Research On Predicting The Risk Of Deterioration Of COVID-19 Patients Based On Interpretable Algorithms

The surge in the number of cases of coronavirus disease 2019(COVID-19)infection poses a significant challenge to the management of healthcare resources.Although roughly 81% of COVID-19 patients exhibit mild or moderate symptoms,some have been observed to deteriorate suddenly,rapidly developing into the severe or critically ill categories Early prediction and aggressive treatment of patients with a high risk of deterioration is therefore essential.Data on patients with COVID-19 from the Wuhan Huoshenshan Hospital database were collected for this study.Inclusion and exclusion criteria were developed based on the clinical typing criteria of the COVID-19 Treatment Protocol(6th edition)combined with guidance from professional clinicians.For the purpose of this study,patients who were classified as mild or normal according to the treatment protocol were considered as mild cases,while patients who were severe or critical were considered as severe cases.For the purpose of this study,a COVID-19 patient is considered to have experienced a deterioration if the patient has experienced a change in condition from mild to severe.The main research objective of this study was to predict whether a deterioration would occur in COVID-19 patients.Ultimately,the experiment extracted 2,071 positive and 2,071 negative samples,and 82 medical indicators.Baseline statistics were performed on the extracted data in this study,and the results proved that the extracted data were in line with reality.The data are preprocessed,including random downsampling to balance the proportion of positive and negative samples and using the random forest method to fill in the gaps.The experiment applied the feature selection method combining the XGBoost algorithm and the sequence forward search method,and screened 15 key risk factors affecting the deterioration of patients with new coronary pneumonia from 82 medical indicators.Global interpretive analysis,personalized attribution analysis and feature interaction analysis were carried out on 15 key risk factors using SHAP method,and the early warning scope of risk factors was further proposed.Using the LIME method,local analysis was performed on two randomly selected positive samples and two negative samples.The analysis results of the local samples were consistent with the global analysis conclusions,which confirmed the reliability and rationality of the analysis conclusions.Finally,the research completed the construction of multi-objective neural network model.Taking the artificial neural network as the model,aiming at model prediction accuracy,model stability,and model complexity,the cellular genetic algorithm was used to optimize the hyperparameters,and a prediction model for the deterioration of patients with new coronary pneumonia was established.After multi-objective neural network learning,a set of Pareto optimal neural network models are obtained.Easy-to-understand diagnostic rules were extracted from the model.In this study,aiming at the question of whether patients with mild symptoms of new coronary pneumonia will deteriorate,combined with machine learning,cellular genetic algorithm and artificial neural network technology,a prediction model with strong interpretability was constructed,and 15 key risk factors were screened.The scope of early warning of key risk factors.The research conclusions can help patients reduce mortality,optimize treatment strategies,normal operation of the medical system,and optimize the allocation of medical resources.