Early Attrition Prediction for Web-Based Interpretation Bias Modification to Reduce Anxious Thinking: A Machine...

Area of Science:

• Digital mental health
• Computational psychiatry
• Human-computer interaction

Background:

• Digital mental health offers personalized, patient-driven healthcare solutions.
• Cognitive bias modification for interpretation (CBM-I) is a web-based intervention targeting interpretation biases.
• High attrition rates and lack of sustained engagement challenge digital mental health interventions.

Purpose of the Study:

• To identify early-stage high-risk dropout participants in web-based CBM-I trials.
• To determine which self-reported and passively detected features best predict dropout.

Main Methods:

• Analyzed community adults with anxiety or negative future thinking across three web-based CBM-I trials (N=1277).
• Created feature sets: baseline demographics, user context/reactions, clinical functioning, and passively detected website behavior.
• Utilized machine learning algorithms to predict participants at high risk of not starting the second CBM-I session.

Main Results:

• Extreme gradient boosting achieved high predictive performance (macro-F1 scores: .832, .770, .917).
• Passively detected user behavior features significantly contributed to dropout prediction.
• Combining all feature sets yielded the best overall predictive accuracy.

Conclusions:

• Integrating passive behavioral indicators with self-reported data enhances early dropout prediction in CBM-I.
• Findings underscore the need for personalized attrition prevention strategies in digital health.
• Generalizability remains a challenge in digital health intervention studies.