Report," which covers important indicators such as the RI distribution range for each risk level, the number of warnings, intervention effectiveness, and equipment failure rate. This detailed data provides a solid basis for optimizing model parameters, helping to continuously improve the accuracy and reliability of the warning system and safeguard patient safety.

[0088] To verify the beneficial effects of the present invention, this embodiment provides the following specific implementation example one: The study included 1200 surgical patients from a tertiary hospital, including high-risk patients: 200 patients with a body mass index (BMI) of less than 18.5, 300 elderly patients aged 65 years or older, and 180 patients who underwent long-term surgery lasting more than 3 hours.

[0089] Based on the above samples, we conducted a series of rigorous controlled experiments. During the experiments, the method proposed in this invention was simultaneously compared with existing methods, focusing on the differences in key performance indicators. At the same time, we closely monitored and meticulously recorded the practical clinical application effects of this invention, comprehensively evaluating its feasibility and practicality from multiple dimensions.

[0090] After systematic and comprehensive experiments and observations, the results are as follows: Figure 5 The figure shows the trend of postoperative hypothermia risk index, with postoperative time (minutes) as the horizontal axis, ranging from 0 to 150 minutes; and the risk index as the horizontal axis. The vertical axis represents the range 0-1. The graph shows five threshold lines: green (RI=0.2), yellow (RI=0.4), orange (RI=0.6), and red (RI=0.8). The blue curve represents a patient undergoing orthopedic surgery. The gray shaded area represents the patient's core body temperature changes.

[0091] Analysis of the patient's condition at different stages after surgery: 0–30 minutes (early postoperative recovery): During this stage, the patient… =0.12, less than 0.2, within the safe range. Predicted core body temperature The temperature was 36.8℃, and the cumulative blood loss was [missing data]. The infusion rate is 150ml, calculated based on body weight. for The patient had no risk characteristics and only required routine warming measures.

[0092] 30-60 minutes: It rose to 0.32, exceeding the safety threshold of 0.2. (Predicted core body temperature) Temperature dropped to 36.3℃, cumulative bleeding volume Increase to 350ml, infusion rate Upgraded to The platform automatically activated the heating blanket and set the temperature to 38℃. After 10 minutes, the patient's core body temperature rose to 36.4℃.

[0093] 60-90 minutes: It further rose to 0.55, reaching the level three warning standard. (Predicted core body temperature) Temperature dropped to 35.8℃, cumulative bleeding volume The infusion rate is 650ml. for At this point, a 40°C infusion warmer and a heater were activated for intervention. After 5 minutes, the patient's core body temperature stopped decreasing.

[0094] 90-120 minutes: The temperature dropped to 0.28, indicating a lower risk level. (Predicted core body temperature) Temperature rose to 36.2℃, cumulative bleeding volume The infusion rate was stabilized at 650ml, and the infusion rate was reduced. The heating blanket is kept at 38°C.

[0095] 120-150 minutes: The temperature dropped to 0.15, which is less than 0.2, indicating a safe level. (Predicted core body temperature) The temperature was stabilized at 36.6℃, all heating equipment was turned off, and only routine insulation was performed.

[0096] Results Analysis Experimental results show that the constructed model can accurately capture The model can detect dynamic changes in trends. It can issue an early warning signal 15 minutes before a risk occurs, buying valuable intervention time for medical staff. For example, at 60 minutes... The temperature reached the level 2 warning threshold of 0.4, but the patient's core body temperature was still not below 36°C. Based on the model's warning, medical staff promptly implemented intervention measures. The rapid decline effectively prevented further deterioration of the risk. This fully validates the effectiveness of the synergistic effect of early warning and quantitative intervention. Furthermore, compared with traditional methods, the time for patients' body temperature to rise was shortened by 22.3% after adopting this model and corresponding intervention measures, further highlighting the significant advantages and important value of this invention in clinical applications.

