Skip to main content
Skip main navigationClose Drawer MenuOpen Drawer Menu
Sections

Background

Readmissions are costly, and often preventable. Very few published models have a C-statistics of >0.8 for readmission. We used advanced machine learning techniques to develop a high performing predictive model for readmissions.

Methods

The OM1™ Cardiology data warehouse contains deep clinical and claims data from cardiology practices across the US. This analysis included a subset of 142,434 patients who a) had linked EMR and claims data; b) were hospitalized between 1/2014 and 8/2017; and c) had at least 12 months of data before the index admission, and 30 days post discharge. The unit of analysis was hospitalization. The outcome was 30-day all-cause unplanned readmission per the CMS definition, or death. The gradient boosted machine learning model was built on a large number of predictive features including the OM1 medical burden index (OM1 MBI), which is a standardized measure of the combined effect of current and prior conditions and treatments on current health status, generated from extensive analysis of OM1's longitudinal patient cohort (n>175M). A random sample of 70% hospitalizations were used for training and the remaining 30% were for validation.

Results

This study included 353,801 index admissions with 76,845 (21.7%) unplanned readmissions or deaths within 30 days of discharge; median age was 56 years. In the validation set, the model correctly identified readmission or death status in at least 9 out of 10 patients (C-statistic=0.90). OM1 MBI was among strongest predictive features; other predictors included care setting, acuity of illness and measures of resource utilization. When precision was fixed at 53% and 75%, the OM1 model correctly identified 19,636 (85%) and 17,220 (74%) of the total 23,082 readmissions or deaths in the validation set, respectively.

Conclusion

Our model generated by advanced machine learning demonstrates superior performance to previously published predictive models, e.g., the established LACE+ index (C-statistic 0.76). This model may be more generalizable as it was built on data from multiple providers and EMR. Integrating these predictive models into clinical workflow will permit timely interventions in high risk patients.

Footnotes

Moderated Poster Contributions

Prevention Moderated Poster Theater, Poster Hall, Hall A/B

Sunday, March 11, 2018, 9:45 a.m.-9:55 a.m.

Session Title: New Insights Into Opportunities For Prevention of CVD Events

Abstract Category: 32. Prevention: Clinical

Presentation Number: 1223M-03