Predicting 30-Day Hospital Readmission in Medicare Patients: Insights from an LSTM Deep Learning Model

Abstract

Background Readmissions among Medicare beneficiaries are a major problem for the US healthcare system from a perspective of both healthcare operations and patient caregiving outcomes. The Centers for Medicare & Medicaid Services (CMS) implemented the Hospital Readmissions Reduction Program (HRRP) to tackle the problem of readmission by penalizing to hospitals with excessive 30-day readmissions. Our study is to analyze hospital readmissions among Medicare patients and provide some suggestions of readmission prevention furthermore. Using LSTM networks with feature engineering, our proposed deep learning model would help us understand the contribution of the features. Design The MIMIC-III clinical database contains information from 53,423 patient admissions at Beth Israel Deaconess Medical Center. We restricted our analysis to 21,002 patient admissions with Medicare coverage, and selected variables from admission-level data, inpatient medical history and patient demography. The baseline model is a logistic-regression model based on the LACE index, and the LSTM model was another model that designed to capture temporal dynamic in the data from admission-level and patient-level data. We used Area Under the Curve (AUC) metric to evaluate the model's performance and leveraged the precision and recall to evaluate how the two models perform in predicting top 10% high-risk decile patients. Results The LSTM model outperformed the logistic regression baseline, accurately leveraging temporal features (time-series data) to predict readmission. The major features used by the model were the Charlson Comorbidity Index, hospital length of stay, the hospital admissions over the past 6 months or the number of medications before discharge, while demographic variables were less impactful Limitations The use of a single-center database from the MIMIC-III database further limits the generalizability of the findings: patients from this database might not fully represent the wider Medicare population. Additionally, the focus on all-cause hospital readmissions without accounting for specific chronic conditions, such as heart failure or diabetes, limits the model's ability to capture the complexities of chronic diseases known to impact readmission rates in older adults. Lastly, the exclusion of external factors, such as environmental quality, proximity to healthcare facilities, and patient behaviors, further constrains the LSTM's predictive accuracy. Conclusions This work suggests that LSTM networks will be a more promising approach to aid efforts to predict the readmission of Medicare patients. Capturing temporal interactions in patient databases can help us arrive at a more nuanced understanding of potential hospital readmission and help healthcare providers make their prediction models better than current approaches do. Implications Adoption of such predictive models into clinical practice may be more effective in identifying high-risk patients to provide earlier and targeted interventions to reduce readmission rates and improve patient outcomes. Future studies need to be conducted to further justify these findings in other patient populations and explore ways to enhance the model's interpretability for use in clinical scenarios.