[0097] To verify the beneficial effects of the present invention, this embodiment provides the following specific implementation example two: A large sample size of 12,000 groups was selected to form the test set. All indicators are presented as mean ± standard deviation to ensure the scientific validity and reliability of the data. The "risk index-related indicators" involved in the method of this invention are all calculated closely around the Risk Index (RI). Through multi-dimensional data analysis and evaluation, the performance of this invention in risk prediction and assessment is comprehensively verified. The results are shown in Table 1 below. Figure 6 As shown: Table 1 Performance Indicators of the Invention and Existing Methods According to the data in Table 1: In terms of accuracy, the method of this invention demonstrates a significant advantage, achieving an accuracy of 96.8%, which is 18.5 percentage points higher than the best existing method's 78.3%. This improvement is attributed to the fusion of multimodal data and the quantification of risk through the RI (Risk Index), effectively avoiding the information blind spots caused by a single parameter.

[0098] In terms of early warning time, this invention also performs exceptionally well, with an average early warning time of 22.5 minutes, which is 14.7 minutes longer than the 7.8 minutes of the traditional Logistic model. This is thanks to the advanced predictive ability of the RI (Regression Analysis), which provides ample time for clinical intervention.

[0099] Regarding the false alarm rate, this invention has a significant advantage, with a false alarm rate of only 3.3%, a reduction of 24.7 percentage points compared to 28.0% for core body temperature monitoring alone. This is attributed to the optimization of the RI calculation logic through dynamic attention fusion and clinical constraints, which effectively reduces the occurrence of false alarms.

[0100] This invention also performs excellently in terms of RI correlation, with an R² value of 0.92 for RI versus actual risk, which is significantly higher than traditional methods. The model's RI is 0.65. This indicates that RI can more accurately quantify risk, avoiding the problem in traditional models where "static parameters cause RI to become disconnected from actual risk."

[0101] This invention also has advantages in terms of body temperature prediction error, with an error of only 0.08℃, compared to traditional methods. The model's temperature is 0.22℃ lower than 0.14℃. High precision. (Predicting core body temperature) ensures the accuracy of RI calculation and avoids RI deviation and misjudgment of risk level due to prediction errors.

[0102] See Figure 6 The diagram shows a five-level risk classification confusion matrix. The horizontal axis represents the predicted risk level, and the vertical axis represents the actual risk level. The values ​​in the matrix represent the number of samples in the test set, which contains a total of 12,000 samples. The values ​​on the diagonal represent the number of correctly classified samples, while the values ​​off-diagonal represent the number of misclassified samples. The color intensity of the cells is positively correlated with the number of samples. In the matrix, the "predicted level" corresponding to each cell is determined based on the RI (Risk Index), while the "actual level" is determined based on... The level determined by (predicted core body temperature).

[0103] The analysis of the forecasts for each risk level is as follows: Level 1 (Safe): The correct identification rate was 90% (180 / 200), meaning that the majority of samples were accurately identified as "safe". Only 10% (20 cases) of the samples were misclassified, and these misclassified samples were concentrated in adjacent Levels 2 and 3, with no instances of skipping levels and being misclassified as Level 4 or 5. This indicates that the model has high stability in identifying "safe" levels.

[0104] Level 2 (Attention): The correct identification rate reached 85% (170 / 200), which is relatively high. False positives were mainly concentrated in Level 1 (10 cases) and Level 3 (15 cases), with only 5 cases misclassified as Level 4. There were no instances of jumping to Level 5. This indicates that the model experiences some confusion when distinguishing between the "Attention" level and the "Safe / Mild Warning" level, but the overall situation is within a controllable range.

[0105] Level 3 (Mild Warning): The correct recognition rate was 80% (160 / 200), a slight decrease. False positives were mainly concentrated in Level 2 (20 cases) and Level 4 (15 cases), with no instances of skipping to Level 1 or Level 5. This reflects the high similarity in features between the "Mild Warning" and "Attention / Moderate Warning" levels, making it difficult for the model to distinguish between these two levels.

