MLHO

MLHO predicts patient-level risks of adverse COVID-19 outcomes, including hospitalization, ICU admission, mechanical ventilation, and mortality, by combining iterative sequential representation mining with feature and model selection.

Key Features:

• Iterative Feature and Algorithm Selection: Employs a dynamic iterative process to select features and statistical learning algorithms to optimize predictive accuracy for specified health outcomes.
• Sequential Representation Mining: Uses sequential representation mining to explore and refine patient data representations derived from historical medical records.
• Parallel and Outcome-Oriented Model Calibration: Supports parallel testing of multiple statistical learning algorithms and feature sets with outcome-oriented calibration for simultaneous evaluation and refinement across different outcomes.
• Use of Pre-COVID-19 Patient Data: Incorporates clinical and demographic data from pre-COVID-19 medical records with over 600 features representing pre-existing health conditions and demographics.

Scientific Applications:

• Risk Prediction for COVID-19 Adverse Outcomes: Predicts hospitalization, ICU admission, mechanical ventilation, and mortality in COVID-19-positive patients.
• Clinical Prioritization and Resource Allocation: Supports healthcare resource allocation and prioritization of interventions by identifying high-risk patients, with a reported mean AUC ROC of 0.91 for mortality prediction.
• Feature Importance and Epidemiologic Insights: Identifies clusters of influential pre-existing conditions and demographic predictors, including age and past medical history, to inform targeted preventive measures and vaccination strategies.

Methodology:

Analyzes a cohort of over 13,000 COVID-19-positive patients, models four adverse outcomes using extensive pre-COVID-19 feature sets, and applies iterative testing and calibration to identify clusters of influential features for each outcome.