Dataset size versus homogeneity: A machine learning study on pooling intervention data in e-mental health dropout...

Area of Science:

• Digital mental health
• Machine learning in healthcare
• Psychological interventions

Background:

• Internet-based Cognitive Behavioral Therapy (iCBT) interventions often face challenges with limited dataset sizes.
• Small datasets can hinder the development and accuracy of machine learning models for predicting patient outcomes.
• Understanding user behavior and symptom data similarities across different iCBT interventions is crucial for data aggregation.

Purpose of the Study:

• To investigate the feasibility and benefits of pooling data from distinct iCBT interventions.
• To examine user behavior and symptom data similarities among iCBT interventions for depression, social anxiety, and panic disorder.
• To determine if pooled data enhances the accuracy of predicting intervention dropout.

Main Methods:

• Analysis of routine care patient data (n=6418) from Internet Psychiatry in Stockholm.
• Application of clustering techniques to identify patient groups based on activity levels.
• Development and comparison of dropout prediction models using individual versus pooled intervention datasets, tested on varying dataset sizes.

Main Results:

• Clustering revealed three distinct patient groups characterized by activity levels, independent of specific interventions.
• Pooling intervention data significantly improved dropout prediction accuracy in eight out of nine tested scenarios.
• Models trained on smaller datasets demonstrated a tendency to overestimate prediction results.

Conclusions:

• Patients with depression, social anxiety, and panic disorder exhibit similar online activity and dropout patterns across iCBT interventions.
• Pooling data from different iCBT interventions is a viable strategy to overcome small dataset limitations in psychological research.
• This approach can enhance the robustness and generalizability of machine learning models in digital mental health.