WHEN DOES MACHINE LEARNING IMPROVE READMISSION PREVENTION IN OLDER ADULTS? INTEGRATING PREDICTIVE PERFORMANCE, CLINICAL EFFECTIVENESS, AND DECISION MODELLING
Author(s)
Somayeh Ghazalbash, PhD, Manaf Zargoush, PhD;
McMaster University, DeGroote School of Business-Health Policy and Management, Hamilton, ON, Canada
McMaster University, DeGroote School of Business-Health Policy and Management, Hamilton, ON, Canada
OBJECTIVES: Hospital readmissions pose ongoing challenges for quality and cost. However, preventive transitional care interventions are limited in capacity and cannot be offered on a wide scale. We examined the effectiveness of machine learning (ML) in preventing readmissions by comparing ML targeting with traditional heuristic methods under budget constraints.
METHODS: We conducted a retrospective cohort study on residents aged 65 and older, focusing on unplanned 30-day readmissions. An eXtreme Gradient Boosting (XGB) model was developed and evaluated, with predictions calibrated using various methods. These predictions were then applied within a budget-constrained decision framework to compare different prioritization strategies: two distinct ML-based policies, LACE, cost-only, and the Random strategy. Policies were assessed for targeting precision and economic efficiency.
RESULTS: The cohort included more than 3 million eligible admissions, of whom approximately 20% were readmitted. XGB achieved an AUC of 0.70 on the test set, showing notable calibration improvements from isotonic regression. In decision analyses, ML-driven policies outperformed heuristic methods, with performance depending on the level of intervention coverage. At low coverage levels, ML policies varied in their balance between precision and economic efficiency, whereas at moderate coverage, their performance converged. As coverage increased, policy differences narrowed, suggesting diminishing returns from more complex targeting. Sensitivity analyses showed that the value of ML-based policies was most affected by assumptions about intervention effectiveness, with economically guided approaches offering greater efficiency gains.
CONCLUSIONS: ML-based policies provide the greatest value in decision support when intervention resources are scarce. The choice of targeting policy should align with program objectives, such as prioritizing reach or economic efficiency, while simpler heuristic approaches may be adequate when coverage is broad. This framework helps decision-makers align targeting policies with real-world constraints.
METHODS: We conducted a retrospective cohort study on residents aged 65 and older, focusing on unplanned 30-day readmissions. An eXtreme Gradient Boosting (XGB) model was developed and evaluated, with predictions calibrated using various methods. These predictions were then applied within a budget-constrained decision framework to compare different prioritization strategies: two distinct ML-based policies, LACE, cost-only, and the Random strategy. Policies were assessed for targeting precision and economic efficiency.
RESULTS: The cohort included more than 3 million eligible admissions, of whom approximately 20% were readmitted. XGB achieved an AUC of 0.70 on the test set, showing notable calibration improvements from isotonic regression. In decision analyses, ML-driven policies outperformed heuristic methods, with performance depending on the level of intervention coverage. At low coverage levels, ML policies varied in their balance between precision and economic efficiency, whereas at moderate coverage, their performance converged. As coverage increased, policy differences narrowed, suggesting diminishing returns from more complex targeting. Sensitivity analyses showed that the value of ML-based policies was most affected by assumptions about intervention effectiveness, with economically guided approaches offering greater efficiency gains.
CONCLUSIONS: ML-based policies provide the greatest value in decision support when intervention resources are scarce. The choice of targeting policy should align with program objectives, such as prioritizing reach or economic efficiency, while simpler heuristic approaches may be adequate when coverage is broad. This framework helps decision-makers align targeting policies with real-world constraints.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
CO159
Topic
Clinical Outcomes
Topic Subcategory
Comparative Effectiveness or Efficacy
Disease
No Additional Disease & Conditions/Specialized Treatment Areas