Interpretable Feature Selection In The Analyses Of COVID-19 And Postoperative Analgesia

The main purpose of medical data analysis is to find relevant factors for disease prediction and provide clinical evidence for diagnosis.With the development of machine learning and the advancement of medical information construction,more and more researches apply machine learning methods to the medical field in order to find several key features from high-dimensional nonlinear features,that is,medical related factors.Feature selection can focus prediction on principal features and help models find relevant factors.Due to the poor quality of medical data,few samples,and high feature dimensionality,existing machine learning and feature selection methods are difficult to find key features and to be accepted by the medical field.To solve this problem,a voting iterative learning framework driven by medical knowledge(VILF-MK)is proposed.The framework is divided into three stages.The first stage is single-feature multi-model selection,that is,use SHAP to measure and sort the contribution of a single feature to different models.Considering medical knowledge,the top features are selected from the sorting result based on voting.The second stage is multi-feature single-model verification,which is to verify the selected feature subset by evaluating the classifier.The process of feature selection and verification can be iterated to improve the performance of the final model.The third stage is multi-view verification,that is,the features of the selected subsets are analyzed from multiple views to verify their clinical significance.The framework has been applied in the analyses of COVID-19 and postoperative analgesia.The performance of the classifier after feature selection is close to the model on the feature set.Compared with other feature selection methods,the classifier trained by this framework has the highest AUC and BAC.And in early warning of severe COVID-19,6key features consistent with evidence-based medicine and 3 potential related features were found.In the analysis of postoperative analgesia,6 key features consistent with evidencebased medicine and 3 potential related features were found respectively.