OBJECTIVES: Use advanced Machine Learning (ML) methodology/techniques to identify the likelihood of hospitalization for high-risk patients diagnosed with COVID 19.

METHODS: We leveraged a de-identified US healthcare claims patient-level database for this study. Patients who have a ICD 10 diagnosis code for COVID 19 (Index Event, IE) in Q4 2020 and having a history of high risk conditions (12 months prior to IE) were included in the sample selection, patients hospitalized were labeled as Targets and the remaining patients were assigned as Controls. The claims data included diagnosis, procedural, surgical information and treatment prescribed to patients. The model training data set is a high dimensional dataset ( 20,000 + variables) and hence we used Boruta for feature elimination/selection and reduced dimensionality. For the final model, Random Forest and eXtreme Gradient Boosting were ensembled assigning appropriate weights to each model.

RESULTS: The model training exercise included a sample of 7,000 targets and 7,000 control and the models were applied on an unseen validation sample of 3000 controls and 3000 targets. The ensemble model was evaluated using various model performance metrics like ROC plot, precision, accuracy and recall, the ensemble model achieved a sensitivity of 76% ( 76% of COVID 19 patients hospitalized were predicted prior to hospitalization) and accuracy of 60%. The final model consist of ~200 important predictors variables such as age, diagnosis like Type 2 diabetes/CKD/ Hypertension , frequency of office visits, Obesity amongst others. The models results can be further improved by engineering newer features , integrating additional datasets and further tunning the models.

CONCLUSIONS: Embedding such models within the point of care facility (EMR or other systems ) can help in identifying cases of hospitalization of COVID , target provider messaging to enable timely care/services/treatments and hence reducing the mortality and burden of the disease.