Search

CN-122025195-A - ECMO offline success rate prediction method, system and storage medium based on dual channels

CN122025195ACN 122025195 ACN122025195 ACN 122025195ACN-122025195-A

Abstract

The ECMO offline success rate prediction method, system and storage medium based on the dual channels comprise the steps of inputting structured time sequence data and multi-mode heterogeneous data into a dual-channel data processing architecture, processing the structured time sequence data by a first channel to obtain a first feature vector, processing the multi-mode heterogeneous data by a second channel to obtain a second feature vector, obtaining a clinical state feature vector and a clinical sensitive index vector according to the structured time sequence data and the multi-mode heterogeneous data, inputting the first feature vector, the second feature vector, the clinical state feature vector and the clinical sensitive index vector into a gating fusion network to obtain fusion features, inputting query sentences and the fusion features into a knowledge injection layer to obtain a plurality of related knowledge segments, inputting the query sentences and the fusion features into a data feedback layer to obtain local rules, constructing the fusion features, the knowledge segments and the local rules into prompt words, and inputting the prompt words into a large language model to obtain ECMO offline success rate and reason.

Inventors

  • WANG HUANLI
  • JIN KUI
  • ZHANG MENGPING

Assignees

  • 安徽省立医院(中国科学技术大学附属第一医院)

Dates

Publication Date
20260512
Application Date
20260415

Claims (10)

