Reinforced risk prediction with budget constraint using irregularly measured data from electronic health records

Area of Science:

• Biomedical Informatics
• Health Services Research
• Predictive Analytics

Background:

• Uncontrolled glycated hemoglobin (HbA1c) in complex diabetic patients leads to adverse events, posing significant health risks and financial burdens.
• Current risk prediction methods may require costly and burdensome biomarker information.
• There is a need for accurate, cost-effective predictive models to identify high-risk diabetic patients for preventative care.

Purpose of the Study:

• To develop and validate a sequential predictive model for classifying complex diabetic patients into high-risk, low-risk, or uncertain categories.
• To optimize information collection by utilizing accumulating longitudinal data for accurate risk prediction.
• To improve patient outcomes and reduce healthcare costs through timely, targeted interventions.

Main Methods:

• A sequential predictive model was developed using longitudinal data from Medicare claims, enrollment files, and Electronic Health Records (EHR).
• Functional principal components analysis was employed to handle noisy longitudinal data.
• Weighting techniques were used to address missing data and sampling bias.

Main Results:

• The proposed sequential model demonstrated higher predictive accuracy compared to competing methods.
• The model achieved lower costs in simulation experiments and real-world data application.
• The approach effectively classifies patients, guiding recommendations for preventative treatment or standard care.

Conclusions:

• The sequential predictive model offers a cost-effective and accurate approach to identifying high-risk diabetic patients.
• This method has the potential to enhance patient care by enabling early, personalized interventions.
• Optimizing data utilization in predictive modeling can lead to significant improvements in healthcare efficiency and patient outcomes.