A Machine Learning Pipeline for Accurate COVID-19 Health Outcome Prediction using Longitudinal Electronic Health...

Current COVID-19 predictive models primarily focus on predicting the risk of mortality, and rely on COVID-19 specific medical data such as chest imaging after COVID-19 diagnosis. In this project, we developed an innovative supervised machine learning pipeline using longitudinal Electronic Health Records (EHR) to accurately predict COVID-19 related health outcomes including mortality, ventilation, days in hospital or ICU. In particular, we developed unique and effective data processing algorithms, including data cleaning, initial feature screening, vector representation. Then we trained models using state-of-the-art machine learning strategies combined with different parameter settings. Based on routinely collected EHR, our machine learning pipeline not only consistently outperformed those developed by other research groups using the same set of data, but also achieved similar accuracy as those trained on medical data that were only available after COVID-19 diagnosis. In addition, top risk factors for COVID-19 were identified, and are consistent with epidemiologic findings.