Anesthesia complication prediction model construction method based on deep learning

Abstract

The invention relates to the technical field of medical systems, and particularly discloses an anesthesia complication prediction model construction method based on deep learning, and the method comprises the following steps: S1, obtaining multi-source heterogeneous anesthesia medical data; s2, constructing a multi-modal feature fusion module; s3, designing a hierarchical deep neural network architecture which comprises sub-networks for processing different modal data in parallel and a full-connection prediction layer fusing multi-modal features; s4, continuously outputting a complication probability curve in a sliding time window mode by adopting a dynamic risk trajectory prediction mechanism instead of a single static prediction result; and S5, deploying a clinical real-time decision interface, and mapping a prediction result to an anesthesia monitoring equipment alarm system in real time. A bidirectional LSTM + 1D-CNN hybrid encoder and a cross-modal attention mechanism are adopted, time sequence dependence of physiological signals and spatio-temporal characteristics of operation events are synchronously captured, deep semantic fusion of multi-source data is achieved, and the characterization capacity of a model for precursor characteristics of complications is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical systems, and specifically to a method for constructing an anesthesia complication prediction model based on deep learning. Background Art

[0002] In clinical anesthesia practice, real-time prediction of complications is crucial to improving surgical safety. Existing technologies have the following shortcomings: Limited data processing capabilities: Traditional models struggle to effectively integrate heterogeneous data from multiple sources (such as physiological waveforms, operation records, and textual assessments). In particular, they fail to capture long-term dependencies in time series data, resulting in incomplete feature representation. Insufficient real-time prediction: Static prediction models cannot dynamically adapt to changes in patient risk and lack an adaptive adjustment mechanism based on risk level, making it difficult to provide frequently updated early warning information in high-risk situations. Weak model interpretability: Deep learning models often operate as "black boxes," lacking the ability to clinically interpret prediction results, making it difficult to gain the trust of medical staff and unable to locate key physiological parameters or operational events that trigger risks. Data imbalance: The scarcity of rare complication samples leads to bias in model training, and traditional data augmentation methods struggle to maintain the physiological authenticity of medical data. Furthermore, cross-hospital data sharing is limited by privacy protection, resulting in insufficient model generalization capabilities. Low edge deployment efficiency: Complex neural networks have high computational loads on edge monitoring devices, resulting in large real-time inference delays, which cannot meet the low-latency requirements in anesthesia scenarios. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides a method for constructing an anesthesia complications prediction model based on deep learning, which solves the problems in the above-mentioned background technology.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for constructing an anesthesia complication prediction model based on deep learning, comprising the following steps: S1: Acquire multi-source heterogeneous anesthesia medical data, including at least patient physiological time series data, anesthesia operation records, drug use records, and preoperative evaluation texts; S2: Construct a multimodal feature fusion module, convert discrete operation events into continuous vectors through the embedding layer, extract local features of physiological signals through the temporal convolution layer, and weight key time segments through the attention mechanism; S3: Design a hierarchical deep neural network architecture, including sub-networks that process data from different modalities in parallel, and a fully connected prediction layer that integrates multimodal features; S4: A dynamic risk trajectory prediction mechanism is used to continuously output complication probability curves in a sliding time window manner, rather than a single static prediction result; S5: Deploy a clinical real-time decision-making interface to map the prediction results to the anesthesia monitoring equipment alarm system in real time.

[0005] Preferably, the multimodal feature fusion module in step S2 specifically includes: A bidirectional LSTM+1D-CNN hybrid encoder is used for physiological time series data. The CNN layer extracts local waveform features, and the LSTM layer captures long-term dependencies. A time-aware embedding layer is used for anesthetic operation events to jointly encode the operation type and the occurrence time into a spatiotemporal vector; The correlation weights between physiological features and operational events are calculated through a cross-modal attention mechanism to achieve feature adaptive fusion. The cross-modal attention mechanism satisfies: ; in, Physiological characteristics Operation events The dependency weight of For the Physiological feature vectors; For the Operation event vector; is a trainable feature concatenation matrix; is a trainable weight vector; To calculate the eigenvector and Scoring function for relevance; is an exponential function; is the hyperbolic tangent activation function; is the eigenvector and splicing operation.

