Derivation and Internal Validation of a Mortality Prognostication Machine Learning Model in Ebola Virus Disease Using Iterative Point-of-Care Biomarkers

Although multiple prognostic models for Ebola Virus Disease (EVD) mortality exist, few incorporate biomarkers and none has used longitudinal point-of-care (POC) serum testing throughout Ebola Treatment Center (ETC) care. This retrospective study evaluated adult EVD patients during the 10th outbreak in the Democratic Republic of Congo. Ebola virus RT-PCR cycle threshold (Ct) and POC serum biomarker data were collected throughout ETC treatment. Four iterative prognostic mortality machine learning (ML) models were created using elastic net regularization. The base model used admission age and Ct as predictors. Ct and biomarkers collected on treatment days one and two (D1,2), days three and four (D3,4) and days five and six (D5,6) that were associated with mortality, were iteratively added to the preceding model to yield temporally dynamic mortality risk estimates. Receiver operating characteristic (ROC) curves were created for each model iteration providing time period specific Area Under Curve (AUC) with 95% confidence intervals (CIs). Of 310 EVD-positive cases, mortality occurred in 46·5% of cases. Biomarkers predictive of mortality were elevated creatinine kinase, aspartate aminotransferase, blood urea nitrogen (BUN), alanine aminotransferase, and potassium, and low albumin during D1,2, elevated c-reactive protein (CRP), BUN, and potassium during D3,4, and elevated CRP and BUN during D5,6. The AUC substantially improved with each iteration: base model AUC 0·74 (95% CI 0·69 – 0·80), D1,2 model AUC 0·84 (95% CI 0·73 – 0·94), D3,4 AUC 0·94 (95% CI 0·88 – 1·0), and D5,6 AUC 0·96 (95% CI 0·90 – 1·0).