[0106] Level 4 (Moderate Warning): The correct recognition rate was 80% (160 / 200), which is close to the performance of Level 3. The misclassified samples were mainly Level 3 (25 cases) and Level 5 (10 cases), with no cases of skipping to Level 1 or Level 2. This indicates that the model needs further optimization in distinguishing the boundary between the "moderate warning" level and the "mild / severe warning" levels.

[0107] Level 5 (Severe Warning): The correct recognition rate was high, at 87.5% (175 / 200), with only 12.5% ​​(25 cases) of samples being misclassified. These misclassified samples were concentrated in the adjacent Level 4 (20 cases), with only 5 samples misclassified as Level 3, and no cases of skipping to a lower level. This indicates that the model has high reliability in recognizing the "severe warning" level.

[0108] Summary of misjudgment characteristics: All misjudged samples were concentrated in adjacent levels, totaling 396 groups, accounting for 3.3% of the total samples. No cross-level misjudgments occurred. This characteristic is consistent with the clinical risk pattern of "gradual rather than abrupt" changes, indicating that the model will not lead to serious decision-making errors due to RI misjudgments, and has a certain degree of reliability and safety in clinical applications.

[0109] To verify the beneficial effects of the present invention, this embodiment provides the following specific implementation example three: A three-month clinical pilot application was conducted in the general surgery department of a tertiary hospital. 400 surgical patients were randomly divided into a routine nursing group, a risk warning group, and a comprehensive intervention group using this method. The incidence of hypothermia and postoperative complications were compared between the two groups. The results are shown in Table 2 below. Table 2 Performance Indicators of the Invention and Existing Methods According to the data in Table 2: Regarding the incidence of hypothermia: the incidence of hypothermia in the experimental group was 12.5%, a decrease of 67.5% compared to 38.5% in the control group. This result fully validates the core value of this invention based on the "RI advance warning + quantitative intervention" model. This model uses RI to identify risks in advance and initiates intervention measures before the patient's body temperature drops below 36°C, effectively preventing the occurrence of hypothermia.

[0110] Regarding the incidence of complications: the incidence of postoperative chills and infection was reduced by 75%. Given that hypothermia can suppress immune and coagulation functions, this invention maintains stable patient body temperature through respiration (RI), thereby reducing the inducing factors of complications at their source.

[0111] In terms of improved medical efficiency: anesthesia recovery time was reduced by 25.6%, and ICU stay was reduced by 46.7%, significantly reducing the occupation of medical resources. This is thanks to precise intervention under the guidance of the RI, which avoided resuscitation delays caused by hypothermia.

[0112] In terms of patient experience optimization: patient satisfaction increased by 15.3%. Because RI can provide accurate early warnings, it reduces the occurrence of discomfort symptoms such as chills and infection, and the intervention measures are more targeted, thereby improving the patient's medical experience.

[0113] Meanwhile, an additional questionnaire survey of medical staff revealed that 74 nurses (92.5% of the total surveyed, out of 80) believed that "RI-based early warning reduces the workload of manual monitoring"; and 70 anesthesiologists (87.5% of the total surveyed, out of 80) believed that "RI's quantitative risk assessment provides a stronger basis for intervention decisions." These results validate the practicality and acceptance of RI in clinical practice.

[0114] Example 2 like Figure 7 As shown, this embodiment provides a postoperative hypothermia early warning system, which includes: The acquisition module is used to acquire multi-source patient data and preprocess the multi-source data, which includes intraoperative core body temperature data, body surface temperature distribution data, intraoperative dynamic parameter data, and patient static baseline data. The feature extraction and fusion module is used to extract features from the multi-source data. It adopts a cross-modal attention mechanism to calculate the weights of each extracted feature vector in real time for adaptive feature fusion, performs constraint optimization on the fused features, and outputs a high-dimensional fused feature vector. The calculation module is used to construct a hybrid time-series prediction model based on the high-dimensional fusion feature vector, so as to output the predicted core body temperature value and low body temperature risk index (RI) of the patient for a certain period of time in the future. Risk warning module: Based on the core body temperature prediction value and the hypothermia risk index RI, risk standards and quantitative intervention strategies are formulated to achieve postoperative hypothermia warning for patients.

