Clinical Scorecard: Predictive Modeling of High Utilization Patients Using Ensemble Machine Learning: Validation and Economic Impact Assessment

At a Glance

Key Highlights

• Boosted tree ensemble models achieved highest predictive performance for inpatient length of stay and emergency department visits.
• Models identified 77% of future inpatient users and 73.9% of future emergency department users correctly in validation.
• Economic impact analysis showed potential cost savings of SGD $152 million (US $111 million) using the best-performing model.

Guideline-Based Recommendations

Diagnosis

• Use multiclass ensemble machine learning models to stratify patients by predicted health care usage levels (LOS and ED visits).

Management

• Apply predictive models to support targeted interventions for high-utilization patients within diabetes population health programs.

Monitoring & Follow-up

• Validate and update predictive models regularly using temporal datasets to maintain accuracy and relevance.

Risks

• Recognize heterogeneity in patient characteristics and usage patterns that may affect prediction accuracy and intervention effectiveness.
• Avoid focusing solely on highest users to prevent overlooking patients with emerging health care needs.

Patient & Prescribing Data

Patients with diabetes registered in a large clinical registry in Singapore

Predictive modeling can guide resource allocation and intervention planning to reduce inpatient and emergency department utilization and associated costs.

Clinical Best Practices

• Incorporate ensemble machine learning predictions into clinical workflows for proactive identification of high-utilization patients.
• Use predicted probabilities from base learner models as inputs for ensemble models to improve multiclass classification.
• Perform economic impact assessments alongside predictive validation to quantify potential cost savings.
• Integrate predictive models within population health initiatives targeting diabetes care to optimize planning and budgeting.

References

• JMIR Medical Informatics, 2026;14:e77202