Interpretable Machine Learning to support early planning of hospital Length Of Stay

Background

Anticipating hospital length of stay is essential for discharge planning, bed management and continuity of care. A useful AI system should not only predict prolonged stay, but also provide understandable evidence supporting that prediction. MARIO was developed as a machine learning framework to support care planning by transforming discharge records into clinically interpretable patient representations.

Methods

We built an anonymized patient-centric dataset of more than 60,000 hospital episodes. Each admission was represented through named features combining three evidence sources: previous hospital use and historical length of stay; administrative and organizational admission context; and early clinical information, including primary diagnosis and procedures recorded shortly after admission. These representations were used to train neural-based classifiers for prolonged stay prediction. The binary task distinguished short from prolonged admissions using a clinically revised threshold of 10 days. Linear models also enabled explanation by ranking the features contributing most to each decision.

Results

The best-performing binary models reached about 82% accuracy. Error analysis showed that many mistakes occurred close to the decision boundary, where differences of one or two days may be clinically and organizationally less relevant. When borderline cases around the threshold were excluded, accuracy approached 90%. Model outputs could also be inspected through feature-level explanations, showing which historical, organizational or early clinical signals drove predictions.

Conclusions

MARIO suggests that interpretable machine learning on routine hospital data can support early planning of hospital stay without replacing clinical judgement. By combining prediction with explainability, the system may help administrators and clinicians understand risk signals, prioritize discharge planning and make more informed decisions within transition pathways.

Key messages

• Interpretable AI using routine hospital data accurately predicts prolonged stays, providing clear clinical evidence to support early discharge planning.

• Combining predictive accuracy with feature-level explanations helps clinicians prioritize bed management and transition pathways without replacing human judgment.