Predicting patient enrollment in a telephone-based principal care management service using topic modeling

Area of Science:

• Ophthalmology
• Health Services Research
• Data Science

Background:

• Diabetic Retinopathy (DR) is a diabetes complication causing vision loss.
• Principal Care Management (PCM) services aim to reduce barriers for DR patients.
• Suboptimal enrollment hinders the effectiveness of DR care management programs.

Purpose of the Study:

• To develop predictive models identifying factors associated with patient enrollment in DR PCM services.
• To analyze call transcripts and metadata to understand enrollment drivers.
• To improve enrollment strategies for DR care management.

Main Methods:

• Feature-engineered call metadata (length, frequency, time intervals, sentiment).
• Extracted discussion topics using Structural Topic Modeling (STM).
• Developed and compared three classification models (metadata, topic-based, topic+metadata) to predict enrollment.

Main Results:

• The topic+metadata model showed superior performance (AUC 0.81-0.99).
• Proactively scheduling appointments post-explanation significantly increases enrollment odds.
• Longer call intervals and comprehensive script coverage positively correlate with enrollment.

Conclusions:

• Enrollment agent strategies, including proactive scheduling and allowing patient reflection, are critical for program uptake.
• Understanding patient conversation topics and metadata enhances predictive accuracy for enrollment.
• Optimizing call strategies can improve access to vital care coordination for DR patients.