Search

CN-121980363-A - Medical text-oriented interpretability high-precision classification model and attribution analysis method

CN121980363ACN 121980363 ACN121980363 ACN 121980363ACN-121980363-A

Abstract

The invention discloses an interpretability high-precision classification model and an attribution analysis method for medical texts, and aims to solve the problems of low medical text data quality, insufficient multi-modal information fusion and lack of interpretability of model decisions in the prior art. The method comprises the steps of cleaning and structuring an original medical examination report text, mining potential diagnosis information which is not recorded in the text in the image by utilizing a multi-mode comparison learning model based on a corresponding medical image, generating an atomized clinical concept, inserting the atomized clinical concept into the text to form an enhanced text report, encoding the original text and the enhanced text into semantic features respectively, and inputting a multi-layer perceptron classification model to predict disease risks. The invention obviously improves the classification precision through deep image-text semantic alignment, and simultaneously enhances the clinical credibility of the model by providing word-level attribution interpretation, thereby providing an efficient and reliable tool for auxiliary diagnosis.

Inventors

  • FENG ZHEN
  • CHEN GANG
  • XU XIAOQUN
  • LIAN YONG

Assignees

  • 温州医科大学附属第一医院

Dates

Publication Date
20260505
Application Date
20260407

Claims (10)

  1. 1. The method for the high-precision interpretive classification model and the attribution analysis for the medical texts is characterized by comprising the following steps of: Step 1, data cleaning and structuring are carried out on an original medical examination report text, and a clean text report is output; Step 2, mining potential diagnosis information which is not recorded by text in the image by utilizing a multi-mode comparison learning model based on the medical image corresponding to the clean text report, generating an atomized clinical concept, and inserting the atomized clinical concept into the clean text report to generate an enhanced text report; step 3, respectively inputting the clean text report and the enhanced text report into a pre-training language model for semantic feature coding to obtain text semantic feature representation, and inputting the text semantic feature representation into a trained multi-layer perceptron classification model to obtain a disease risk classification prediction result; and 4, calculating contribution values of each dimension in the text semantic feature representation to the prediction result of the multi-layer perceptron classification model by adopting an integral gradient method, and mapping the contribution values back to words of an original text to generate a visualized attribution report.
  2. 2. The medical text-oriented high-precision interpretable classification model and the attribute analysis method according to claim 1, wherein the data cleaning and structuring of the original medical examination report text in the step 1 specifically comprises: step 1.1, uniformly converting an original medical examination report text into a preset character coding format; Step 1.2, identifying and removing prefix phrase character strings which occur frequently in the original medical examination report text and are irrelevant to a core diagnosis conclusion; Step 1.3, adopting a regular expression technology to carry out format regulation on the text from which the prefix phrase character string is removed, wherein the format regulation comprises the steps of removing inconsistent line feed sequences, compressing a plurality of continuous space characters into single space and unifying paragraph separators; And 1.4, outputting the text with the regular completion format as the clean text report.
  3. 3. The medical text-oriented interpretable high-precision classification model and the attribute analysis method according to claim 1, wherein the generating an atomized clinical concept in the step 2 specifically includes: Analyzing standard medical documents related to the target diseases by using a large language model, and extracting clinical manifestation description paragraphs related to disease diagnosis; Designing concept decomposition prompt words, and guiding the large language model to decompose a section of complex clinical presentation description into a plurality of independent atomization clinical concepts, wherein the atomization clinical concepts are phrases for describing single and clear pathological signs or physiological states; And adjusting all the atomized clinical concepts output by the large language model based on historical data or expert data, wherein the manual auditing and sorting comprises deleting nonsensical character strings, re-expressing the fuzzy concepts, and merging synonymous concept expressions to form an atomized clinical concept candidate feature library.
  4. 4. The medical text-oriented high-precision interpretable classification model and the attribute analysis method according to claim 3, wherein the mining of the potential diagnosis information based on the medical image in the step 2 specifically includes: Acquiring a medical image corresponding to the clean text report; respectively extracting a visual feature vector of the medical image and a text feature vector of each concept in the candidate feature library of the atomized clinical concept by using a pre-trained multi-mode contrast learning model; calculating semantic similarity between the visual feature vector of the medical image and the text feature vector of each atomized clinical concept; And screening a plurality of atomized clinical concepts with highest semantic similarity according to a preset similarity threshold value to serve as the potential diagnosis information.
  5. 5. The medical text-oriented interpretable high-precision classification model and the attribute analysis method according to claim 1, wherein the semantic feature codes in the step 3 are specifically: The method comprises the steps of inputting a clean text report or an enhanced text report into a Chinese pre-training language model based on a transducer architecture, converting an input text into a sub-word mark sequence by a built-in word segmentation device of the Chinese pre-training language model, adding position codes for each sub-word mark, processing the sub-word mark sequence by a multi-layer transducer encoder, modeling semantic dependency relations among marks by a self-attention mechanism on each layer, carrying out nonlinear transformation by a feedforward neural network, and finally extracting hidden state vectors corresponding to specific category marks output by the Chinese pre-training language model to be used as text semantic feature representation.
  6. 6. The medical text-oriented interpretable high-precision classification model and the attribute analysis method according to claim 1, wherein the training process of the multi-layer perceptron classification model comprises the following steps: step 3.1, constructing a multi-layer perceptron classification model, wherein the multi-layer perceptron classification model comprises a hidden layer, a batch normalization layer is connected behind the hidden layer, and an output layer is a linear layer of a single node; Step 3.2, forming a training set by a plurality of text semantic feature representations corresponding to the clean text reports or the enhanced text reports, and labeling a disease risk real label corresponding to each report; Step 3.3, using a two-class cross entropy loss function, adopting an optimization algorithm with weight attenuation, and carrying out parameter optimization on the multi-layer perceptron classification model by using the training set; And 3.4, monitoring verification loss by adopting an independent verification set in the optimization process, and stopping training in advance when the verification loss is not reduced in a plurality of continuous training rounds to obtain the trained multi-layer perceptron classification model.
  7. 7. The medical text-oriented interpretable high-precision classification model and the attribute analysis method according to claim 5, wherein the calculating the contribution value by using the integral gradient method in the step 4 is specifically: And integrating the gradient of the output of the multi-layer perceptron classification model relative to the input characteristic along the linear path, and calculating to obtain the contribution value of each dimension in the text semantic characteristic representation to the prediction result.
  8. 8. The medical text-oriented high-precision interpretable classification model and attribute analysis method according to claim 7, wherein the mapping the contribution value back to the word of the original text in step 4 specifically comprises: Step 4.1, positioning a sub-word mark sequence corresponding to each word in the original text according to the word segmentation result of the Chinese pre-training language model on the input text; And 4.2, aiming at each word, marking all corresponding sub words on a gradient contribution value generated by the last layer of the Chinese pre-training language model, and aggregating in a mode of taking an average value or a maximum value, wherein the aggregated value is the importance score of the word on a final classification prediction decision.
  9. 9. The medical text-oriented high-precision interpretable classification model and the attribute analysis method according to claim 8, wherein the generating the visual attribute report in the step 4 is specifically: according to the calculated importance score of each word, highlighting a preset number of words with the highest importance score in an original text; The highlighting is performed in such a way that the highlighting is performed using a color gradient, the color being graded from red, which represents a high positive contribution, to blue, which represents a high negative contribution; And meanwhile, labeling specific importance score values beside the highlighted words to form the visual attribution report.
  10. 10. The medical text-oriented high-precision interpretable classification model and the causal analysis method according to claim 3, wherein the original medical examination report text is a lower limb vein ultrasound examination report text, the medical image is a lower limb vein ultrasound image, the target disease is pulmonary thromboembolism, and the disease risk classification prediction result is the risk probability of occurrence of pulmonary thromboembolism.

