JMIR Publications

ABSTRACT

Background:

Amid growing demands and constrained healthcare resources, effective hospital bed capacity management is crucial. Delayed hospital discharge, where patients remain in the hospital beyond the need for acute care, strains resources, affects patient outcomes, and reduces system efficiency. Predicting such delays facilitates early interventions to avert them and alleviate burdens on patients, care partners, hospitals, and the broader healthcare system.

Objective:

This study aimed to develop comprehensive predictive analytics for delayed discharges using explainable machine learning to boost transparency and interpretability, while integrating fairness to mitigate algorithmic biases.

Methods:

Leveraging longitudinal data from over two decades in Ontario, Canada, we applied extreme gradient boosting (XGB) and logistic regression (LR) models to predict delayed discharges within 90 days post-acute care. Data preprocessing included a two-year look-back for clinical histories and balanced sampling to address class imbalance. Model performance was assessed via area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and F1-score. Fairness was evaluated across sex, urban/rural residence, and residential instability using metrics like equal opportunity and predictive equality. Explainability was examined globally (via partial dependence plots and permutation feature importance) and locally (via Shapley Additive Explanation, breakdown, and ceteris paribus methods), with principal component analysis clustering key features for high-risk patients.

Results:

The XGB model outperformed LR, achieving an AUC of 0.82 and sensitivity of 0.79 on the test set, with acceptable fairness (ratios 0.85-1.15) across subgroups. Explainability clustering analyses identified functional and cognitive declines (e.g., care support needs, dementia, mobility issues) and regional disparities as primary drivers of high-risk predictions. Bias mitigation improved equal opportunity for sex (from 1.10 to 0.99) but slightly reduced recall and worsened predictive parity (1.06 to 1.23), underscoring trade-offs between accuracy, fairness, and explainability that policymakers must weigh.

Conclusions:

This study demonstrates the potential of responsible artificial intelligence in healthcare, emphasizing the need to balance predictive accuracy, equity, and interpretability. It uncovers systemic gaps and offers actionable insights for enhanced discharge planning, resource optimization, and equitable care delivery.