[0006] Preferably, the processing of physiological time series data specifically includes: The input data at least includes electrocardiogram, blood oxygen saturation, invasive blood pressure, and end-tidal carbon dioxide waveform; A differentiable wavelet denoising module is added before the LSTM layer, and the optimal wavelet basis function is adaptively selected through training. The loss function of the differentiable wavelet denoising module is: ; in, is the loss function of the wavelet denoising module; The original physiological signal at time The value of For the A learnable wavelet basis function; For the The wavelet coefficients corresponding to the wavelet basis functions; is the number of wavelet basis functions; is the regularization parameter; is a vector consisting of wavelet coefficients, ; is the L2 norm, which is used to measure the signal reconstruction error; is the L1 norm, which is used to constrain the sparsity of wavelet coefficients; For all time points sum; For all wavelet basis functions sum; A multi-scale feature pyramid structure is used to extract features at time scales of 1 second, 5 seconds, and 30 seconds in parallel. The output satisfies the following requirements: ; in, is the output feature vector of the multi-scale feature pyramid; It is a 1×1 convolution operation for feature fusion and dimensionality reduction; is the feature vector extracted from the 1 second time scale; is the feature vector extracted from the 5-second time scale; is the feature vector extracted from the 30-second time scale; It is a channel splicing operation that splices features of different scales in the channel dimension; It is a linear interpolation upsampling operation that unifies the feature resolution to 1 second scale.

[0007] Preferably, the hierarchical network architecture in step S3 includes: Event Reasoning Sub-Network: Uses Transformer decoder to model the causal relationship of anesthesia operation sequences and predict potential operation risks; Physiological state sub-network: Use graph convolutional networks to model the biological correlations between multiple physiological parameters; Cross-modal fusion layer: Design a gated cross network to control the direction of information flow through a learnable gating mechanism. The gating mechanism of the gated cross network is defined as: ; in, is the gating vector, used to control the direction of information flow; for Activation function, mapping vector elements to the interval [0,1]; is a trainable gating weight matrix; is a trainable gate bias vector; It is the concatenation vector of the output of the event reasoning sub-network and the output of the physiological state sub-network; is the output feature vector of the event reasoning sub-network; is the output feature vector of the physiological state sub-network; is the feature vector after cross-modal fusion; is the element-wise multiplication operator; It is the complement operation of the gated vector, which negates each element.

[0008] Preferably, the dynamic risk trajectory prediction in step S4 is specifically implemented as follows: Define a risk-sensitive sliding window, the window length is dynamically adjusted according to the patient's current risk level, and a short window is used for high-frequency updates under high-risk conditions. The risk-sensitive sliding window length Dynamically adjusted to: ; in, is the length of the sliding window after dynamic adjustment; is the minimum length of the sliding window, which is used in high-risk situations; is the maximum length of the sliding window, which is used in low-risk situations; For the current moment Predicted probability of complication risk; is a high risk threshold, when When this value is exceeded, it is considered a high-risk state; is a low risk threshold, when When the value is lower than this, it is judged as a low-risk state; a dual-branch prediction head is designed in the output layer, the main branch predicts the complication probability curve, and the auxiliary branch predicts the risk change trend derivative; an uncertainty quantification module is introduced, and the Monte Carlo Generate confidence intervals for prediction results. The uncertainty quantification module outputs confidence intervals: ;in, is the mean of the Monte Carlo Dropout sampling results, representing the expectation of the predicted probability; is the standard deviation of the Monte Carlo Dropout sampling results, representing the uncertainty of the prediction; is the sampling number of Monte Carlo Dropout; For the The prediction results of Monte Carlo Dropout sampling; For The sub-sampling results are summed.

[0009] Preferably, the data preprocessing in step S1 further includes: Constructing a knowledge graph in the field of anesthesia to link medical terms in preoperative text assessments to standardized entities; Adopting generative adversarial networks to synthesize rare complication samples and solve the problem of data imbalance; Design a timing alignment interpolation module and establish a dynamic time warping model for device data with different sampling frequencies.

[0010] Preferably, the training strategy of the model includes: A multi-task joint learning framework is used, with the main task being to predict complications, and auxiliary tasks including anesthesia depth classification and blood flow stability assessment. Designing a clinical interpretability constraint loss function to force high-risk prediction results to be associated with specific physiological patterns; Using the federated learning mechanism, model parameters can be updated across hospitals without sharing the original data.

[0011] Preferably, the clinical decision interface of step S5 is implemented as follows: Develop a risk visualization engine to overlay prediction curves on the anesthesia monitor timeline; Establish a hierarchical alarm trigger logic: when the predicted probability exceeds the first threshold, an early warning is issued; when it exceeds the second threshold, the infusion pump is decelerated. The hierarchical alarm trigger conditions are: ; For the moment Predicted probability of complications; It is the integral of the predicted probability in the past 30 seconds and is used to measure the cumulative risk; The threshold for triggering warnings. When the cumulative risk exceeds this value, a yellow warning is triggered. It is the maximum value of the predicted probability in the last 10 seconds and is used to capture sudden high risks; The device linkage trigger threshold is set when the probability peak exceeds this value and the risk change rate meets the conditions. is the time derivative of the predicted probability, which is used to measure the rate of risk change; is the risk change rate threshold. When the absolute value of the derivative exceeds this value, it is determined to be a sudden risk change. It is an absolute value operator, used to calculate the magnitude of the risk change rate; The reverse interpretation module is integrated to highlight the physiological parameters and operation events with the highest contribution when an alarm is triggered. The contribution of the reverse interpretation module is calculated as: ; in, For the An indicator of the contribution of an input feature to the predicted probability; is the predicted probability For the first Input feature vector The gradient of , which reflects the influence of feature changes on the prediction results; For the input feature vectors; is an element-wise multiplication operator used to convert the gradient With the original feature vector multiply to highlight the impact of key features; is the L2 norm, used to quantify the feature contribution the overall size.

