Interpretable Machine Learning for COVID-19: An Empirical Study on Severity Prediction Task

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• The "black-box" nature of machine learning (ML) models impedes trust and deployment in critical medical applications like COVID-19 diagnosis.
• Interpreting ML models can reveal crucial biomarkers, aiding clinicians overwhelmed during pandemics.
• Identifying reliable biomarkers is essential for predicting COVID-19 severity and patient outcomes.

Purpose of the Study:

• To interpret machine learning models for identifying biomarkers associated with COVID-19 infection severity.
• To validate the identified biomarkers using independent datasets.
• To enhance the clinical utility of high-accuracy ML diagnostic tools.

Main Methods:

• Utilized four ML models: decision trees, random forests, gradient boosted trees, and neural networks.
• Employed various interpretation techniques: permutation feature importance, partial dependence plots, individual conditional expectation, accumulated local effects, LIME, and SHAP.
• Analyzed two datasets: 92 patients in Zhuhai, China, and 5644 patients from Kaggle (Hospital Israelita Albert Einstein).

Main Results:

• Identified increased N-terminal pro-brain natriuretic peptide, C-reaction protein, and lactic dehydrogenase, alongside decreased lymphocytes, as indicators of severe COVID-19 and mortality risk.
• Validated findings on a larger dataset, confirming the association with severe infection.
• Discovered leukocytes, eosinophils, and platelets as additional indicative biomarkers for COVID-19.

Conclusions:

• Model interpretation techniques effectively identify key biomarkers for COVID-19 severity prediction.
• Biomarkers such as N-terminal pro-brain natriuretic peptide, C-reaction protein, lactic dehydrogenase, and lymphocytes are crucial for assessing COVID-19 prognosis.
• Leukocytes, eosinophils, and platelets also serve as significant indicators for COVID-19, supporting the use of interpretable AI in clinical settings.