CN-122024998-A - Anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology

CN122024998ACN 122024998 ACN122024998 ACN 122024998ACN-122024998-A

Abstract

The invention provides an anesthesia depth assessment method based on Mamba architecture and a multi-mode fusion technology. The method comprises the steps of constructing a sliding window based on physiological time lag and enhancing pharmacokinetic characteristics, preprocessing multi-modal data to solve the problems of drug effect lag and characteristic sparseness, constructing a dual-flow characteristic extraction network based on Mamba and a drug effect-physical sign interaction fusion module based on cross attention to capture long-range drug metabolism states and physiological stress reactions respectively, fusing multi-modal physiological information and injecting demographic priors, constructing a physiological consistency constraint loss of fused medical ordinal logic to jointly optimize the model to enhance predicted physiological rationality, and finally introducing a time smoothing strategy and an integral gradient algorithm to output anesthesia depth classification results and an interpretable decision basis to assist clinical accurate administration.

Inventors

WANG JIA
HE YUANYUAN
FAN BINGJUAN
WU JUN

Assignees

杭州市萧山区第一人民医院
浙江大学滨江研究院

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (9)

1. The anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology is characterized by comprising the following steps: collecting multi-mode physiological data, drug infusion data and demographic characteristics of a patient through monitoring equipment, constructing an asymmetric sliding window of the multi-mode physiological data of the patient, and enhancing the pharmacokinetic characteristics of the drug infusion data; Constructing a double-flow feature extraction network based on Mamba state space model according to the data obtained in the first step to independently extract physiological data and drug data features of a patient, introducing a drug effect-physical sign interaction fusion module based on a cross attention mechanism, and carrying out patient fusion state characterization of the physiological features, the drug data features and the demographic features of the patient through the module; thirdly, constructing a physiological consistency constraint loss function fused with the medical ordinal logic, and performing end-to-end training optimization on the model constructed in the second step so as to improve the accuracy of the result and the self-consistency of the medical logic; and step four, inputting time sequence data to be detected into a trained model, outputting three classification probabilities of anesthesia depth, and finishing anesthesia depth assessment.
2. The anesthesia depth assessment method based on Mamba architecture and multi-modal fusion technique as set forth in claim 1, wherein the step one of performing asymmetric sliding window construction on multi-modal physiological data of a patient and performing pharmacokinetic feature enhancement on drug infusion data is performed by the following specific modes The characteristic value of the medicine at the moment is Enhanced differential features The following can be calculated: ; then at the current time As a benchmark, the forward interception length is A slice of historical data for seconds and up-sampling techniques are used to fill vital sign data at a lower sampling rate to the same 1Hz temporal resolution as the drug data.
3. The method for assessing the anesthesia depth based on Mamba architecture and multi-modal fusion technique as set forth in claim 1 or 2, wherein the patient state characterization by the injection of demographic features is obtained in step two by first letting Is that The potential state variable for the time of day, And Is a matrix obtained by discretizing continuous system parameters through a zero-order holding technology, Is a matrix of learnable parameters, and is used for extracting the concentration of the medicine or vital signs after one-dimensional convolution Its potential status updates and outputs the final result The calculation formula of (2) is as follows: and in obtaining potential state sequences of drugs And physiological potential state sequence And then constructing a pesticide effect-physical sign interaction fusion module based on a cross attention mechanism to simulate a causal logic of physical sign change generated by the action of a drug on a human body so as to extract fusion characteristics with causal interpretation.
4. The method for evaluating the anesthesia depth based on Mamba architecture and multi-modal fusion technique of claim 3, wherein the drug effect-physical sign interaction fusion module in the second step simulates causal logic of the drug acting on the human body to generate physical sign changes, and the specific method for extracting the fusion characteristics with causal interpretation comprises the following steps of 、、 As a matrix of parameters that can be learned, As dimension-related scaling factor, the output feature vector after interactive fusion The following can be calculated: Then, let the In order to achieve a vectorized demographic, And Is a scaling factor and a shifting factor derived therefrom; Finally, affine transformation is carried out on the fusion vector obtained in the last step by utilizing two factors so as to assist the model to adaptively adjust the characteristic distribution according to the individual characteristics of the patient and generate personalized patient state representation finally used for evaluation : 。
5. The anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology according to claim 1 or 2, wherein the specific way of constructing a physiological consistency constraint loss function of fusion medical ordinal logic in the third step, and performing end-to-end training optimization on the model to improve the accuracy of the result and the self-consistency of the medical ordinal logic is as follows: constructing a composite objective function consisting of classification loss, regression loss and physiological consistency alignment loss, wherein the total loss function is defined as follows: 。
6. The anesthesia depth assessment method based on Mamba architecture and multi-modal fusion technique, characterized in that the classification Loss in the third step takes focus Loss-based class optimization as a main task, and Focal Loss is introduced as a Loss function of the classification main task To reduce the weight of the easily separable samples and force the model to pay attention to the critical states which are difficult to distinguish, and to receive by using a multi-layer perceptron As input and activated by Softmax function to obtain predictive classification probability result, let For true category Is used to determine the prediction probability of (1), To adjust the parameters, the term loss can be calculated as follows: 。
7. The anesthesia depth assessment method based on Mamba architecture and multi-modal fusion technique, characterized in that the regression loss in the third step uses BSI numerical regression based on mean square error as an auxiliary task, and uses a multi-layer perceptron to receive As input, applying Sigmoid activation function to output to obtain normalized BSI prediction result And scale patient BSI labels to Interval as an optimization target The mean square error loss for the auxiliary BSI regression task may be calculated as follows: 。
8. the method for assessing the depth of anesthesia based on Mamba framework and multimodal fusion technique, wherein the physiological consistency alignment loss in the third step converts the physiological consistency alignment loss based on medical thresholds into a mathematical penalty term, and sets two medical thresholds 0.4、 And according to the prediction result of the classification head Predictive value for regression head Dynamic constraint is carried out: 。
9. The anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology as set forth in claim 1 or 2, wherein the step four is characterized in that the time series data to be tested is input into a trained model, and the specific way of outputting the anesthesia depth three-classification probability is as follows, the model outputs the current time step Is the original classification probability distribution of (1) The system maintains a length of And calculating the smoothed probability by adopting an exponential weighted moving average algorithm, and obtaining an interpretable decision basis of the classification result by using an output model through an integral gradient algorithm. Order the In order to balance the real-time and stability smoothing coefficients, For the probability of the classification result output by the model at the moment, the algorithm flow is as follows: 。

