Artificial intelligence-enabled digital health systems for personalized and adaptive exercise management: A scoping review
Mirjam Bonanno, Andrea Calderone, Giovanni Lonia, Silvia Migale, Daniele Ravì, Angelo Quartarone, Demetrio Milardi, Antonio Celesti, Rocco Salvatore CalabròBackground
Artificial intelligence (AI) is increasingly used in digital exercise systems, but technical outputs are often described as personalized without showing how they inform an exercise-management decision.
Objective
To map empirical evidence on AI-enabled systems that generate, adapt, evaluate, or support individualized exercise management in adults and to examine validation, actionability, and methodological limitations.
Methods
We searched five databases through 28 May 2026. Eligible studies linked an AI output to exercise planning, dose/load adjustment, coaching, rehabilitation feedback, adherence, safety, or professional decisions. Appraisal used RoB 2 for randomized trials, ROBINS-I for nonrandomized intervention studies, PROBAST+AI for prediction models, and MMAT for other designs.
Results
Of 15,865 records, 14,584 were screened, 234 full reports were assessed, and 54 studies were included: 19 generative/large language model studies, 12 adaptive recommender, reinforcement-learning or mobile-health studies, 11 sensor-driven monitoring studies, and 12 clinical or rehabilitation-delivery studies. Among 14 randomized trials, three were at low risk of bias, six raised some concerns, and five were at high risk. All six nonrandomized intervention studies were at serious risk of bias; prediction-model and descriptive studies most often had concerns related to sample representativeness, internal validation, missing prespecification, or limited reproducibility. Clinical effectiveness, external validation, adverse-event reporting, accessibility, and workflow integration were inconsistent.
Conclusions
Current evidence supports bounded, supervised uses of AI for communication, behavioral adaptation, monitoring, and decision support. It does not establish routine autonomous exercise prescription. External validation, explicit safety escalation, accessible design, and prospective evaluation of patient-relevant decisions remain priorities.