Description

Medical text-oriented interpretability high-precision classification model and attribution analysis method Technical Field The invention relates to the technical field of artificial intelligence and medical information processing, in particular to an interpretability high-precision classification model and an attribution analysis method for medical texts. Background In the field of clinical auxiliary diagnosis, the automatic analysis and classification of medical examination reports by using natural language processing technology is an important means for improving diagnosis efficiency and realizing early warning of disease risks. For example, predicting the risk of a patient to develop pulmonary thromboembolism based on lower extremity venous ultrasound report text is of significant clinical value. However, building a high-performance, reliable medical text classification model faces multiple technical challenges. First, the original medical text data is ragged in quality. The inspection report typically contains a large number of non-standardized representations, redundant formatting indicia, and descriptive prefixes that are not relevant to the core diagnostics. These data noises can directly interfere with the capture of key medical concepts by the model, limiting the upper limit of classification performance. The traditional text cleaning method depends on fixed rules and dictionaries, is difficult to adapt to diversified expression habits of different medical institutions and different doctors, and causes insufficient generalization capability. Second, medical diagnostics is essentially a process of multimodal information fusion. Taking ultrasonic examination as an example, the morphology, hemodynamics and other abundant information contained in the image often cannot be completely and accurately described in a short text report. Most of the prior art schemes only use text mode, or simply process text and images in parallel, and fail to realize deep and semantic graphic information complementation and alignment, so that a great number of potential visual clues which are critical to classification are omitted. Furthermore, existing deep learning-based medical text classification models, particularly schemes using pre-trained language models in combination with multi-layer perceptrons, are generally subject to the problem of "black boxes" although there is some improvement in performance. The internal decision logic of the model is opaque, and the doctor cannot understand which specific medical description words or concepts the model makes a judgment of "high risk" or "low risk". This lack of interpretability severely hampers the practical application of the model in serious medical scenarios and the trust of the clinician. Therefore, the problems in the prior art can be summarized as firstly that the model input quality is limited by the isomerism and noise of the original medical text data, secondly that the model input information is incomplete due to the fact that potential diagnosis information which is not expressed in the text in an inspection image cannot be effectively mined and fused, and thirdly that a model prediction result lacks visual decision basis interpretation which accords with medical cognition. Disclosure of Invention Based on the above object, the invention provides an interpretable high-precision classification model and an attribution analysis method for medical texts, wherein the method comprises the following steps: Step 1, data cleaning and structuring are carried out on an original medical examination report text, and a clean text report is output; Step 2, mining potential diagnosis information which is not recorded by text in the image by utilizing a multi-mode comparison learning model based on the medical image corresponding to the clean text report, generating an atomized clinical concept, and inserting the atomized clinical concept into the clean text report to generate an enhanced text report; step 3, respectively inputting the clean text report and the enhanced text report into a pre-training language model for semantic feature coding to obtain text semantic feature representation, and inputting the text semantic feature representation into a trained multi-layer perceptron classification model to obtain a disease risk classification prediction result; and 4, calculating contribution values of each dimension in the text semantic feature representation to the prediction result of the multi-layer perceptron classification model by adopting an integral gradient method, and mapping the contribution values back to words of an original text to generate a visualized attribution report. Preferably, in the step 1, the data cleaning and structuring of the original medical examination report text specifically includes: step 1.1, uniformly converting an original medical examination report text into a preset character coding format; Step 1.2, identifying and removing prefix phrase