CN-122000062-A - Sepsis risk multi-modal prediction method and system based on deep learning

CN122000062ACN 122000062 ACN122000062 ACN 122000062ACN-122000062-A

Abstract

The invention discloses a sepsis risk multi-mode prediction method and a system thereof based on deep learning, belonging to the technical field of medical data intelligent analysis, wherein the method comprises the steps of acquiring vital sign sequences, bedside instant inspection, laboratories, microorganism culture, images and medical history data by multi-mode data acquisition; the method comprises a time sequence feature encoding step, a cross-mode fusion prediction step, an organ function dynamic evaluation step, a risk layering early warning step and a risk feedback optimization encoding parameter, wherein the time sequence feature encoding step extracts a time sequence embedded vector through a multi-scale time sequence encoding network, the cross-mode fusion prediction step fuses features based on a dynamic attention mechanism and predicts sepsis occurrence probability, the organ function dynamic evaluation step evaluates a missing test result and calculates organ function scores, and the risk layering early warning step outputs a four-level risk layering result and feeds back optimized encoding parameters.

Inventors

ZHANG SHUANG
JIANG RONG
YANG HANG
LI YUN
HU QIAN

Assignees

珠海市金湾中心医院

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. The sepsis risk multi-modal prediction method based on deep learning is characterized by comprising the following steps of: A multi-mode data acquisition step of acquiring vital sign sequence data, bedside instant test data, conventional laboratory data, microorganism culture data, infection focus image data and past medical history data of an ICU patient, and performing time stamp alignment and missing value marking on the acquired multi-source heterogeneous data to generate a standardized feature set; a time sequence feature coding step, namely inputting time sequence data in the standardized feature set into a multi-scale time sequence coding network, wherein the multi-scale time sequence coding network carries out parallel feature extraction on the vital sign sequence data respectively through a first time window, a second time window and a third time window, carries out cascade fusion on features of different time scales, and outputs a time sequence embedding vector; A cross-modal fusion prediction step, namely calculating a modal weight coefficient between the time sequence embedded vector and static feature vectors corresponding to the bedside instant test data and the conventional laboratory data based on a cross-modal dynamic attention mechanism, carrying out weighted fusion on each modal feature according to the modal weight coefficient to generate a fusion feature representation, and inputting the fusion feature representation into a sepsis prediction network to output sepsis occurrence probability; An organ function dynamic evaluation step, namely estimating unobtained test results in real time through an organ function deficiency value dynamic estimation model according to the fusion characteristic representation and organ function related parameters in the standardized characteristic set, and calculating organ function scores of a respiratory system, a circulatory system, a liver, a coagulation system, a kidney and a nervous system based on the estimated values and the measured values; And a risk layering early warning step, namely dividing the patient into four risk grades of low risk, medium risk, high risk and high risk by a multi-factor linkage layering model according to the sepsis occurrence probability and the organ function score, outputting risk grade labels, risk evolution trend curves and key risk factor ordering, and feeding early warning results back to the time sequence feature encoding step to update encoding parameters.
2. The method of claim 1, wherein in the multi-modality data acquisition step, the vital sign sequence data comprises heart rate, systolic pressure, diastolic pressure, mean arterial pressure, respiratory rate, and body temperature, the sampling frequency is once per minute, and the bedside on-the-fly test data comprises lactate value, procalcitonin concentration, and C-reactive protein concentration.
3. The method of claim 1, wherein in the time series feature encoding step, the first time window is 30 minutes in length, the second time window is 2 hours in length, the third time window is 6 hours in length, and the three time windows capture short term fluctuation features, mid term trend features, and long term evolution features, respectively.
4. The method of claim 1, wherein in the risk stratification pre-warning step, the patient is classified as a low risk level when the sepsis occurrence probability is less than a preset first probability threshold, the patient is classified as a medium risk level when the sepsis occurrence probability is greater than or equal to the first probability threshold and less than a preset second probability threshold, the patient is classified as a high risk level when the sepsis occurrence probability is greater than or equal to the second probability threshold and less than a preset third probability threshold, and the patient is classified as a very high risk level when the sepsis occurrence probability is greater than or equal to the third probability threshold.
5. The method according to claim 1, wherein in the cross-modal fusion prediction step, the cross-modal dynamic attention mechanism comprises taking the time sequence embedded vector as a query vector, taking the static feature vector as a key vector and a value vector, calculating a correlation score between modalities by scaling dot product attention, and normalizing the correlation score to obtain the modal weight coefficient.
6. The method according to claim 1, wherein in the organ function dynamic evaluation step, the organ function deficiency value dynamic estimation model estimates a posterior distribution of non-acquired test results by a bayesian inference network based on the acquired vital sign data and a part of the test results, taking an expected value of the posterior distribution as an estimated value of the deficiency value.
7. The method according to claim 1, wherein in the risk stratification pre-warning step, the multi-factor linkage stratification model uses the sepsis occurrence probability, the organ function score and the trend of the organ function score as input features, calculates a comprehensive risk score by a gradient-lifting decision tree, and determines a risk level according to the comprehensive risk score.
8. The method of claim 1, wherein the multi-scale temporal coding network comprises a position coding layer that adds temporal position information to the input temporal data, a multi-headed self-attention layer that captures long-range dependencies in the temporal data, and a feed-forward neural network layer that non-linearly transforms the attention output.
9. The method of claim 1, wherein the key risk factor ranking is generated by calculating gradient contribution values of feature dimensions in the fused feature representation to the probability of sepsis occurrence, ranking from large to small absolute values of gradient contribution values, and outputting a pre-set number of features with the contribution values ranked as key risk factors.
10. A deep learning based sepsis risk multimodal prediction system for implementing the method of any one of claims 1-9, comprising: The multi-mode data acquisition module is used for acquiring vital sign sequence data, bedside instant test data, conventional laboratory data, microorganism culture data, infection focus image data and past medical history data of an ICU patient, and performing time stamp alignment and missing value marking on the acquired multi-source heterogeneous data to generate a standardized feature set; The time sequence feature coding module is used for inputting the time sequence data in the standardized feature set into a multi-scale time sequence coding network, respectively extracting and cascading and fusing the parallel features in different time windows, and outputting a time sequence embedded vector; The cross-modal fusion prediction module is used for calculating a modal weight coefficient based on a cross-modal dynamic attention mechanism, carrying out weighted fusion on each modal characteristic according to the modal weight coefficient to generate fusion characteristic representation, and outputting sepsis occurrence probability; The organ function dynamic evaluation module is used for estimating unobtained test results in real time through the organ function deficiency value dynamic estimation model, and calculating organ function scores of six organ systems based on the estimated values and the measured values; The risk layering early warning module is used for dividing a patient into four risk grades of low risk, medium risk, high risk and high risk through the multi-factor linkage layering model according to the sepsis occurrence probability and the organ function score, outputting risk grade labels, risk evolution trend curves and key risk factor ordering, and feeding early warning results back to the time sequence feature encoding module to update encoding parameters.