  1. 1. The ECMO offline success rate prediction method based on the double channels is characterized by comprising the following steps of: The method comprises the steps of inputting structured time sequence data and multi-mode heterogeneous data into a dual-channel data processing architecture, wherein the dual-channel data processing architecture comprises a first channel and a second channel, the first channel is used for processing the structured time sequence data to obtain a first characteristic vector, and the second channel is used for processing the multi-mode heterogeneous data to obtain a second characteristic vector; Inputting the first feature vector, the second feature vector, the clinical state feature vector and the clinical sensitive index vector into a gating fusion network to obtain fusion features; Inputting the query statement and the fusion feature into a knowledge injection layer, and taking the query statement and the fusion feature as search conditions to perform similarity search in a vector database to obtain a plurality of related knowledge segments; inputting the query statement and the fusion feature into a data feedback layer, and retrieving a matched local rule from a local rule base; Constructing the fusion characteristics, the knowledge segments, the local rules, the clinical state characteristic vectors and the clinical sensitive index vectors into prompt words, wherein the prompt words comprise a plurality of modules, a state statement module, a knowledge reference module, a local reference module and an reasoning instruction module; And inputting the prompt word into a large language model to obtain the ECMO offline success probability and reason.
  2. 2. The dual-channel-based ECMO offline success rate prediction method of claim 1, wherein the processing the structured time series data by the first channel to obtain a first feature vector comprises: the structured time sequence data comprises static structured data and dynamic time sequence data, wherein the static structured data is an index which does not change with time, and the dynamic time sequence data is a physiological index with a time stamp, which is acquired in real time in the treatment process; The first channel identifies ECMO clinical stages and ECMO modes corresponding to the structured time series data, wherein the ECMO modes comprise a first mode and a second mode; The first mode is a venous-venous adventitia pulmonary oxygenation mode, and the second mode is a venous-arterial adventitia pulmonary oxygenation mode; the first channel divides a feature extraction window according to the clinical stages of ECMO; when the structured time sequence data corresponds to a first mode, a first channel defines a first key index in the first mode, and a feature value corresponding to the first key index is calculated in a corresponding feature extraction window to obtain a first feature vector; When the structured time sequence data corresponds to a second mode, a first channel defines a second key index in the second mode, and a feature value corresponding to the second key index is calculated in a corresponding feature extraction window to obtain a first feature vector; Each dimension of the first feature vector corresponds to an ECMO offline evaluation dynamic index with definite clinical significance, and in different ECMO modes, the same dimension of the first feature vector corresponds to different key indexes, wherein the key indexes in the VV-ECMO mode comprise oxygenation index improvement slope, dynamic lung compliance change curve area and carbon dioxide partial pressure reduction rate, and the key indexes in the VA-ECMO mode comprise lactic acid clearance rate, vasoactive drug score reduction trend and average arterial pressure stability.
  3. 3. The dual-channel-based ECMO offline success rate prediction method of claim 1, wherein the processing the multi-modal heterogeneous data by the second channel to obtain a second feature vector includes: the multi-modal heterogeneous data comprises natural language text, image data and structured nonstandard data; The second channel adopts a large language model to extract a semantic embedded vector of the natural language text, adopts a visual model to extract a visual feature vector of the image data, and converts the structured nonstandard data into a text feature vector; and the semantic embedded vector, the visual feature vector and the text feature vector are integrated into a second feature vector through vector splicing.
  4. 4. The dual-channel ECMO offline success rate prediction method according to claim 1, wherein said obtaining a clinical status feature vector and a clinical sensitivity index vector according to the structured time series data and the multi-modal heterogeneous data comprises: calculating matching degrees of three clinical stages of ECMO according to the structured time sequence data and the multi-mode heterogeneous data, wherein the matching degrees comprise an on-machine preparation period matching degree, a stable period matching degree and an off-machine preparation period matching degree; Combining the on-machine preparation period matching degree, the stable period matching degree and the off-machine preparation period matching degree into a three-dimensional column vector to obtain the clinical state feature vector; According to the structured time sequence data and the multi-mode heterogeneous data, calculating risk sensitivity of six ECMO indexes, namely lactic acid level, APACHE II score, diastolic pressure, blood pH value, albumin level and activated partial thromboplastin time through a Sigmoid risk normalization formula calibrated by clinical data; And combining the risk sensitivities into a six-dimensional column vector to obtain the clinical sensitivity index vector.
  5. 5. The dual-channel-based ECMO offline success rate prediction method of claim 1, wherein the inputting the query sentence and the fusion feature into the knowledge injection layer, performing similarity search in a vector database using the query sentence and the fusion feature as search conditions, and obtaining a plurality of related knowledge segments includes: The vector database is composed of knowledge segments formed by embedding clinical guidelines, expert consensus and latest documents; the plurality of related knowledge pieces are marked with time stamps and confidence degrees; and updating the vector database through real-time networking when the relevant knowledge segments do not exist in the vector database.
  6. 6. The dual-channel ECMO offline success rate prediction method according to claim 1, wherein said inputting the query statement and the fusion feature to the data feedback layer, retrieving the matching local rule from the local rule base, comprises: the local rule base is composed of a plurality of local rules; And analyzing the ECMO offline patient history data by the local rule through a retrieval enhancement generation technology, and filtering through a time window to obtain the ECMO offline patient history data.
  7. 7. The dual-channel ECMO offline success rate prediction method according to claim 1, wherein the constructing the fusion feature, the knowledge segment, the local rule, the clinical status feature vector, the clinical sensitivity index vector as a hint word includes a plurality of modules, a state statement module, a knowledge reference module, a local reference module, and an inference instruction module, including: Filling the fusion feature and clinical state feature vector into the state statement module, filling the knowledge segment and clinical sensitive index vector into the knowledge reference module, and filling the local rule into the local reference module; The reasoning instruction module comprises a multi-step reasoning instruction and comprises a knowledge fragment evaluation step, a local rule evaluation step, a conflict resolution step, a probability calculation step and a reason generation step.
  8. 8. The dual channel-based ECMO offline success rate prediction method as described in claim 7, wherein, The state statement module combines the ECMO mode identification vector to generate a patient state abstract matched with the mode; The knowledge quotation module sets confidence coefficient for the knowledge segments according to the dimension values of the corresponding knowledge segment association indexes in the clinical sensitive index vector, and adjusts the quotation strength of the corresponding knowledge segments in the prompt words according to the confidence coefficient; The local reference module screens the local rules according to the ECMO mode identification vector, reserves the local rules related to the mode, and removes the local rules irrelevant to the mode.
  9. 9. The ECMO offline success rate prediction system based on the double channels is characterized by comprising a double-channel data processing module, an enhanced reasoning module and a reasoning output module: The system comprises a dual-channel data processing module, a gate control fusion network, a data processing module and a data processing module, wherein the dual-channel data processing module is used for inputting structured time sequence data and multi-mode heterogeneous data into the dual-channel data processing architecture, the dual-channel data processing architecture comprises a first channel and a second channel, the first channel processes the structured time sequence data to obtain a first characteristic vector, and the second channel processes the multi-mode heterogeneous data to obtain a second characteristic vector; the system comprises a knowledge injection layer, a data feedback layer, a query statement and fusion feature, a local rule library, a prompt word, a state statement module, a knowledge reference module, a local reference module and an inference instruction module, wherein the knowledge injection layer is used for inputting the query statement and the fusion feature into the knowledge injection layer, taking the query statement and the fusion feature as search conditions, and carrying out similarity search on a vector database to obtain a plurality of related knowledge fragments; and the reasoning output module is used for inputting the prompt word into the large language model to obtain the ECMO offline success probability and reason.
  10. 10. A storage medium having stored thereon a computer program which, when run, performs the two-channel based ECMO offline success rate prediction method according to any of claims 1-8.

