Data Mining to Guide a Program to Prevent Infection Related Readmissions From Skilled Nursing Facilities

Abstract

Background: Readmissions to hospitals are common, costly and often preventable, notably readmissions due to infections. A 30-day readmission analysis following hospital discharges, found much of the variation in Medicare spending between hospitals was related to readmissions and skilled nursing facility (SNF) care. Although some readmissions of patients with advanced disease are not preventable, efforts to decrease readmission are most effectively directed towards those patients with intermediate levels of a specific risk. A prediction model to identify patients at highest (or intermediate) risk of infection readmission will help healthcare administrators and providers to allocate appropriate resources. Hospitals should use electronic health record (EHR) data with modern data mining techniques to create more curated, sophisticated models as part of a comprehensive transitional care program. We propose using the risk estimates of a validated prediction model to notify stakeholders and develop readmission rate reports by SNF or discharging physician. Methods: We applied machine learning (ML) methods to predict the risk of 30-day readmission due to sepsis and pneumonia of patients discharged to SNF. We used our EHR data during 2012–2017 to train and data from 2018 to validate. We applied ML algorithms to data including logistic regression, random forest, gradient boosting trees, and support vector machine. Data from EDW and EPIC clarity tables were extracted and managed using SAS Base 9.4 and Enterprise Miner 14.3 (SAS Institute, Cary, NC). We assessed the discrimination and calibration to select the most effective prediction model. Using the resulted risk estimates, we created a notification system and reports for key stakeholders. Results: Figures 1 and 2 show the discrimination and calibration results of the final selected gradient boosting model (GBM). For predicting unplanned readmissions with sepsis and with pneumonia within 30 days after discharge to SNF, the c-statistic for final GBM model with 140 features was 0.69 (95% CI 0.65-0.73) and 73 features was 0.71 (95% CI 0.66-0.75), respectively. Table 1 lists features important to the validation set of the prediction model. We used estimates from these models to develop a daily email notification of patients discharged to SNF stratified into a low, medium, and high risk group for sepsis and pneumonia. We additionally created reports with case-mix adjustments to benchmark SNFs and discharging physicians to monitor and understand performance. Conclusions: Hospitals should leverage the plethora of data found in EHRs to curate readmission prediction models, and promote collaboration among transitional care teams and IPC to ultimately reduce readmissions due to sepsis and pneumonia.Funding: NoneDisclosures: None