[0115] A postoperative hypothermia early warning system in this application embodiment can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), etc. This application embodiment does not impose specific limitations.

[0116] One embodiment of the postoperative hypothermia early warning system in this application can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application does not specifically limit the specific operating system used.

[0117] The postoperative hypothermia early warning system provided in this application embodiment can achieve... Figure 1 The various processes involved in implementing a postoperative hypothermia early warning method are described in the following embodiments. To avoid repetition, they will not be repeated here.

[0118] The postoperative hypothermia early warning method and system according to the present invention have the following advantages compared with the prior art: (1) By using a multimodal static and dynamic data fusion strategy, the patient's static baseline data, body surface temperature data, and intraoperative dynamic data are correlated and fused in the spatiotemporal dimensions. This method effectively solves the "information silo" problem that exists in traditional methods that only measure body surface temperature data, expands the monitoring dimensions, and thus improves the accuracy of postoperative risk assessment.

[0119] (2) A feature fusion module based on the attention mechanism was designed. This module can calculate the weight of each parameter in real time during the dynamic changes of intraoperative parameters, avoiding the error caused by fixed weights. At the same time, combined with the LSTM-GRU hybrid network to capture the temporal trend of body temperature, the core body temperature of the next 15 minutes can be predicted with an error of less than 0.1℃ and an advance warning time of 15-30 minutes, which is 10-15 minutes earlier than the traditional method, providing a sufficient time window for clinical intervention.

[0120] (3) A hypothermia risk index (RI) integrating "predicted body temperature + dynamic parameters" was constructed, elevating the risk assessment from "qualitative" to "0-1 continuous quantification," and a five-level risk standard was formulated in conjunction with clinical scenarios. The method reduced the false negative rate to below 2.1% and the false positive rate to below 3.3%, significantly lower than the false negative rate of 29.5% and the false positive rate of 22.4% for manual monitoring, and the false negative rate of 18.7% and the false positive rate of 28% for single-modality monitoring. In addition, each risk level corresponds to a clear quantitative intervention measure, which can be directly applied to clinical practice. Optionally, embodiments of this application also provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of a postoperative hypothermia early warning method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0121] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of a postoperative hypothermia early warning method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0122] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0123] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention.

[0124] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0125] Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The reference to "embodiment" herein means that a specific feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily indicate the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0126] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for early warning of postoperative hypothermia, characterized in that, include: S1: Acquire multi-source patient data and preprocess the multi-source data, which includes intraoperative core body temperature data, body surface temperature distribution data, intraoperative dynamic parameter data, and patient static baseline data. S2: Extract features from the multi-source data, employ a cross-modal attention mechanism, calculate the weights of each extracted feature vector in real time to perform adaptive feature fusion, perform constraint optimization on the fused features, and output a high-dimensional fused feature vector. S3: Construct a hybrid time-series prediction model based on the high-dimensional fusion feature vector to output the predicted core body temperature and low body temperature risk index (RI) of the patient for a certain period in the future; S4: Based on the core body temperature prediction value and the hypothermia risk index RI, formulate risk standards and quantitative intervention strategies to achieve early warning of hypothermia after surgery.