Description

ECMO offline success rate prediction method, system and storage medium based on dual channels Technical Field The application relates to the technical field of artificial intelligence, in particular to a method, a system and a storage medium for predicting ECMO offline success rate based on double channels. Background External membrane oxygenation (Extracorporeal Membrane Oxygenation, ECMO) is an artificial heart-lung machine with the core function of replacing the heart and lung of the human body for oxygen exchange and blood circulation, but an invasive life support technique, itself, can bring about a number of complications. If the blind attempt is offline, the patient needs to be re-on-line once it fails. Repeating the machine can not only add the complications, but also cause secondary injury to the heart and lung functions, and remarkably improve the death rate of patients. The off-line timing of ECMO has no absolute unified standard, and the disease conditions, basic diseases and cardiopulmonary function recovery speed of different patients are greatly different. The ECMO off-line success rate refers to the proportion of patients treated with ECMO (extracorporeal membrane oxygenation) assistance that successfully departed from ECMO equipment after improvement of the condition, and is one of the core indicators for assessing the efficacy of ECMO treatment. For patients with high offline success probability, the offline flow can be started in time, so that unnecessary ECMO support time is reduced, and the rehabilitation process of the patients is accelerated. For patients with low off-line success probability, off-line treatment schemes (such as optimizing breathing machine parameters, reinforcing myocardial nutrition, controlling infection and the like) can be suspended, and the patients are assessed after the cardiopulmonary function of the patients is improved, so that treatment failure caused by the early off-line is avoided. Disclosure of Invention The present application has been made to solve the above-described problems. According to an aspect of the present application, there is provided a dual channel-based ECMO offline success rate prediction method, the method including: The method comprises the steps of inputting structured time sequence data and multi-mode heterogeneous data into a dual-channel data processing architecture, wherein the dual-channel data processing architecture comprises a first channel and a second channel, the first channel is used for processing the structured time sequence data to obtain a first characteristic vector, and the second channel is used for processing the multi-mode heterogeneous data to obtain a second characteristic vector; Inputting the first feature vector, the second feature vector, the clinical state feature vector and the clinical sensitive index vector into a gating fusion network to obtain fusion features; Inputting the query statement and the fusion feature into a knowledge injection layer, and taking the query statement and the fusion feature as search conditions to perform similarity search in a vector database to obtain a plurality of related knowledge segments; inputting the query statement and the fusion feature into a data feedback layer, and retrieving a matched local rule from a local rule base; Constructing the fusion characteristics, the knowledge segments, the local rules, the clinical state characteristic vectors and the clinical sensitive index vectors into prompt words, wherein the prompt words comprise a plurality of modules, a state statement module, a knowledge reference module, a local reference module and an reasoning instruction module; And inputting the prompt word into a large language model to obtain the ECMO offline success probability and reason. In one embodiment of the present application, the processing, by the first channel, the structured time series data to obtain a first feature vector includes: the structured time sequence data comprises static structured data and dynamic time sequence data, wherein the static structured data is an index which does not change with time, and the dynamic time sequence data is a physiological index with a time stamp, which is acquired in real time in the treatment process; The first channel identifies ECMO clinical stages and ECMO modes corresponding to the structured time series data, wherein the ECMO modes comprise a first mode and a second mode; The first mode is a venous-venous adventitia pulmonary oxygenation mode, and the second mode is a venous-arterial adventitia pulmonary oxygenation mode; the first channel divides a feature extraction window according to the clinical stages of ECMO; when the structured time sequence data corresponds to a first mode, a first channel defines a first key index in the first mode, and a feature value corresponding to the first key index is calculated in a corresponding feature extraction window to obtain a first feature vector; When the structur