Bayesian strategy selection identifies optimal solutions to complex problems using an example from GP prescribing.

Area of Science:

• Health Services Research
• Applied Machine Learning
• Behavioral Science

Background:

• Complex health issues necessitate multi-faceted interventions.
• General Practitioner (GP) patient discussions on physical activity (PA) are crucial but often insufficient.
• Identifying optimal interventions requires efficient testing methodologies.

Purpose of the Study:

• To employ machine learning for selecting optimal interventions to increase GP-patient PA discussions.
• To evaluate the efficiency of a multi-arm bandit algorithm in identifying effective health strategies.
• To determine the most impactful interventions for promoting PA conversations in primary care settings.

Main Methods:

• Developed interventions based on a causal loop diagram with 26 GPs.
• Utilized a multi-arm bandit algorithm for weekly strategy assignment and optimization over seven weeks.
• Assessed GP PA discussion rates as the primary outcome measure, recording 11,176 conversations.

Main Results:

• The machine learning approach efficiently identified the top three performing strategies within seven weeks.
• Clinic reception staff providing PA information and PA screening questionnaires in waiting rooms were the most effective interventions.
• Identified 15 factors influencing GP PA discussion rates, including GP skills and care fragmentation.

Conclusions:

• Machine learning, specifically multi-arm bandit algorithms, offers an efficient method for testing and optimizing multiple health interventions.
• Simple, systematically implemented strategies involving support staff and patient-initiated screening can significantly increase GP-patient PA discussions.
• This approach can accelerate the identification of evidence-based practices in primary care.