[0012] Preferably, the complications include at least: malignant hyperthermia, anaphylactic shock, difficult airway, cardiac arrest after spinal anesthesia, and local anesthetic poisoning, and the model sets independent prediction channels for each type of complication, and each channel shares a feature extraction layer.

[0013] The present invention provides a method for constructing an anesthesia complication prediction model based on deep learning, which has the following beneficial effects: This invention achieves the following breakthroughs through systematic technological innovation: 1. Multimodal feature fusion architecture: This architecture uses a bidirectional LSTM+1D-CNN hybrid encoder and a cross-modal attention mechanism to simultaneously capture the temporal dependencies of physiological signals and the spatiotemporal characteristics of operational events, enabling deep semantic fusion of multi-source data and improving the model's ability to characterize precursor features of complications.

[0014] 2. Dynamic risk trajectory prediction: The risk-sensitive sliding window adaptively adjusts the window length based on the real-time risk probability, and automatically switches to a short-window, high-frequency update mode in high-risk conditions to ensure early identification of sudden complications. The dual-branch prediction head simultaneously outputs the probability curve and the risk change rate, providing clinicians with a more comprehensive assessment of risk evolution trends.

[0015] 3. Interpretability-enhanced design: The reverse interpretation module locates high-contribution physiological parameters and operational events through a gradient attribution mechanism. Combined with a clinical interpretability constraint loss function, it forcibly associates model prediction results with specific pathophysiological patterns, thereby improving decision-making transparency.

[0016] 4. Data optimization and privacy protection: The knowledge graph in the field of anesthesia standardizes medical terminology, and GAN technology synthesizes rare complication samples to balance data distribution. The federated learning framework supports cross-hospital model updates without sharing original data, balancing model generalization capabilities and data security.

[0017] 5. Efficient edge computing deployment: Model distillation technology and dynamic computation offloading strategies reduce the computational load on edge devices. Privacy protection mechanisms ensure local processing of sensitive data, enabling low-latency output of prediction results and device linkage control.

[0018] 6. Clinical decision-making system integration: The risk visualization engine integrates prediction results with the monitor timeline for display, and the hierarchical alarm logic links infusion pumps and other equipment to perform intelligent intervention, building a closed-loop anesthesia complication warning and treatment system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the method for constructing an anesthesia complications prediction model based on deep learning described in the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, the present invention provides a technical solution: a method for constructing an anesthesia complication prediction model based on deep learning, comprising the following steps: S1: Acquire multi-source heterogeneous anesthesia medical data, including at least patient physiological time series data, anesthesia operation records, drug use records, and preoperative evaluation texts; S2: Construct a multimodal feature fusion module, convert discrete operation events into continuous vectors through the embedding layer, extract local features of physiological signals through the temporal convolution layer, and weight key time segments through the attention mechanism; S3: Design a hierarchical deep neural network architecture, including sub-networks that process data from different modalities in parallel, and a fully connected prediction layer that integrates multimodal features; S4: A dynamic risk trajectory prediction mechanism is used to continuously output complication probability curves in a sliding time window manner, rather than a single static prediction result; S5: Deploy a clinical real-time decision-making interface to map the prediction results to the anesthesia monitoring equipment alarm system in real time.

[0022] More specifically, the multimodal feature fusion module in step S2 specifically includes: A bidirectional LSTM+1D-CNN hybrid encoder is used for physiological time series data. The CNN layer extracts local waveform features, and the LSTM layer captures long-term dependencies. A time-aware embedding layer is used for anesthetic operation events to jointly encode the operation type and the occurrence time into a spatiotemporal vector; The cross-modal attention mechanism is used to calculate the correlation weights between physiological features and operational events to achieve adaptive feature fusion. The cross-modal attention mechanism satisfies: ; in, Physiological characteristics Operation events The dependency weight of For the Physiological feature vectors; For the Operation event vector; is a trainable feature concatenation matrix; is a trainable weight vector; To calculate the eigenvector and Scoring function for relevance; is an exponential function; is the hyperbolic tangent activation function; is the eigenvector and splicing operation.

[0023] The physiological time series data encoding processing of the multimodal feature fusion module includes: inputting waveform data such as electrocardiogram (ECG) and blood oxygen saturation (SpO2) into the bidirectional LSTM+1D-CNN hybrid encoder in time series, performing convolution operations through the 1D-CNN layer with a window length of 512, and the convolution kernel sizes are 16, 32, and 64, respectively, with a step size of 8, to extract local waveform features; inputting the feature sequence output by the CNN layer into the bidirectional LSTM layer with 256 hidden layer neurons, capturing long-term dependencies through forward and backward LSTM units, and outputting the physiological feature sequence. .

