Randomized trial of two artificial intelligence coaching interventions to increase physical activity in cancer survivors

Obesity is a significant public health concern, linked to increased risk of cardiovascular disease, diabetes, and cancer. The prevalence of obesity is high, resulting in substantial economic burden. Among cancer survivors, obesity reduces survivorship and increases CVD-related mortality. Increasing physical activity is crucial for improving health outcomes in this population. While personalized coaching can improve adherence to physical activity recommendations, scalability remains a challenge. This study aimed to evaluate the potential of novel AI-based coaching interventions to address this challenge, leveraging the scalability and personalization offered by smart speaker and text message technologies. The study hypothesized that AI-driven coaching would lead to significant increases in physical activity compared to a control group receiving standard health information. The high prevalence of physical inactivity and the limitations of traditional, person-to-person interventions highlight the need for scalable solutions. AI offers a potential solution by providing personalized, on-demand coaching adapted to individual behavior using data from fitness trackers and health records.

The introduction cites several studies highlighting the negative health consequences of obesity, particularly among cancer survivors. It emphasizes the effectiveness of behavioral interventions, but notes the limitations of traditional, person-to-person coaching due to cost and time constraints. The introduction lays the groundwork for the study by highlighting the potential of AI-powered interventions to overcome these limitations by providing scalable and personalized support. The increased availability of smart devices and data analytics technologies is emphasized as key to the feasibility of the proposed AI interventions. It also mentions previous research in the use of technology for increasing physical activity.

This study employed a three-arm, randomized, controlled pilot trial with 42 participants (14 per arm). Participants were overweight or obese cancer survivors who were physically inactive. The three arms were: (1) voice-assisted AI coaching (MyCoach) delivered via a smart speaker; (2) autonomous AI text message coaching (SmartText); and (3) a control group receiving standard health information. Data collection was automated via wearable sensors (Fitbit Charge 2 HR) and voice technology, ensuring blinded outcome ascertainment. The primary outcome measure was change in mean steps per day from baseline (1 week) to the end of the 4-week intervention period. Stratified permuted block randomization was used, considering age, sex, and BMI. Baseline characteristics were similar across groups. The interventions were delivered as intended. MyCoach involved interactive conversations with the AI agent, while SmartText provided unidirectional text messages. The control group received educational materials on physical activity. Statistical analysis included intention-to-treat analysis using regression models to assess the effects of each intervention compared to control and to compare the two active interventions. Missing data (2.3% of person-days) were imputed with the average steps from valid person-days. Sample size calculations aimed to detect a 2000 steps/day difference with 80% power and a 2-sided alpha of 0.05.

At 4 weeks, the MyCoach arm showed a significant increase in average steps per day (3618.2, 95% CI: 2490–4764) compared to the control arm (net difference = 3568.9 steps/day, P<0.001) and showed a marginally significant increase compared to the SmartText arm (net difference = 2160.6 steps/day, P=0.049). While both intervention groups showed increased steps compared to baseline, MyCoach demonstrated a sustained increase throughout the intervention period, unlike SmartText, which showed an early increase followed by a decline. The percentage of person-days achieving ≥10,000 steps was substantially higher in the MyCoach arm (61%) compared to SmartText (41%) and control (28%). These findings remained consistent when analyzing data from the final week of the intervention. There were no adverse events reported.

The study's primary finding is the superior effectiveness of voice-assisted AI coaching (MyCoach) in increasing physical activity among overweight or obese, inactive cancer survivors. The sustained engagement observed with MyCoach may be attributed to its bidirectional communication approach, allowing for user-driven interaction, compared to the unidirectional nature of SmartText. The results suggest that interventions allowing users to control the frequency and mode of interaction might be more effective in achieving and maintaining behavior change. The findings highlight the potential of AI-powered coaching to address the public health challenge of physical inactivity among cancer survivors, emphasizing the benefits of increased physical activity in reducing cancer recurrence and CVD-related morbidity and mortality. While the study focused on cancer survivors, the methodology could be adapted to other populations. The short-term nature of the pilot study limits the ability to assess long-term outcomes and the generalizability of the findings to broader populations.

The study demonstrates the promising potential of AI-based voice-assisted coaching (MyCoach) as a scalable and effective method for increasing physical activity in sedentary cancer survivors. The superior performance of MyCoach compared to SmartText highlights the importance of bidirectional communication and user-driven interaction in AI-based interventions. Future research should focus on replicating these findings in larger, more diverse populations, investigating long-term effects, and exploring the application of this approach to other health behaviors and populations. Further investigation into the mechanisms underlying the differences between MyCoach and SmartText is needed.

This study was a small-scale pilot trial with a short (4-week) follow-up period. The sample was predominantly female and comprised mainly of breast cancer survivors, limiting the generalizability of the findings. Inclusion criteria required access to technology, which may limit the applicability to populations lacking such access. The study did not assess the impact on weight change or other health outcomes. The use of a secure local server for data hosting could limit the scalability of the intervention. However, the high follow-up rate and automated data collection are strengths.