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

The study addresses the problem of low physical activity among overweight and obese cancer survivors, a group at elevated risk for cardiovascular disease, diabetes, cancer recurrence, and mortality. Traditional behavioral interventions involving individualized coaching are effective but resource-intensive and difficult to scale. With the proliferation of intelligent voice assistants, wearables, and connected technologies, AI-driven coaching may provide scalable, personalized interventions. The research question was whether AI-based coaching—via a voice-assisted agent (MyCoach) or an autonomous text messaging agent (SmartText)—can increase physical activity compared with control educational materials among sedentary cancer survivors with overweight/obesity. The purpose was to evaluate feasibility and short-term efficacy over 4 weeks, with the significance tied to addressing a major public health need for scalable physical activity interventions in a high-risk population.

The paper situates the trial within evidence showing rising obesity prevalence and associated poor outcomes, including among cancer survivors. Prior technology-supported interventions for cancer survivors have often been short in duration, limiting assessment of long-term outcomes. The discussion references a systematic review concluding most technology-based physical activity studies in adults with cancer are of short duration, emphasizing the need for scalable, effective approaches and longer-term evaluations.

Design: Single-center, three-arm, randomized parallel pilot trial (PATH) with 1:1:1 allocation to MyCoach (voice-assisted AI coaching via Amazon Echo), SmartText (autonomous AI text messaging), or control (NCI educational materials). IRB-approved; written informed consent obtained. Participants and eligibility: Maryland adults with history of breast, prostate, colon, lung, cervical, oral, or melanoma cancer; BMI ≥25 kg/m²; completed cancer treatment ≥3 months prior (except ongoing anti-hormonal therapy allowed); access to a smartphone (Android or iOS); able to perform mild-to-moderate activity (e.g., walking); physician clearance; willing to be randomized; willing to wear an activity tracker for 5 weeks. Exclusions: ≥150 min/week routine physical activity in prior 4 weeks; stage 4 cancer; plans to relocate; participation in structured physical activity programs or consumer tech-guided plans; psychiatric conditions interfering with participation; pregnancy. Timeline: 1-week baseline (no new activities; safety monitoring; establish baseline PA) followed by 4-week intervention. Randomization: Stratified permuted block randomization (factors: age, sex, BMI). Six blocks of size 3. Sequence concealed using Microsoft 365 smart forms; assignments conducted using random.org. Interventions:

• MyCoach (voice-assisted AI): On-demand, bidirectional coaching via Amazon Echo/Alexa smart speaker installed at home. Reinforcement learning–based recommendation system with reward signals tied to PA goals (e.g., ≥10,000 steps/day). Real-time feedback from Fitbit Charge 2 HR via secure linkage; hosted on secure Johns Hopkins server.
• SmartText (autonomous texting): Goal-based, supervised AI agent delivering three personalized text messages daily. Content adapted using participant schedule, anthropometrics, wearable feedback, preferences, and progress. Unidirectional messaging; computations on secure server with minute-by-minute wearable data.
• Control: Printed and emailed NCI educational materials about PA for cancer survivors recommending 10,000 steps/day. Data collection: Fitbit Charge 2 HR provided to all arms; minute-by-minute data transmitted via research API to a secure database. Remote, automated data capture; no in-person follow-up needed. Wear-time validated using heart rate sensor. Blinding: Participants not blinded due to intervention nature; outcome ascertainment effectively blinded via automated sensor-based data collection and transfer. Outcomes: Primary endpoint analyzed as change in average steps per day from 1-week baseline to end of 4-week intervention. Also assessed change using only the final (fourth) week. Process measures: control material engagement, number of SmartText messages sent/received, and MyCoach interaction sessions per day. Sample size: Planned 39 (13/arm) to detect 2000 steps/day between-arm difference (SD 1800), 80% power, two-sided alpha 0.05; enrolled 42 to account for dropouts. Missing data: Of 1470 expected person-days, 34 (2.3%) missing; imputed using each participant’s mean of valid person-days. Statistical analysis: Intention-to-treat. Within-arm changes estimated via regression with participant-level clustering; between-arm differences in change estimated via multiple linear regression with period-by-arm interaction, adjusting for baseline and clustering by participant ID. Reported means, SDs, 95% CIs, and P values. Analyses repeated using only week 4 follow-up.