[0024] The anesthesia operation event encoding process of the multimodal feature fusion module includes: for anesthesia operation events, the operation type is mapped into a category vector through an embedding layer of dimension 128 The operation time is converted into a timestamp code through sine-cosine position coding ; Concatenate the category vector and the timestamp code and input them into the fully connected layer to generate a spatiotemporal vector of dimension 256 , realizing the joint encoding of operation type and time information.

[0025] Input data and preprocessing: The input includes waveform data such as electrocardiogram (ECG) and blood oxygen saturation (SpO2), which are organized into raw signal sequences in time series.

[0026] 1D-CNN layer: A convolutional layer with a window length of 512 is used, with convolution kernel sizes of 16, 32, and 64, respectively, with a step size of 8. Local waveform features (such as the QRS complex of the ECG and the inflection point of the SpO2 waveform) are extracted through a sliding window.

[0027] Bidirectional LSTM layer: The feature sequence output by the CNN is input into a bidirectional LSTM layer with 256 hidden layer neurons. The forward and backward neurons simultaneously capture the long-term dependencies of the signal (such as the trend of blood pressure changes and the periodicity of respiratory rate).

[0028] Output: Generate physiological feature sequence , each eigenvector It integrates local waveform and timing dependency information.

[0029] Time-aware embedding layer: discrete operation types (such as "endotracheal intubation" and "drug administration") are mapped into category vectors through a 128-dimensional embedding layer , capturing the semantic features of the operation. Using sine-cosine position encoding, the operation time is converted into a timestamp code , preserving the time sequence information. After splicing, it is input into the fully connected layer and a 256-dimensional space-time vector is generated through linear transformation. , realizing the joint representation of “operation type-time”.

[0030] Correlation weight calculation, scoring function score(a,b); input physiological feature vector Operation event vector ; Through the splicing matrix Concatenate the two vectors into , after the hyperbolic tangent activation function After processing, with the weight vector Multiply by the transpose of to get the correlation score: ; Score the score of all operation events Normalize and get the physiological feature i for the operation event Dependency weight : ;pass Weighted fusion of physiological features and operational event features strengthens key time segments that have a significant impact on complications (such as the association between drug administration time and physiological signal changes) to achieve semantic-level fusion of multimodal data.

[0031] More specifically, the processing of physiological time series data includes: The input data at least includes electrocardiogram, blood oxygen saturation, invasive blood pressure, and end-tidal carbon dioxide waveform; A differentiable wavelet denoising module is added before the LSTM layer. The optimal wavelet basis function is adaptively selected through training. The loss function of the differentiable wavelet denoising module is: ; in, is the loss function of the wavelet denoising module; The original physiological signal at time The value of For the A learnable wavelet basis function; For the The wavelet coefficients corresponding to the wavelet basis functions; is the number of wavelet basis functions; is the regularization parameter; is a vector consisting of wavelet coefficients, ; is the L2 norm, which is used to measure the signal reconstruction error; is the L1 norm, which is used to constrain the sparsity of wavelet coefficients; For all time points sum; For all wavelet basis functions sum; A multi-scale feature pyramid structure is used to extract features at time scales of 1 second, 5 seconds, and 30 seconds in parallel. The output satisfies the following requirements: ; in, is the output feature vector of the multi-scale feature pyramid; It is a 1×1 convolution operation for feature fusion and dimensionality reduction; The feature vector is extracted from the 1-second time scale (such as a single cardiac cycle of ECG and sudden arrhythmia, using 1D-CNN window length 128 and convolution kernel size 8); is the feature vector extracted from the 5-second time scale (e.g., blood pressure fluctuation amplitude within 5 seconds, respiratory rate change, CNN window length 640, step size 128); is the feature vector extracted from the 30-second time scale (e.g., the downward trend of blood pressure within 30 seconds of anesthesia induction, the slow decrease of SpO2, CNN window length 3840, step size 640); It is a channel splicing operation that splices features of different scales in the channel dimension; It is a linear interpolation upsampling operation that unifies the feature resolution to 1 second scale.

[0032] Differentiable wavelet denoising processing includes: Construct a set of learnable wavelet basis functions initialized to Haar wavelet basis , optimize the basis function by gradient descent; calculate the wavelet coefficients Minimize the L2 norm of the reconstructed signal and the original signal, combined Regularization constrains sparsity, and the loss function is ; Normalize the noise-reduced signal to generate preprocessed physiological time series data.

