Data Analytics and Machine Learning Models on COVID-19 Medical Reports Enhanced with XAI for Usability

Area of Science:

• Medical Informatics
• Data Science
• Machine Learning

Background:

• Effective data analytics and machine learning (ML) solutions are crucial for medical decision-making and understanding COVID-19.
• Analyzing electronic medical records (EMRs) can provide valuable insights into disease progression and patient outcomes.

Purpose of the Study:

• To identify effective data analytics and ML solutions for medical decision-making in the context of COVID-19.
• To analyze factors influencing COVID-19 disease severity and mortality risk.
• To develop and validate predictive models for COVID-19 severity and mortality.

Main Methods:

• Utilized various ML models including LightGBM, XGBoost, random forest, SVM, and logistic regression on EMR data from 4711 COVID-19 patients.
• Employed statistical methods, LASSO regression, and SHAP values for feature selection and model explainability.
• Evaluated model performance using accuracy, F1-score, and ROC AUC. Validated mortality prediction models on an external cohort.

Main Results:

• The LightGBM model demonstrated superior performance in predicting COVID-19 severity (89.5% accuracy with top features) and mortality risk (83.7% ROC AUC).
• Key factors influencing severity and mortality were identified through feature importance analysis.
• Web-based applications integrating simplified predictive models were developed and tested by medical experts, achieving a 73.3% overall prediction accuracy.

Conclusions:

• Machine learning, especially LightGBM, offers effective solutions for predicting COVID-19 severity and mortality risk.
• The developed predictive models and web application can support clinical decision-making and enhance the understanding of COVID-19.
• Explainable AI methods provide insights into the factors driving disease outcomes.