2. The postoperative hypothermia early warning method according to claim 1, characterized in that, In S2, feature extraction of the multi-source data specifically includes: S211: Principal component analysis is used to reduce the dimensionality of multiple features in the patient's static baseline data. The number of principal components is determined through a graph to output a three-dimensional static feature vector. ∈ ; S212: A one-dimensional convolutional neural network is used to extract local trend features from the time-series variables of multiple dynamic parameters in the intraoperative core body temperature data, and outputs a 64-dimensional dynamic local feature vector. ∈ ; S213: A Transformer encoder is used to capture long-term dependencies in the temporal data of intraoperative dynamic parameter data. The association weights of data at different time steps are calculated through a self-attention mechanism, and a 128-dimensional long-term feature vector is output. ∈ ; S214: The U-Net semantic segmentation network is used to extract temperature distribution features from infrared thermal images in body surface temperature distribution data. The enhanced thermal image is input, and the encoder extracts temperature gradient features. The decoder generates pixel-level temperature anomaly region masks and finally outputs three core features: mean body surface temperature, standard deviation of body surface temperature, and percentage of high-temperature area, which constitute a three-dimensional body surface feature vector. ∈ .

3. The postoperative hypothermia early warning method according to claim 2, characterized in that, S2 also includes: S221: Define the attention weight matrix , in For static feature weights, For dynamic local feature weights, For long-term dependent feature weights, Weights for body surface features; S222: Concatenate the 3D static feature vector, dynamic local feature vector, long-term feature vector, and 3D surface feature vector according to their respective weights to output a 256-dimensional fused feature vector. ∈ The formula is: Where ⊕ represents the vector concatenation operation.

4. The postoperative hypothermia early warning method according to claim 1, characterized in that, In S2, the constraint optimization of the fusion features specifically includes body temperature coupling constraints, bleeding volume constraints, and anesthesia depth constraints, which are used to calculate the total constraint loss.

5. The postoperative hypothermia early warning method according to claim 1, characterized in that, The total loss of the loss function in the hybrid temporal prediction model Represented as: in, , , All are weighting coefficients. To constrain losses; The loss in body temperature prediction, denoted as RMSE (Reflection on the Deviation Between Predicted and Actual Body Temperature Values), is expressed as: For the predicted core body temperature of the i-th sample, Let N be the actual core body temperature of the i-th sample, and N be the number of samples in the test set. Binary cross-entropy loss (BCE): in, =1 indicates hypothermia. =0 indicates normal body temperature, and RI is the risk index for low body temperature.

6. The postoperative hypothermia early warning method according to claim 1, characterized in that, In S3, the hypothermia risk index RI is calculated in two steps: weighted summation and sigmoid normalization. Specifically: Calculate the unnormalized risk value : in, Let be the normalized predicted core body temperature value for the i-th sample. For the first Normalized cumulative bleeding volume for each sample. For the first Normalized infusion rate for each sample. For the first Normalized BIS values ​​for each sample; Sigmoid normalization: Will Mapping to the [0,1] interval yields the low body temperature risk index for the i-th sample. The formula is: Where the steepness coefficient k=8, and the threshold , That is, the quantitative risk of postoperative hypothermia occurring in the i-th sample within the next 15 minutes. The closer it is to 1, the higher the risk.

7. A postoperative hypothermia early warning system, characterized in that, For implementing the postoperative hypothermia early warning method as described in any one of claims 1 to 6, it includes: The acquisition module is used to acquire multi-source patient data and preprocess the multi-source data, which includes intraoperative core body temperature data, body surface temperature distribution data, intraoperative dynamic parameter data, and patient static baseline data. The feature extraction and fusion module is used to extract features from the multi-source data. It adopts a cross-modal attention mechanism to calculate the weights of each extracted feature vector in real time for adaptive feature fusion, performs constraint optimization on the fused features, and outputs a high-dimensional fused feature vector. The calculation module is used to construct a hybrid time-series prediction model based on the high-dimensional fusion feature vector, so as to output the predicted core body temperature value and low body temperature risk index (RI) of the patient for a certain period of time in the future. Risk warning module: Based on the core body temperature prediction value and the hypothermia risk index RI, risk standards and quantitative intervention strategies are formulated to achieve postoperative hypothermia warning for patients.

8. A computer device, characterized in that, The device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the postoperative hypothermia early warning method as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, The computer storage medium stores instructions that, when executed on the computer, cause the computer to perform the postoperative hypothermia early warning method as described in any one of claims 1 to 6.

Images