[0033] Using a differentiable wavelet module, the optimal wavelet basis function is dynamically learned. Compared to traditional fixed wavelet bases (such as Daubechies), this model better captures physiological signal characteristics, effectively preserving key features such as the QRS complex and respiratory waveform while suppressing high-frequency noise (such as electrode interference) and low-frequency drift (such as baseline fluctuation). L1 regularization constraints sparsify the wavelet coefficients, improving the model's robustness to noise and avoiding overfitting to random fluctuations in the physiological signal. 1-second-scale features capture transient physiological anomalies (such as arrhythmias), 5-second-scale features reflect short-term trends (such as blood pressure fluctuations), and 30-second-scale features reveal long-term physiological states (such as changes in anesthesia depth). The integration of these three features comprehensively characterizes the temporal hierarchical characteristics of complication precursor signals. Through a feature pyramid structure, the model simultaneously addresses the "transient anomaly, short-term trend, and long-term state." For example, in the prediction of malignant hyperthermia, it can capture both a sudden increase in EtCO2 (1-second features) and a gradual increase in body temperature (30-second features), improving prediction sensitivity. The wavelet denoising module is trained end-to-end through gradient descent, avoiding the manual parameter adjustment required in traditional wavelet transforms. Multi-scale feature extraction uses parallel processing combined with 1×1 convolution dimensionality reduction to reduce computational complexity while ensuring feature integrity, thus meeting the real-time processing requirements of edge monitoring equipment.

[0034] More specifically, the hierarchical network architecture in step S3 includes: an event reasoning subnetwork, a physiological state subnetwork, and a cross-modal fusion layer; the event reasoning subnetwork uses a Transformer decoder to model the causal relationship of anesthesia operation sequences and predict potential operation risks; the physiological state subnetwork uses a graph convolutional network to model the biological correlation between multiple physiological parameters; the cross-modal fusion layer is used to design a gated cross network, which controls the direction of information flow through a learnable gating mechanism. The gating mechanism of the gated cross network is defined as: ; in, is the gating vector, used to control the direction of information flow; for Activation function, mapping vector elements to the interval [0,1]; is a trainable gating weight matrix; is a trainable gate bias vector; It is the concatenation vector of the output of the event reasoning sub-network and the output of the physiological state sub-network; is the output feature vector of the event reasoning sub-network; is the output feature vector of the physiological state sub-network; is the feature vector after cross-modal fusion; is the element-wise multiplication operator; It is the complement operation of the gated vector, which negates each element.

[0035] Event reasoning features and physiological status characteristics Splice along the channel dimension to form a joint feature vector ; Generate gate vector through fully connected layer : ; in, is the weight matrix, is the bias vector, Activation function, each element of g takes a value in the range [0,1], representing the weight of the event feature.

[0036] The weights of the two types of features are dynamically adjusted through the gating vector to output the fusion features: ; in, is element-wise multiplication, when When it is close to 1, the model is more dependent on the characteristics of the operation event (such as the impact of intubation operation on airway complications); When it is close to 0, more attention is paid to physiological status characteristics (such as the independent risk of sudden drop in blood oxygen).

[0037] The causal modeling advantages of the event-based reasoning subnetwork include its use of a Transformer decoder rather than a traditional RNN. This multi-head attention mechanism concurrently captures long-range operational dependencies, such as the temporal relationship between medication administration during induction, intubation, and hemodynamic fluctuations. This improves the efficiency of operational sequence modeling compared to LSTM. A masking mechanism ensures the physical plausibility of causal relationships, prevents future operational information leakage, and conforms to the temporal order of clinical procedures.

[0038] Graph convolution innovation in the physiological state subnetwork: This constructs a physiological parameter association graph based on a medical knowledge graph, rather than being entirely data-driven. This ensures the model aligns with pathophysiological mechanisms (such as the correlation between ECG and EtCO2 in the respiratory and circulatory systems) and avoids "black box" predictions. Graph convolution operations can capture nonlinear synergistic effects between parameters, such as the predictive weight of the combined features of decreased IBP, increased heart rate, and decreased SpO2 for anaphylactic shock, improving diagnostic accuracy compared to independent parameter analysis.

[0039] Dynamic fusion value of gated cross-network: The gated mechanism can be learned to achieve adaptive adjustment of modal weights, and the gate vector can be optimized through gradient descent. This approach is relevant to complication type, for example, by weighting body temperature (physiological characteristics) in malignant hyperthermia prediction and intubation (event characteristics) in difficult airway prediction. The fusion formula uses linear combination rather than simple concatenation to reduce feature dimensionality expansion (maintaining 256 dimensions) while preserving modality-specific information. Clinical data has validated this fusion approach, demonstrating improved AUC values for complication prediction compared to traditional fully connected fusion.

