In a retrospective study, Pandata partnered with a cancer hospital to develop an oncology-specific predictive model to identify patients who are at the highest risk of an unplanned 30-day readmission at the point of discharge.
We designed a model that identifies 2 out of 3 high risk cancer patients likely to experience an unplanned readmission by applying data preprocessing, feature engineering and machine learning to 45,500 patient encounters and 100+ clinical and socioeconomic variables. This solution offers personalized predictions summarizing the top clinical and social determinants of health (SDoH) factors that impact risk scores for each patient. Explainable AI allows frontline users to understand model results and identify the factors that contributed to each prediction, solving for the “black box” problem in healthcare.
A high unplanned readmissions rate is a quality metric for hospitals that can indicate increased costs, poor quality of care and decreased patient satisfaction. In recent years, value-based Medicare programs have penalized hospitals for excess readmissions. While much work has gone into developing predictive risk models for the surgery population, these tools do not extend well to the cancer population. This study analyzed the top features impacting readmission risk for patients at this cancer center and provided individualized insights on social and clinical factors that contribute to the patient’s risk. In the future, predictions can be integrated into case management workflows to create targeted patient follow-up and improve discharge planning, which would translate to estimated cost savings of $2.2 million annually.