Evaluating Biomedical Feature Fusion on Machine Learning’s Predictability and Interpretability of COVID-19 Severity Types: Model Development, Interpretation, and Validation

Background: Accurately differentiating severe from nonsevere COVID-19 clinical types is critical for the health care system to optimize workflow. Current techniques lack the ability to accurately classify COVID-19 clinical types in patients, especially as SARS-CoV-2 continues to mutate. Objective: We explore the predictability and interpretability of multiple state-of-the-art machine learning (ML) techniques trained and tested under different biomedical data types and SARS-CoV-2 variants. Methods: Comprehensive patient-level data were collected from 362 patients (severe COVID-19: n=148; nonsevere COVID-19: n=214) infected with the original SARS-CoV-2 strain in 2020 and 1000 patients (severe COVID-19: n=500; nonsevere COVID-19: n=500) infected with the Omicron variant in 2022-2023. The data included 26 biochemical features from blood testing and 26 clinical features from patients' clinical characteristics and medical history. Different ML techniques, including penalized logistic regression, random forest, k-nearest neighbors, and support vector machines, were applied to build predictive classification models based on each data modality separately and together for each variant. Fifty randomized train-test splits were conducted per scenario, and performance results were recorded. Results: The fusion (hybrid) characteristic modality yielded the highest mean area under the curve (AUC) in this study, achieving 0.915, while the biochemical and clinical modalities had AUCs of 0.862 and 0.818, respectively. All ML models performed similarly under different testing scenarios and were consistent when cross-tested with data of patients infected with the original strain and those infected with the Omicron variant. Our models ranked elevated d-dimer (biochemical), elevated high sensitivity troponin I (biochemical), and age greater than 55 years (clinical) as the most positively predictive features of severe COVID-19. Conclusions: These results are compatible with the hypothesis that ML is a useful tool for predicting severe COVID-19 based on comprehensive individual patient-level data. Further, ML models trained on the biochemical and clinical modalities together show patterns consistent with enhanced predictive performance. The improved performance observed with Omicron variant data agrees with the hypothesis that ML approaches may retain utility across variants in this study setting, although further validation is required before clinical application. Future work using larger datasets with more ethnic variation and investigating unbiased ML interpretation methods may be able to provide further validation.