Search

CN-121983309-A - ICU severe patient multi-mode risk early warning system based on deep learning

CN121983309ACN 121983309 ACN121983309 ACN 121983309ACN-121983309-A

Abstract

The invention belongs to the technical field of artificial intelligence and medical health information processing, in particular relates to an ICU severe patient multi-mode risk early warning system based on deep learning, and aims to solve the problems of delayed disease prediction, difficult multi-source data fusion and low single index early warning precision in intensive care. The system collects high-frequency physiological signals, clinical observation values, inspection images and unstructured texts through a data perception layer, cross-modal embedding and contrast learning alignment are carried out through a heterogeneous fusion layer, a shared semantic space is constructed, a time sequence reasoning layer utilizes an improved space-time diagram neural network modeling dynamic knowledge graph to capture causal and time sequence dependence among clinical events, and a risk decision layer outputs multidimensional risk probabilities of sepsis, respiratory failure and sudden cardiac arrest in parallel. The system supports online incremental learning and individualized time attenuation modeling, improves timeliness, accuracy and interpretability of early warning, and is obviously superior to the traditional method in clinical verification.

Inventors

  • LIU JIANWEI

Assignees

  • 刘建卫

Dates

Publication Date
20260505
Application Date
20260122

Claims (10)

  1. 1. ICU severe patient multi-modal risk early warning system based on deep learning, characterized by comprising: the data perception layer is used for accessing a high-frequency physiological signal stream, a structured clinical observation value, a semi-structured inspection and image report and an unstructured medical text in parallel from the hospital information system so as to obtain four types of heterogeneous data streams; The heterogeneous fusion layer is used for respectively carrying out mode special coding on the four types of heterogeneous data streams and executing cross-mode contrast learning so as to obtain a unified characterization vector sequence aligned in a shared semantic space; the time sequence reasoning layer is used for constructing an individualized dynamic knowledge graph based on the unified characterization vector sequence, and carrying out message transmission and state updating through a space-time graph neural network integrating medical priori knowledge, a gating mechanism and an individualized time attenuation function so as to obtain a comprehensive embedded vector for characterizing the health state of a patient; And the risk decision layer is used for receiving the comprehensive embedded vector, driving a plurality of prediction branches aiming at different clinical events in parallel, and integrating through a learnable weighted fusion unit capable of dynamically adjusting the weight of each branch according to the diagnosis of the patient complaints so as to generate a hierarchical early warning index and pushing the hierarchical early warning index to a medical care workstation.
  2. 2. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 1, wherein the heterogeneous fusion layer comprises: the physiological signal coding subunit is used for processing the high-frequency physiological signal flow through a parallel one-dimensional convolutional neural network and a self-attention mechanism so as to obtain a time sequence feature vector sequence; The structured observation coding subunit is used for carrying out nonlinear mapping on the standardized structured clinical observation value through a fully-connected network so as to obtain a dense characterization vector; The report semantic coding subunit is used for carrying out semantic coding on the semi-structured inspection and the image report through a pre-training medical language model so as to obtain pathological indication feature vectors; The text event extraction subunit is used for processing the unstructured medical text through a field-adaptive two-way long-short-term memory network and a named entity recognition module so as to extract a key event characterization vector with a time stamp.
  3. 3. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 2, wherein the heterogeneous fusion layer further comprises: the public projection subunit is used for respectively and linearly mapping the time sequence feature vector sequence, the dense characterization vector, the pathological characterization feature vector and the key event characterization vector to the same high-dimensional semantic space; And the cross-modal alignment subunit is used for constructing positive sample pairs based on different modal data items of the same patient in the same time window, constructing negative sample pairs based on different time data items of different patients, and optimizing parameters of the public projection subunit through contrast learning loss functions so as to shorten the distance between the positive sample pairs and push away the negative sample pairs.
  4. 4. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 1, wherein the temporal inference layer comprises: The dynamic diagram construction subunit is used for dividing a patient hospitalization time line into continuous analysis windows, taking the unified characterization vector in each window as a node, and initializing directed edges and weights thereof according to the time co-occurrence probability among the nodes and a preset medical causal rule matrix so as to construct the personalized dynamic knowledge graph; and the space-time diagram neural network subunit is used for aggregating neighbor node information at each time step to update the current node state, wherein the aggregation process is controlled by a message commonly modulated by the edge weight and the attention coefficient, and the integration of the history memory and the new information is managed through the gating circulation unit.
  5. 5. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 4, wherein the space-time diagram neural network subunit further comprises: an individuation time attenuation subunit for dynamically adjusting the attenuation coefficient of the time attenuation function according to the age, the basic disease burden index and the admission score of the patient, so that the gating circulation unit gives exponentially decreasing weight to the long-term observation, and the history information retention period of the chronic patient is longer than that of the young patient.
  6. 6. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 1, wherein the risk decision layer comprises: a sepsis risk prediction branch for determining a risk probability of sepsis episodes within a first predetermined time period in the future based on the comprehensive embedding vector using a classifier with a class imbalance compensation mechanism; A respiratory failure risk prediction branch, which is used for estimating a cumulative risk function of the tracheal intubation to be implemented in a second preset time period in the future based on the comprehensive embedding vector by adopting a depth part logistic regression model under a survival analysis framework; and the cardiac arrest risk prediction branch is used for quantifying the crisis probability of cardiac arrest in a third preset time period in the future by comparing reconstruction deviation of the original multi-mode input by the self-encoder by adopting a reconstruction error estimator based on an abnormality detection principle.
  7. 7. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 6, wherein the learnable weighted fusion unit is configured to dynamically calculate and assign respective fusion weights of the sepsis risk prediction branch, the respiratory failure risk prediction branch, and the cardiac arrest risk prediction branch according to a complaint diagnostic tag of a current patient using a gated attention mechanism.
  8. 8. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 1, wherein the risk decision layer further comprises: the evidence chain generation subunit is used for synchronously generating the grading early warning indexes, extracting and organizing key evidence supporting early warning judgment, wherein the key evidence comprises physiological waveform fragments with the most discriminant, abnormal laboratory numerical trend graphs and dangerous signal keywords in text records.
  9. 9. The deep learning based ICU critically ill patient multimodal risk early warning system of claim 1, further comprising: And the online incremental learning module is used for automatically triggering a small batch retraining process after the complete diagnosis and treatment path data of the newly diagnosed case is backfilled into the database so as to periodically fine-tune the model parameters of the heterogeneous fusion layer, the time sequence reasoning layer and the risk decision layer.
  10. 10. The ICU severe patient multi-mode risk early warning system based on deep learning according to claim 1 is characterized in that the system is deployed on a private cloud platform of a hospital, a containerized micro-service architecture is adopted, all functional modules communicate through well-defined application program interfaces, a national secret SM4 encryption algorithm is adopted for data transmission, and a transparent database encryption technology is adopted for static data storage.

