Predictors of hospital-at-home feasibility. Data-driven insights

Abstract

The hospital-at-home (HaH) model enables patients to receive specialized care at home as an alternative to inpatient treatment while supporting a structured transition back to routine outpatient care. In this pilot study, we evaluated a telemedicine supported transition from hospital to home with subsequent handover to the primary general practitioner, with the aim of identifying which early routinely collected signals are most informative for assessing HaH feasibility. We analyzed data from 234 patients enrolled in a Swiss HaH pilot program, defining regular versus early study termination as the primary outcome (early termination defined a priori as rehospitalization, loss to follow-up, or consent withdrawal, reflecting programlevel feasibility). Candidate variables included demographic and clinical enrollment characteristics, physiological measurements, and early patient- and clinician-reported measures collected within a predefined day 0-1 assessment window. Machine learning models were applied within a stratified cross-validation framework to compare feature sets and to support systematic feature relevance analysis under class imbalance. Models based solely on demographic and routine clinical enrollment data showed limited discriminative ability, with balanced accuracy values close to chance level (up to 0.58). In contrast, models incorporating early patient- and clinician-reported measures achieved markedly improved performance, with balanced accuracies up to 0.79 and precision-recall AUCs approaching 0.90, alongside strong recall for early termination. Feature importance and SHAP analyses confirmed that completion and results of early patient- and clinician-reported measures were more informative for HaH completion than baseline demographic, physiological, or routine clinical characteristics. Overall performance was limited by sample size, and the outcome represents a pragmatic proxy for successful HaH transition. Still, the findings indicate that early routinely collected operational and telemedical assessments were associated with HaH feasibility, with machine learning serving as an analytical tool despite modest predictive gains, and should be interpreted as predictive rather than causal.