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

Steven Allender, Joshua Hayward, Sunil Gupta, A. Sanigorski, Santu Rana, Hugh Seward, Stephan Jacobs, Svetha Venkatesh

Introduction

Public health challenges, such as physical inactivity, are often classified as "wicked problems"—complex systems characterized by multiple causes, feedback loops, and significant resistance to simple interventions. While General Practitioners (GPs) are strategically positioned to encourage exercise, physical activity (PA) is discussed in fewer than 20% of consultations. Barriers such as time constraints, lack of specialized training, and fear of adverse outcomes hinder routine prescribing.

Traditionally, identifying effective interventions requires Randomized Controlled Trials (RCTs) or factorial designs. However, these methods scale poorly; testing nine interventions against a control would necessitate 256 separate experimental groups. This paper presents a novel approach using a Bayesian multi-arm bandit (MAB) algorithm to navigate this "combinatorial explosion," identifying optimal solutions in real-time while the experiment is still running.

Methods

The researchers employed a three-stage methodology combining systems science with reinforcement learning:

  1. Participatory Mapping: Using Causal Loop Diagrams (CLDs) developed with 26 GPs, the team mapped the systemic factors influencing PA prescribing. This process identified 15 key factors and over 80 potential interventions, which were then narrowed down to 8 priority strategies for testing.
  2. The Multi-Arm Bandit (MAB): Instead of equal allocation, the study utilized Thompson Sampling. This Bayesian algorithm treats each intervention as an "arm" of a slot machine, updating its belief about strategy effectiveness weekly based on incoming data and prioritizing clinics for the most promising strategies.
  3. Exploration vs. Exploitation: The MAB framework balances "exploitation" (applying the best-known strategy) with "exploration" (testing less certain strategies to uncover potential improvements).
Causal Loop Diagram (CLD) mapping the complex system of physical activity prescribing in general practice.

The study was conducted across 13 clinics over 12 weeks, divided into a 2-week baseline, a 7-week algorithmic phase, and a 3-week "power" phase to validate the top-performing strategies.

Data and Reward Metrics

The primary dataset comprised 11,176 GP-patient conversations from 13 clinics in Geelong, Australia. The algorithm's "reward" metric was defined as the fraction of relevant consultations that included a PA discussion, measured against each clinic's specific baseline.

Key Results and Insights

The MAB algorithm successfully identified the two most effective strategies within seven weeks, while deprioritizing ineffective interventions by week five.

Posterior mean estimates for the eight tested strategies over the 12-week study period, showing the emergence of optimal interventions.

Winning Strategies

  • Strategy 7 (Waiting room questionnaires): Achieved a mean increase in PA discussion rate of 0.15 ($p < 0.001$).
  • Strategy 3 (Receptionist info packs): Achieved a mean increase of 0.09 ($p = 0.002$).

Interestingly, the most successful interventions were "upstream" solutions that reduced the GP's cognitive load. By shifting the initial engagement to reception staff or patient self-screening, the system bypassed the need for GPs to remember to initiate the conversation under pressure. Conversely, strategies like GP mentoring (Strategy 1) or software integration (Strategy 6) showed negligible or even slightly negative effects.

Discussion

This study demonstrates the power of Bayesian strategy selection for organizational learning. By treating clinic operations as a dynamic system, the researchers identified effective interventions with a fraction of the resources required for a standard factorial design—which would have needed 512 experimental conditions to test these combinations.

Limitations and Future Work

While efficient, Thompson Sampling introduces a "rich get richer" bias, meaning less is learned about "losing" strategies that might have succeeded with more time. Additionally, the model assumes clinic independence, which may not account for regional information sharing among GPs. Future research should explore testing combinations of interventions simultaneously and measuring the long-term quality of these conversations on patient health outcomes.

Summary updated

Abstract

Complex health problems require multi-strategy, multi-target interventions. We present a method that uses machine learning techniques to choose optimal interventions from a set of possible interventions within a case study aiming to increase General Practitioner (GP) discussions of physical activity (PA) with their patients. Interventions were developed based on a causal loop diagram with 26 GPs across 13 clinics in Geelong, Australia. GPs prioritised eight from more than 80 potential interventions to increase GP discussion of PA with patients. Following a 2-week baseline, a multi-arm bandit algorithm was used to assign optimal strategies to GP clinics with the target outcome being GP PA discussion rates. The algorithm was updated weekly and the process iterated until the more promising strategies emerged (a duration of seven weeks). The top three performing strategies were continued for 3 weeks to improve the power of the hypothesis test of effectiveness for each strategy compared to baseline. GPs recorded a total of 11,176 conversations about PA. GPs identified 15 factors affecting GP PA discussion rates with patients including GP skills and awareness, fragmentation of care and fear of adverse outcomes. The two most effective strategies were correctly identified within seven weeks of the algorithm-based assignment of strategies. These were clinic reception staff providing PA information to patients at check in and PA screening questionnaires completed in the waiting room. This study demonstrates an efficient way to test and identify optimal strategies from multiple possible solutions.
BibTeX
@article{allender2020bayesian,
  author = {Steven Allender and Joshua Hayward and Sunil Gupta and A. Sanigorski and Santu Rana and Hugh Seward and Stephan Jacobs and Svetha Venkatesh},
  title = {{Bayesian strategy selection identifies optimal solutions to complex problems using an example from GP prescribing}},
  journal = {Npj Digit. Medicine},
  year = {2020},
  doi = {10.1038/s41746-019-0205-y},
}
Figure