With an aim to improve management of 30-day readmission rates and associated high costs of cardiac care, Edj’s custom readmission model helps hospital systems and care providers predict which cardiac patients are most likely to be readmitted and manage factors that can reduce readmission rates.
A healthcare system serving a tristate area engaged EdjAnalytics to analyze their Congestive Heart Failure (CHF) patients covered by Medicare. Prior to this project, the hospital system had been penalized through the Hospital Readmissions Reduction Program for a heightened number of readmissions among CHF patients. The goal of this analysis was to anticipate and reduce the number of readmissions within 30 days of hospital discharge.
Edj received three years of data on Medicare patients with CHF who had hospital stays within the hospital system’s network. Variables provided for the analysis included the Hospital Specific Report compiled by Medicare indicating unplanned readmissions, diagnoses, procedures, labs, vitals, discharge disposition and demographics. The Edj team of data scientists tested multiple advanced modeling techniques to score readmission risk and narrowed down to a final methodology given the system’s desire to integrate the final model into their Electronic Medical Records system. The methodology selected places foremost importance on interpretability, which is crucial in a clinical setting.
Edj revealed that the following variables were most important in predicting readmission risk:
- The type of facility where a patient is discharged
- Number of previous encounters where CHF was listed as a co-morbidity
- Number of times the patient visited the hospital in the previous 90 days
- Potassium, Hemoglobin and Creatinine levels
As is often the case in predictive modeling, these findings both supported conventional wisdom and shed light on new factors that were not given prior consideration. While the project validated that vitals are important, Edj also revealed that discharge plans and visit history are, in fact, more predictive of a patient’s readmission risk.
The final model achieved an AUC of 0.7 on unseen data, which indicates that the model is highly accurate in predicting patient readmission risk and outperforms other published research models designed to predict similar outcomes.
The model enables healthcare providers to assess readmission risk for each of their Medicare CHF patients through their Electronic Medical Records system and adjust discharge plans accordingly. The visual to the right showcases what the system displays for a sample patient with high risk of readmission.