[0040] More specifically, the dynamic risk trajectory prediction in step S4 is implemented as follows: Define a risk-sensitive sliding window. The window length is dynamically adjusted according to the patient's current risk level. In high-risk situations, a short window with high frequency update is used. The risk-sensitive sliding window length is Dynamically adjusted to: ; in, is the dynamically adjusted sliding window length (number of time steps); is the minimum length of the sliding window, which is used in high-risk situations; is the maximum length of the sliding window, which is used in low-risk situations; For the current moment Predicted probability of complication risk; is a high risk threshold, when When this value is exceeded, it is considered a high-risk state; is a low risk threshold, when When the value is lower than this, it is judged as a low-risk state; a dual-branch prediction head is designed in the output layer, the main branch predicts the complication probability curve, and the auxiliary branch predicts the risk change trend derivative; an uncertainty quantification module is introduced, and the Monte Carlo Generate confidence intervals for prediction results. The uncertainty quantification module outputs confidence intervals: ;in, is the mean of the Monte Carlo Dropout sampling results, representing the expectation of the predicted probability; is the standard deviation of the Monte Carlo Dropout sampling results, representing the uncertainty of the prediction; is the sampling number of Monte Carlo Dropout; For the The prediction results of Monte Carlo Dropout sampling; For The sub-sampling results are summed.

[0041] The risk-sensitive sliding window adjustment of dynamic risk trajectory prediction includes: (threshold such as 0.8), the minimum window length is used Achieve high-frequency updates; when (threshold such as 0.3), use the maximum window length Reduce the computational load; use the linear interpolation formula within the threshold range Dynamically adjust the window length.

[0042] Window length dynamic adjustment mechanism, where is the dynamically adjusted window length (number of time steps), (High Risk Minimum Window), (Low Risk Maximum Window); For the moment The predicted probability of complications, (high risk threshold), (Low risk threshold).

[0043] Threshold trigger logic: When When you switch to , to update the forecast every 10 time steps (such as 10 seconds) to capture sudden risks; when When using , updated every 100 time steps to reduce the computing load of edge devices; when When , the window length is dynamically adjusted according to the linear interpolation formula, for example hour, .

[0044] Each time step corresponds to 1 second of physiological data, and the window length L represents processing of the time series data of the last L seconds; an overlapping sliding window is used, and the step length of each movement is (For example, when L=10, move 1 step each time) to ensure data continuity.

[0045] Monte Carlo Dropout Sampling: When the model is forward propagated, each input sample is Dropout sampling (such as the hidden layer Dropout rate is set to 0.2), each sampling randomly discards some neurons to generate Prediction results ; Calculation formula: ; in, is the mean of the predicted probabilities, Confidence interval generation and clinical application: Generate The confidence interval is , for example when When , the confidence interval is [0.504,0.896]; when When the prediction is made, it prompts "high uncertainty in prediction" and it is recommended to make judgment based on clinical experience.

[0046] More specifically, the data preprocessing of step S1 also includes: constructing a knowledge graph in the field of anesthesia to link medical terms in preoperative text evaluations to standardized entities; using a generative adversarial network to synthesize rare complication samples to solve the problem of data imbalance; and designing a timing alignment interpolation module to establish a dynamic time warping model for device data with different sampling frequencies.

[0047] More specifically, the model's training strategy includes: adopting a multi-task joint learning framework, with the main task being to predict complications, and auxiliary tasks including anesthesia depth classification and blood flow stability judgment; designing a clinically interpretable constrained loss function to force high-risk prediction results to be associated with specific physiological patterns; and using a federated learning mechanism to update model parameters across hospitals without sharing original data.

[0048] More specifically, the clinical decision interface of step S5 is implemented as follows: Develop a risk visualization engine to overlay prediction curves on the anesthesia monitor timeline; Establish a hierarchical alarm trigger logic: when the predicted probability exceeds the first threshold, an early warning is issued; when it exceeds the second threshold, the infusion pump is decelerated. The hierarchical alarm trigger conditions are: ; For the moment Predicted probability of complications; It is the integral of the predicted probability in the past 30 seconds and is used to measure the cumulative risk; The threshold for triggering warnings. When the cumulative risk exceeds this value, a yellow warning is triggered. It is the maximum value of the predicted probability in the last 10 seconds and is used to capture sudden high risks; The device linkage trigger threshold is set when the probability peak exceeds this value and the risk change rate meets the conditions. is the time derivative of the predicted probability, which is used to measure the rate of risk change; is the risk change rate threshold. When the absolute value of the derivative exceeds this value, it is determined to be a sudden risk change. It is an absolute value operator, used to calculate the magnitude of the risk change rate; The integrated reverse interpretation module highlights the physiological parameters and operational events with the highest contribution when an alarm is triggered. The contribution of the reverse interpretation module is calculated as: ; in, For the An indicator of the contribution of an input feature to the predicted probability; is the predicted probability For the first Input feature vector The gradient of , which reflects the influence of feature changes on the prediction results; For the input feature vectors; is an element-wise multiplication operator used to convert the gradient With the original feature vector multiply to highlight the impact of key features; is the L2 norm, used to quantify the feature contribution the overall size.

