Introduction
A machine learning algorithm has been developed using a combination of:
- Real time accessed patient reported symptoms and vital signs
- CDW derived medication lists
- Admission history
- Lab values
- Historical conditions
- Historical vital signs
- Vaccine history
This algorithm has an AUC of 0.89 for predicting hospital admission within 30 days of COVID-19 test collection determined at the time of test result. This is a dynamic risk score that benefits from live data because with the update of labs and vital signs, the risk of admission can change for the patient.
Additionally, self reported vital signs and symptoms from home can change the model output which arrive in the health record at unpredictable intervals. The model output is used to guide clinical care by identifying patients at high risk of clinical worsening and directing clinical interventions such as monoclonal antibody treatment, antiviral therapies, etc.
For other disease states the intervention could be an in person assessment for volume status and imaging or a medication change. With the advent of remote monitoring devices such as continuous glucose monitors, access to a steady stream of patient glucose values coupled with our CDW data could be used to predict hospital admission for DKA or severe hypoglycemia conditions.