Toward a Clinically Actionable, Electronic Health Record-Based Machine Learning Model to Forecast 90-Day Change in...

Area of Science:

• Pediatric Endocrinology
• Data Science in Healthcare
• Clinical Informatics

Background:

• Clinicians need tools to identify youth with type 1 diabetes (T1D) at risk of worsening glycemic control between visits.
• Current methods are insufficient for proactive intervention, limiting care for rising HbA1c levels.

Purpose of the Study:

• To assess the feasibility of using electronic health record (EHR) data to build a machine learning (ML) model.
• To predict 90-day changes in HbA1c in youth (9-18 years) with T1D.
• To enhance clinical decision-making by identifying youth with clinically significant HbA1c increases.

Main Methods:

• Utilized longitudinal EHR data from 1743 youth with T1D (2012-2017).
• Transformed over 17,000 features from EHR data for model training.
• Employed random forest regression with 3-fold cross-validation to predict HbA1c unit-change.

Main Results:

• The ML model demonstrated strong correlation between predicted and actual HbA1c (r=0.79).
• Key predictors included postal code, HbA1c metrics, and treatment engagement difficulty.
• At a ≥0.3% HbA1c rise threshold, the model achieved 60.3% positive predictive value, a 1.5-fold enrichment.

Conclusions:

• Routinely collected EHR data can effectively train ML models to predict HbA1c changes in youth with T1D.
• This predictive capability can aid in timely interventions to manage glycemic control.
• Future research will focus on model optimization and validation in diverse cohorts.