Participants: 42 randomized (14/arm). Mean age 62.1 years; mean BMI 32.9 kg/m²; 91% female; 36% Black; 85.7% stage 1 or 2 breast cancer. Baseline steps similar across arms. Within-arm changes (baseline to all 4 weeks):

• Control: +886.1 steps/day (95% CI: -894.9 to 2667.1)
• SmartText: +1619.0 steps/day (95% CI: -328.1 to 3566.2)
• MyCoach: +3618.2 steps/day (95% CI: 2490.1 to 4764.2) Within-arm changes (baseline to last week only):
• Control: +746.6 steps/day (95% CI: -1544.1 to 3037.4)
• SmartText: +1402.4 steps/day (95% CI: -1025.6 to 3830.4)
• MyCoach: +3585.0 steps/day (95% CI: 2303.6 to 4866.4) Between-arm differences in change (all 4 weeks):
• MyCoach vs Control: +3568.9 steps/day (95% CI: 1482.7 to 5655.0), P=0.001
• MyCoach vs SmartText: +2160.6 steps/day (95% CI: 11.4 to 4309.7), P=0.049
• SmartText vs Control: +1408.2 steps/day (95% CI: -1312.4 to 4128.9), P=0.30 Between-arm differences in change (last week only):
• MyCoach vs Control: +3675.1 steps/day (95% CI: 1304.7 to 6045.5), P=0.003
• MyCoach vs SmartText: +2344.1 steps/day (95% CI: 152.8 to 4535.3), P=0.037
• SmartText vs Control: +1331.0 steps/day (95% CI: -1731.9 to 4393.9), P=0.39 Proportion of person-days with ≥10,000 steps during entire intervention: Control 28%, SmartText 41%, MyCoach 61% (week 4 only: 31%, 34%, 58%, respectively). Process and safety: Control participants received materials; SmartText participants received 3 texts/day (one participant missed 1 day); MyCoach participants averaged two interactions/day. No adverse events reported.

The trial demonstrates that AI-based, voice-assisted, bidirectional coaching (MyCoach) significantly increased physical activity among sedentary, overweight/obese cancer survivors compared with both control and a unidirectional autonomous text messaging intervention (SmartText). While all arms showed some increase in steps, SmartText did not significantly outperform control, and its early gains waned over time, whereas MyCoach’s gains were early and sustained across 4 weeks. The greater efficacy of MyCoach may be attributable to bidirectional, on-demand interaction facilitating user agency, tailored timing, and flexible contact frequency, contrasted with SmartText’s fixed, unidirectional schedule. These findings support the research hypothesis that AI-powered, interactive voice coaching can deliver scalable, individualized behavior change support with meaningful short-term increases in physical activity in a high-risk population. The results highlight the potential of AI coaching to address public health needs where traditional person-to-person models are not scalable, with implications for extending such technology to broader populations and other lifestyle behaviors.

AI voice-assisted coaching (MyCoach) produced substantial and sustained increases in daily steps over 4 weeks among sedentary, overweight/obese cancer survivors, outperforming both control and an autonomous text messaging intervention. This suggests voice-based, bidirectional AI coaching is a promising, scalable approach to individualized physical activity promotion. Future research should replicate findings in larger, more diverse samples, extend follow-up to assess long-term maintenance and health outcomes (e.g., weight, cardiometabolic health, cancer recurrence), evaluate cost-effectiveness, and test generalizability to the broader population and other behaviors.

• Short duration (4-week intervention) limits assessment of long-term maintenance and downstream health outcomes; last-week analyses suggest sustained effect but longer follow-up is needed.
• Small sample size (n=42) limits precision and generalizability; replication in larger cohorts warranted.
• Generalizability may be limited to sedentary, overweight/obese cancer survivors; results may not extend to all sedentary individuals or normal-weight survivors.
• Requirement for access to and comfort with technology (smartphone, smart speaker) may limit applicability; technology access is increasing but not universal.
• Data privacy and security considerations; at the time, hosting was required on a secure institutional server; HIPAA-compliant commercial solutions now exist.
• The trial was not designed to assess effects on weight or broader health outcomes.