Feasibility of a Reinforcement Learning-Enabled Digital Health Intervention to Promote Mammograms: Retrospective, Single-Arm, Observational Study

Abstract

Background: Preventive screenings such as mammograms promote health and detect disease. However, mammogram attendance lags clinical guidelines, with roughly one-quarter of women not completing their recommended mammograms. A scalable digital health intervention leveraging behavioral science and reinforcement learning and delivered via email was implemented in a US health system to promote uptake of recommended mammograms among patients who were 1 or more years overdue for the screening (ie, 2 or more years from last mammogram).

Objective: The aim of this study was to establish the feasibility of a reinforcement learning-enabled mammography digital health intervention delivered via email. The research aims included understanding the intervention's reach and ability to elicit behavioral outcomes of scheduling and attending mammograms, as well as understanding reach and behavioral outcomes for women of different ages, races, educational attainment levels, and household incomes.

Methods: The digital health intervention was implemented in a large Catholic health system in the Midwestern United States and targeted the system's existing patients who had not received a recommended mammogram in 2 or more years. From August 2020 to July 2022, 139,164 eligible women received behavioral science-based email messages assembled and delivered by a reinforcement learning model to encourage clinically recommended mammograms. Target outcome behaviors included scheduling and ultimately attending the mammogram appointment.

Results: In total, 139,164 women received at least one intervention email during the study period, and 81.52% engaged with at least one email. Deliverability of emails exceeded 98%. Among message recipients, 24.99% scheduled mammograms and 22.02% attended mammograms (88.08% attendance rate among women who scheduled appointments). Results indicate no practical differences in the frequency at which people engage with the intervention or take action following a message based on their age, race, educational attainment, or household income, suggesting the intervention may equitably drive mammography across diverse populations.

Conclusions: The reinforcement learning-enabled email intervention is feasible to implement in a health system to engage patients who are overdue for their mammograms to schedule and attend a recommended screening. In this feasibility study, the intervention was associated with scheduling and attending mammograms for patients who were significantly overdue for recommended screening. Moreover, the intervention showed proportionate reach across demographic subpopulations. This suggests that the intervention may be effective at engaging patients of many different backgrounds who are overdue for screening. Future research will establish the effectiveness of this type of intervention compared to typical health system outreach to patients who have not had recommended screenings as well as identify ways to enhance its reach and impact.

Keywords: artificial intelligence; behavioral intervention; cancer screening; digital health; email; feasibility studies; health equity; mammograms; nudging; reinforcement learning.

Figure 1

Sample assembled messages from the Precision Nudging mammography intervention.

Figure 2

An example of the messaging cadence by cohort showing alternating pulses of one message per week for 5 weeks and an 8-week pause.

Figure 3

Bar graphs showing the proportion of each demographic subgroup who engaged and took action after receiving an intervention. Relatively small changes down the funnel from reached to opened to clicked to scheduled to attended suggest, from a practical perspective, proportional response to the intervention for people in that group.