[0049] The logic execution of graded alarm includes: when the probability integral in the past 30 seconds (Threshold value is 15), a yellow warning is triggered and a risk warning is displayed; when the probability peaks in the last 10 seconds (threshold value such as 0.9) and the risk change rate When the speed reaches 0.05 / s (threshold value, for example, 0.05 / s), a red alarm is triggered and the infusion pump is decelerated by 50%.

[0050] Clinical decision-making adaptation of graded alarms: The cumulative risk score mechanism captures persistent risks (such as persistently high blood pressure due to insufficient anesthesia depth), and the peak + derivative mechanism identifies sudden risks (such as a sudden drop in blood oxygen due to intubation). The combination of the two improves the accuracy of early warnings; the infusion pump linkage adopts a "probability peak + risk growth rate" dual-condition trigger to avoid false triggering caused by a single threshold (the false trigger rate is lower than the traditional single threshold).

[0051] Clinical interpretability of reverse interpretation: Gradient attribution methods locate specific physiological waveform segments (e.g. waveform suddenly increased) and operation events (such as succinylcholine injection), helping doctors understand the basis of prediction and improve the trust of the model; contributing quantitative indicators The key features annotated by clinical experts are highly consistent. For example, in the prediction of malignant hyperthermia, the core feature "body temperature rise rate" is Values are always ranked in the top 2.

[0052] Real-time and reliability optimization: The alarm logic calculation time is less than 15ms (based on the ARM Cortex-A53 processor), meeting clinical real-time requirements; the infusion pump linkage adopts a double confirmation mechanism (model trigger + nurse confirmation) to avoid system malfunction and ensure patient safety.

[0053] More specifically, complications include at least: malignant hyperthermia, anaphylactic shock, difficult airway, cardiac arrest after spinal anesthesia, and local anesthetic poisoning. The model sets up independent prediction channels for each type of complication. Each channel shares the feature extraction layer but has a dedicated output layer.

[0054] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing an anesthesia complication prediction model based on deep learning, characterized in that: The following steps are involved: S1: Acquire multi-source heterogeneous anesthesia medical data, including at least patient physiological time series data, anesthesia operation records, drug use records, and preoperative evaluation texts; S2: Construct a multimodal feature fusion module, convert discrete operation events into continuous vectors through the embedding layer, extract local features of physiological signals through the temporal convolution layer, and weight key time segments through the attention mechanism; S3: Design a hierarchical deep neural network architecture, including sub-networks that process data from different modalities in parallel, and a fully connected prediction layer that integrates multimodal features; S4: A dynamic risk trajectory prediction mechanism is used to continuously output complication probability curves in a sliding time window manner, rather than a single static prediction result; S5: Deploy a clinical real-time decision-making interface to map the prediction results to the anesthesia monitoring equipment alarm system in real time.

2. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 1, characterized in that: The multimodal feature fusion module in step S2 specifically includes: A bidirectional LSTM+1D-CNN hybrid encoder is used for physiological time series data. The CNN layer extracts local waveform features, and the LSTM layer captures long-term dependencies. A time-aware embedding layer is used for anesthetic operation events to jointly encode the operation type and the occurrence time into a spatiotemporal vector; The correlation weights between physiological features and operational events are calculated through a cross-modal attention mechanism to achieve feature adaptive fusion. The cross-modal attention mechanism satisfies: ； in, Physiological characteristics Operation events The dependency weight of For the Physiological feature vectors; For the Operation event vector; is a trainable feature concatenation matrix; is a trainable weight vector; To calculate the eigenvector and Scoring function for relevance; is an exponential function; is the hyperbolic tangent activation function; is the eigenvector and splicing operation.

3. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 2, characterized in that: The processing of physiological time series data specifically includes: The input data at least includes electrocardiogram, blood oxygen saturation, invasive blood pressure, and end-tidal carbon dioxide waveform; A differentiable wavelet denoising module is added before the LSTM layer, and the optimal wavelet basis function is adaptively selected through training. The loss function of the differentiable wavelet denoising module is: ； in, is the loss function of the wavelet denoising module; The original physiological signal at time The value of For the A learnable wavelet basis function; For the The wavelet coefficients corresponding to the wavelet basis functions; is the number of wavelet basis functions; is the regularization parameter; is a vector consisting of wavelet coefficients, ; is the L2 norm, which is used to measure the signal reconstruction error; is the L1 norm, which is used to constrain the sparsity of wavelet coefficients; For all time points sum; For all wavelet basis functions sum; A multi-scale feature pyramid structure is used to extract features at time scales of 1 second, 5 seconds, and 30 seconds in parallel. The output satisfies the following requirements: ； in, is the output feature vector of the multi-scale feature pyramid; It is a 1×1 convolution operation for feature fusion and dimensionality reduction; is the feature vector extracted from the 1 second time scale; is the feature vector extracted from the 5-second time scale; is the feature vector extracted from the 30-second time scale; It is a channel splicing operation that splices features of different scales in the channel dimension; It is a linear interpolation upsampling operation that unifies the feature resolution to 1 second scale.

4. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 3, characterized in that: The hierarchical network architecture in step S3 includes: Event Reasoning Sub-Network: Uses Transformer decoder to model the causal relationship of anesthesia operation sequences and predict potential operation risks; Physiological state sub-network: Use graph convolutional networks to model the biological correlations between multiple physiological parameters; Cross-modal fusion layer: Design a gated cross network to control the direction of information flow through a learnable gating mechanism. The gating mechanism of the gated cross network is defined as: ； in, is the gating vector, used to control the direction of information flow; for Activation function, mapping vector elements to the interval [0,1]; is a trainable gating weight matrix; is a trainable gate bias vector; It is the concatenation vector of the output of the event reasoning sub-network and the output of the physiological state sub-network; is the output feature vector of the event reasoning sub-network; is the output feature vector of the physiological state sub-network; is the feature vector after cross-modal fusion; is the element-wise multiplication operator; It is the complement operation of the gated vector, which negates each element.

5. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 4, characterized in that: The dynamic risk trajectory prediction of step S4 is specifically implemented as follows: Define a risk-sensitive sliding window, the window length is dynamically adjusted according to the patient's current risk level, and a short window is used for high-frequency updates under high-risk conditions. The risk-sensitive sliding window length Dynamically adjusted to: ； in, is the length of the sliding window after dynamic adjustment; is the minimum length of the sliding window, which is used in high-risk situations; is the maximum length of the sliding window, which is used in low-risk situations; For the current moment Predicted probability of complication risk; is a high risk threshold, when When this value is exceeded, it is considered a high-risk state; is a low risk threshold, when When the value is lower than this, it is judged as a low-risk state; a dual-branch prediction head is designed in the output layer, the main branch predicts the complication probability curve, and the auxiliary branch predicts the risk change trend derivative; an uncertainty quantification module is introduced, and the Monte Carlo Generate confidence intervals for prediction results. The uncertainty quantification module outputs confidence intervals: ;in, is the mean of the Monte Carlo Dropout sampling results, representing the expectation of the predicted probability; is the standard deviation of the Monte Carlo Dropout sampling results, representing the uncertainty of the prediction; is the sampling number of Monte Carlo Dropout; For the The prediction results of Monte Carlo Dropout sampling; For The sub-sampling results are summed.

6. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 5, characterized in that: The data preprocessing of step S1 also includes: Constructing a knowledge graph in the field of anesthesia to link medical terms in preoperative text assessments to standardized entities; Adopting generative adversarial networks to synthesize rare complication samples and solve the problem of data imbalance; Design a timing alignment interpolation module and establish a dynamic time warping model for device data with different sampling frequencies.

7. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 6, characterized in that: The model training strategy includes: A multi-task joint learning framework is used, with the main task being to predict complications, and auxiliary tasks including anesthesia depth classification and blood flow stability assessment. Designing a clinical interpretability constraint loss function to force high-risk prediction results to be associated with specific physiological patterns; Using the federated learning mechanism, model parameters can be updated across hospitals without sharing the original data.

8. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 7, characterized in that: The clinical decision interface implementation of step S5 is as follows: Develop a risk visualization engine to overlay prediction curves on the anesthesia monitor timeline; Establish a hierarchical alarm trigger logic: when the predicted probability exceeds the first threshold, an early warning is issued; when it exceeds the second threshold, the infusion pump is decelerated. The hierarchical alarm trigger conditions are: ； For the moment Predicted probability of complications; It is the integral of the predicted probability in the past 30 seconds and is used to measure the cumulative risk; The threshold for triggering warnings. When the cumulative risk exceeds this value, a yellow warning is triggered. It is the maximum value of the predicted probability in the last 10 seconds and is used to capture sudden high risks; The device linkage trigger threshold is set when the probability peak exceeds this value and the risk change rate meets the conditions. is the time derivative of the predicted probability, which is used to measure the rate of risk change; is the risk change rate threshold. When the absolute value of the derivative exceeds this value, it is determined to be a sudden risk change. It is an absolute value operator, used to calculate the magnitude of the risk change rate; The reverse interpretation module is integrated to highlight the physiological parameters and operation events with the highest contribution when an alarm is triggered. The contribution of the reverse interpretation module is calculated as: ； in, For the An indicator of the contribution of an input feature to the predicted probability; is the predicted probability For the first Input feature vector The gradient of , which reflects the influence of feature changes on the prediction results; For the input feature vectors; is an element-wise multiplication operator used to convert the gradient With the original feature vector multiply to highlight the impact of key features; is the L2 norm, used to quantify the feature contribution the overall size.

9. The method for constructing an anesthesia complication prediction model based on deep learning according to claim 8, characterized in that: The complications include at least: malignant hyperthermia, anaphylactic shock, difficult airway, cardiac arrest after spinal anesthesia, and local anesthetic poisoning, and the model sets independent prediction channels for each type of complication, and each channel shares a feature extraction layer.

Images