Description

Anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology Technical Field The invention relates to the technical field of digital medical treatment, in particular to an anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology. Background In modern medical systems, accurate assessment of anesthesia depth is a key link in ensuring surgical safety and improving long-term prognosis of patients. Investigation reports show that the number of operations performed annually worldwide is approaching 3.13 billion. The extremely high risks faced by anesthesia management, accompanied by an increasing amount of surgery, are important concerns in the medical field. In particular, too shallow anesthesia may lead to the patient becoming aware of or eliciting severe stress and increase the risk of cardiovascular accidents, while too deep anesthesia is demonstrated to have a strong correlation with postoperative delirium and cognitive dysfunction. Clinical data indicate that in elderly patients undergoing cardiac surgery, cognitive dysfunction can occur at up to 40% within one week after surgery, as well as more than 30% after delirium in hip fracture surgery. The postoperative nervous system complications not only obviously prolong the hospitalization time, but also seriously affect the life quality of patients. Therefore, implementing accurate anesthesia status monitoring under complex pharmacological environments and fragile physiological conditions and avoiding excessive or insufficient anesthetic drugs has become a critical public health problem to be solved urgently in current perioperative medicine. Although the clinical practice at present widely uses the auxiliary monitoring means such as physical sign observation and brain electrical double frequency index (BIS) based on the experience of anesthesiologists, a significant technical bottleneck still exists in realizing accurate and personalized anesthesia evaluation. The traditional physical sign monitoring is deeply restricted by the drug covering effect and the physiological hysteresis, and the significant time delay of the hemodynamic change relative to the neural activity also causes that the traditional physical sign monitoring cannot meet the real-time early warning requirement. Although the brain electrical monitoring device represented by BIS is regarded as a gold standard, the algorithm of the brain electrical monitoring device generally adopts a smooth window of 15 to 30 seconds to maintain stable values, so that serious calculation lag exists in the monitoring result, and sudden change of conscious state is difficult to capture. In addition, the existing anesthesia depth evaluation scheme based on artificial intelligence mostly adopts simple multi-mode feature stitching, lacks explicit modeling of a complex interaction mechanism between drug infusion and physiological reaction, and further limits popularization and application due to insufficient clinical interpretability and physiological consistency constraint. Therefore, developing an intelligent evaluation method which can integrate the drug metabolism history and the physiological time sequence characteristics, overcome the signal lag and have strict medical logic constraint is a technical difficulty in the field which needs to be broken through. Disclosure of Invention Aiming at the defects existing in the prior art, the invention aims to provide an anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology, which comprises the steps of firstly carrying out asymmetric sliding window construction and pharmacokinetic characteristic enhancement based on physiological time lags on collected drug infusion history and multi-parameter vital sign data, then constructing a drug effect-vital sign double-flow interaction network based on a Mamba state space model, realizing causal association modeling of drug effect on vital signs through a cross attention mechanism, carrying out multi-task joint optimization on the model by adopting a physiological consistency constraint loss function fused with medical ordinal logic on the basis, and finally realizing anesthesia depth real-time assessment with high precision, logic self-consistency and clinical transparency by combining an interpretability analysis module. In order to achieve the purpose, the invention provides the following technical scheme that the anesthesia depth assessment method based on Mamba architecture and multi-mode fusion technology is characterized by comprising the following steps: collecting multi-mode physiological data and drug infusion data of a patient through monitoring equipment, constructing an asymmetric sliding window of the multi-mode physiological data of the patient, and enhancing pharmacokinetic characteristics of the drug infusion data; Constructing a double-flow feature extraction network based on Mamba state space mod