Evaluating biomedical feature fusion on machine learning’s predictability and interpretability of COVID-19 severity types

Accurately differentiating severe from non-severe COVID-19 clinical types is critical for the healthcare system to optimize workflow. Current techniques lack the ability to accurately predict COVID-19 patients’ clinical type, especially as SARS-CoV-2 continues to mutate. We explore predictability and interpretability of multiple state-of-the-art machine learning (ML) techniques trained and tested under different biomedical data types and COVID-19 variants. Comprehensive patient-level data were collected from 362 patients (214 severe, 148 non-severe) with the original SARS-CoV-2 variant in 2020 and 1000 patients (500 severe, 500 non-severe) 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 (LR), random forest (RF), k-nearest neighbors (kNN), and support vector machines (SVM) were applied to build predictive 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. The fused (hybrid) characteristic modality yielded the highest mean area under the curve (AUC) achieving 0·915, while the biochemical modality alone and the clinical modality alone had AUCs of 0·862 and 0·818 respectively. All ML models performed similarly under different testing scenarios and were robust when cross-tested with original and Omicron variant patient data. Our models ranked elevated d-dimer (biochemical), elevated high sensitivity troponin I (biochemical), and age greater than 55 years (clinical) as the most predictive features of severe COVID-19. ML is a powerful tool for predicting severe COVID-19 based on comprehensive individual patient-level data. Further, ML models trained on the biochemical and clinical modalities together witness enhanced predictive power. The improved performance of these ML models when trained and cross-tested with Omicron variant data supports the robustness of ML as a tool for clinical decision support. U.S. Centers for Disease Control and Prevention (CDC) We searched the PubMed database for publications investigating the use of machine learning (ML) in predicting severe COVID-19 types using patient-level data. We found studies published from the beginning of the COVID-19 pandemic in 2020 up to February 2023 using keywords such as “severe COVID-19”, “SARS-CoV-2”, “multimodal”, “machine learning”, “prediction”, and “data-driven.” The resulting studies were overall limited in scope, as they focused on single data modalities or uninterpretable models. Nearly all studies found only used patient data obtained from the outbreak of COVID-19 and lacked data from the later variants, such as Omicron. These limitations prevent identification of the data modalities and ML techniques most suitable for predicting severe types, as well as the generalizability of these models to multiple variants. We built end-to-end machine learning pipelines with a variety of ML techniques, data modalities (biochemical, clinical, and fusion), and SARS-CoV-2 variants (original and Omicron) to compare the predictive power of each model type. Our study shows these models have strong predictive power severe COVID-19 when trained on multiple modalities and robustness across different variants of the virus, with two models achieving an AUC > 0·90. We compared feature rankings of models trained with the different variants and found overall agreement that the following features are highly predictive of severe COVID-19: elevated coagulation markers (d-dimer), indicators for heart damage (hsCRP, hsTNI), and age >55 years. These findings result from a thorough analysis of the effect of data type, ML technique, and SARS-CoV-2 variant on the power to predict severe COVID-19. To our knowledge, no other work has provided analysis of the effect of these characteristics, particularly the SARS-CoV-2 variants, on the performance of ML models. This model yields a powerful framework for healthcare providers seeking clinical decision support tools for not only COVID-19, but many other viral respiratory illnesses. Our work demonstrates a need for further testing with larger datasets to confirm the benefits of biomedical feature fusion.