Using Machine Learning to Identify Social Determinants of Health that Impact Discharge Disposition for Hospitalized Patients

Objective

To identify self-reported social determinants of health (SDOH) among hospitalized patients that predict discharge to a skilled nursing facility (SNF).

Design

A retrospective cohort analysis of 134,807 hospitalized patients from electronic medical records.

Setting and Participants

All patients admitted to hospitals within a large multistate tertiary health system.

Methods

The primary outcome was hospital disposition (home discharge vs SNF). The cohort was split into derivation and validation sets (75/25). We adopted 2 regularized regression-based statistical approaches, namely, the stacked elastic net (SENET) and bootstrap imputation-stability selection (BISS), to implement variable selection with incomplete data. After variable selection, logistic regression with the selected variables was conducted to create the final predictive model. The prediction accuracy and model fairness were evaluated on the test dataset using the area under the curve (AUC), equal AUC, and calibration.

Results

In the sample, 8.72% of patients were discharged to an SNF. The final models included between 11 and 15 variables. Significant SDOH variables included alcohol consumption, dental check, employment status, financial resources, nutrition, physical activities, social connection, and transportation needs. The final models also included 1 clinical (Charlson Comorbidity Index) and 2 demographic (marital status and education level) characteristics. The final models were confirmed across methods and datasets, predicted well in the validation cohort (AUC around 0.77), and were well calibrated.

Conclusions and Implications

Multiple SDOH characteristics predict SNF disposition, especially the lack of a life partner or spouse, are potentially mitigable (nutrition, physical activities, and transportation needs), and offer actionable targets to increase home discharge rates. The collection and integration of SDOH data may optimize the appropriateness and efficiency discharge planning.