Description

Sepsis risk multi-modal prediction method and system based on deep learning Technical Field The invention relates to the technical field of medical data intelligent analysis, in particular to a sepsis risk multi-mode prediction method and a sepsis risk multi-mode prediction system based on deep learning. Background Sepsis is a systemic inflammatory response syndrome caused by infection and is one of the leading causes of death in intensive care patients. Early identification of sepsis is of great importance for improving patient prognosis, whereas sepsis onset is occult, early symptoms are not specific, and traditional identification methods have obvious hysteresis. Existing sepsis risk assessment methods rely primarily on sequential organ failure Scores (SOFA) and rapid sequential organ failure scores (qSOFA). SOFA scoring requires complete laboratory test results, including multiple indicators of blood gas analysis, liver function, kidney function, clotting function, and the like, and typically takes hours or even more from blood sampling to test results, which results in a hysteresis in identification. qSOFA scoring, while simplifying the evaluation procedure, is not sensitive enough to permit missed diagnosis of early sepsis patients. Chinese invention CN119025109a discloses an embedded interface design method and system for sepsis management software. The technical scheme mainly solves the problem of integration between sepsis management software and an original medical system, and seamless integration of the sepsis management software and the original medical system is realized by designing an embedded interface of two modes of a folding state and an unfolding state. The data analysis module of the scheme is used for acquiring medical record data and analyzing the medical record data, and automatically carries out intelligent evaluation and early warning on the risk of the patient developing sepsis. However, the invention has the technical defects that firstly, at the data acquisition level, only medical record data including basic information of a patient, laboratory examination results, image reports and doctor orders are mentioned, real-time acquisition and time sequence analysis of continuous vital sign monitoring data are not involved, and dynamic change trend of the physiological state of the patient cannot be captured. Secondly, on the aspect of risk prediction, a specific risk assessment algorithm is not disclosed, only the key information is extracted for analysis, whether the patient has high risk of sepsis is assessed, a deep learning prediction model with multi-mode data fusion is lacked, and early warning before the sepsis diagnosis standard is met is difficult to realize. Third, in the organ function evaluation level, the real-time evaluation of the degree of organ dysfunction is not involved, and the problem of evaluation lag caused by waiting for a complete test result cannot be solved. Fourth, in the risk stratification layer, only the binary judgment that the risk preset threshold is reached is mentioned, a multi-level risk stratification system is not established, and the risk evolution trend and the key risk factor ordering cannot be output. Therefore, there is a need for an ICU sepsis risk intelligent assessment method that can integrate multidimensional clinical data, achieve early risk prediction, support organ function real-time assessment, and provide refined risk stratification. Disclosure of Invention Aiming at the problems of recognition lag, insufficient data fusion, dependence on complete test results for organ function evaluation, rough risk stratification and the like in sepsis risk assessment in the prior art, the invention provides a sepsis risk multi-mode prediction method and a sepsis risk multi-mode prediction system based on deep learning. The invention provides a deep learning-based sepsis risk multi-modal prediction method, which comprises a multi-modal data acquisition step, wherein vital sign sequence data, bedside instant test data, conventional laboratory data, microorganism culture data, infection focus image data and past medical history data of an ICU patient are acquired, and time stamp alignment and missing value marking are carried out on the acquired multi-source heterogeneous data to generate a standardized feature set. And a time sequence feature coding step, namely inputting time sequence data in the standardized feature set into a multi-scale time sequence coding network, respectively extracting parallel features of vital sign sequence data in time windows with different lengths, carrying out cascade fusion on features with different time scales, and outputting time sequence embedded vectors. And a cross-modal fusion prediction step, namely calculating a modal weight coefficient between the time sequence embedded vector and the static feature vector based on a cross-modal dynamic attention mechanism, carrying out weighted fusion on each modal feat