The Centers for Medicare and Medicaid Innovation (CMMI) Bundled Payments for Care Improvement Advanced (BPCI-A) initiative aims to promote value-based delivery of patient care. (1) We sought to test the hypothesis that by (a) integrating clinical and socioeconomic patient characteristics and by (b) employing machine learning tools, we would develop predictive models for the outcomes of inpatient charges for each index admission and 90-day readmissions. We studied patients admitted to a large academic tertiary care center with an index hospitalization due to heart failure (HF), myocardial infarction (AMI) without coronary intervention, and arrhythmias collectively described as Cardiac Care Bundle patients.
Model 1 (prediction of charges for index admission) comprised clinical variables, namely history of cancer, discharge disposition, and Better Outcomes by Optimizing Safe Transitions (BOOST) 8 P’s scores. (2) Model 2 (prediction of 90-day readmission) comprised clinical and demographic variables, namely black race, discharge disposition, and BOOST 8 P’s checklist.
The goal of this project is to understand predictors of (a) inpatient charges incurred during index admission and (b) 90-day readmissions for each episode of care of patients under the Cardiac Care Bundle. Insights into predictors of these financial and clinical outcomes may optimize value-based care for our patients.
From August 2020 to August 2021, a total of 460 patients in a Cardiac Care Bundle, as defined above, were retrospectively studied. Clinical, demographic and socioeconomic variables were extracted using a combination of proprietary natural language processing (NLP) system and by traditional chart review.
A multivariate logistic regression and log-normal model were created using a causal discovery algorithm. Model 1 selected variables that predict index admission charges. Model 2 selected variables that estimate the likelihood for 90-day readmission. Odds ratios were calculated for both models.
One important obstacle is that our data only reflects patient care at our site and does not capture readmissions outside our health system. Our models are short, with clinically relevant information and can be used to predict charges for index admissions and 90-day readmissions in patients hospitalized with heart disease.