Description

ICU severe patient multi-mode risk early warning system based on deep learning Technical Field The invention belongs to the technical field of artificial intelligence and medical health information processing, and particularly relates to an ICU severe patient multi-mode risk early warning system based on deep learning. Background The advanced fusion of artificial intelligence technology in the field of medical health is pushing a clinical decision support system to develop rapidly in an intelligent and accurate direction. As an important branch of artificial intelligence, deep learning is widely applied to multiple scenes such as medical image analysis, electronic medical record mining, physiological signal processing and the like by virtue of strong nonlinear mapping capability and multi-level feature extraction advantages, and the efficiency and accuracy of disease prediction, diagnosis assistance and treatment scheme recommendation are remarkably improved. The multi-modal data fusion is used as a key path for realizing comprehensive patient state characterization, and aims to integrate heterogeneous information from different sources, such as vital sign monitoring data, laboratory test results, imaging reports, nursing records and the like, so as to construct a more complete and dynamic disease evolution model. Among them, intensive Care Unit (ICU) is used as the clinical unit with the most critical illness state and the most rapid change in hospitals, and has extremely high timeliness requirements for early risk identification and intervention. In this environment, the physiological state of the patient exhibits highly non-linear and complex coupling characteristics, and it is difficult for single modality data to accurately capture potentially deteriorating precursor signals. Therefore, the risk early warning method based on the multi-mode data becomes a core research direction for improving ICU management quality, and the basic aim is to realize early prediction of severe clinical events such as sepsis, acute respiratory failure, arrhythmia and the like by modeling space-time correlation among multi-source information. In the prior art, although the deep neural network is tried to be introduced into an ICU risk early warning task and trend prediction and anomaly detection of partial physiological parameters are primarily realized, a series of key technical bottlenecks are faced, namely the problems of inconsistent time scale, large sampling frequency difference, uneven missing value distribution and the like of multi-mode data generally exist, so that the traditional fusion strategy is difficult to effectively align semantic information, most models adopt a static weight distribution mechanism to carry out modal fusion, the contribution degree of each mode cannot be dynamically adjusted according to the real-time state of a patient, the adaptability of the models in individuation early warning is weakened, meanwhile, the deep network structure is easily influenced by the problems of small samples and unbalanced categories in the training process, the generalization performance is limited, the false alarm rate is high, and in addition, the traditional system lacks the interpretable output of early warning results, and is difficult to provide a credible decision basis for clinicians. The defects are particularly prominent in the clinical practice of ICU with high strength and fast pace, and the practical floor application of an intelligent early warning system is severely restricted, so that a new architecture capable of realizing efficient, robust and interpretable multi-modal risk early warning is needed. Disclosure of Invention The invention aims to provide an ICU severe patient multi-modal risk early warning system based on deep learning, which aims to solve the technical problems of untimely prediction of patient disease deterioration, weak multi-source heterogeneous data fusion capability and insufficient sensitivity and specificity of single-modal monitoring indexes in the current intensive care environment. In modern ICU clinical practice, although vital sign monitoring equipment is highly developed, physiological parameters such as heart rate, blood pressure, blood oxygen saturation and the like can be acquired in real time, and meanwhile, an electronic medical record system continuously records laboratory test results, imaging reports and nursing documents, the existing early warning mechanism still generally depends on a single-parameter alarm triggered by a fixed threshold or an early warning scoring system based on a simple scoring rule. The method is difficult to capture nonlinear dynamic characteristics in the complex pathological evolution process, is easy to generate a large number of false positive alarms, and causes alert fatigue of medical staff, and more importantly, the method lacks the capability of carrying out deep semantic association and cross-domain collaborative modeling on