Data Driven Predictions of Hospitalizations

In a new study published in the American Journal of Managed Care, researchers Morawski, Dvorkis, and Monsen use electronic health record (EHR) data and claims-derived data to predict whether individuals will be hospitalized.

Hospitalization Predictions

The ability to predict hospitalization despite diverse groups of patients from readily available data is an important contribution of this study and can help inform hospitals about how to allocate resources. The authors use data from 185,388 patients of Atrius Health collected between June 2013 and November 2015 and find that EHR data, claims data, and combined EHR and claims data nearly equally predict hospitalization.

Data Study

This study highlights the importance of using data for bettering medical care. Electronic data are constantly collected when a patient encounters the healthcare system through medical transcription and other means. While the data has historically been used to document the patient’s health, ensure all medical professionals are aware of the patient’s health, and report procedures and visits to billing, the changing technology with machine learning techniques allows medical professionals to use these data beyond the standard purposes. By utilizing machine learning through the use of logistic regression models, this study predicts a patient’s likelihood of being hospitalized within 6 months given the patient’s demographic characteristics, prior hospital visits, prescribed medications, and prior clinical diagnoses. The authors find that age and prior diagnoses are the most predictive factors for whether an individual will be hospitalized in the future.

While the model is highly predictive of hospitalizations…

It is more accurate for some groups than for others. The authors find that the model performs very well for those that are less likely to be hospitalized, but performs less well for those that have a predicted risk of hospitalization that is greater than 30%. The authors believe this is likely due to the small sample size of patients that have a hospitalization risk of greater than 30%.

The results of this study are limited by the use of only one health system. However, the authors believe that the approach used to modeling hospitalization risk can be used for other health systems with some calibration adjustments to the health system’s own patients and care.

Overall, this study’s approach to using machine learning with the abundance of EHR data-collected in part from medical transcription – and claims data is a meaningful step in predicting patient’s hospitalization risk. The authors suggest that this predictive tool can be used by clinicians to address patient concerns and health problems in a manner that can lessen the likelihood that the client will be hospitalized in the future.