JMIR Publications

ABSTRACT

Background:

Robust and accurate prediction of severity for patients with COVID-19 is crucial for patient triaging decisions. Current prognostic models rarely have both high discrimination and clinical interpretability.

Objective:

We developed and validated a machine learning (ML)-based Robust and Interpretable Early Triaging Support system (RIETS) that predicts severity based on routinely available biomarkers upon hospitalization.

Methods:

We used 5,945 hospitalized patients with COVID-19 from 19 general and tertiary care hospitals in South Korea collected between January 2020 and August 2022. We conducted a comprehensive search of all possible models from candidate feature subsets and six ML algorithms to discover RIETS. The Shapley method and patient clustering were used for model interpretability. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC). Bias risk was assessed using Prediction model Risk Of Bias ASsessment Tool.

Results:

RIETS, based on a deep neural network of 11 readily available biomarkers, had best predictive discrimination (AUROC, 0.938; 95% CI, 0.935-0.938]) with high calibration (integrated calibration index [ICI], 0.041) compared to existing low bias risk prognostic models (AUROC=0.60-0.80). Moreover, it had sustainable prediction on Omicron cases despite limited availability (AUROC=0.903 [0.897-0.910]); was rated as a low bias risk model; and showed substantial interpretability through patient clustering and characterization.

Conclusions:

RIETS had robust ability to predict severity of hospitalized patients with COVID-19, along with substantial interpretability. These findings suggest that incorporating RIETS in routine clinical practice can support efficient and effective early patient triaging.