FITTER combines AI-generated exercise plans with explicit safety checks

A research preprint posted on November 2, 2026, presents FITTER, a framework for drafting personalized exercise prescriptions from clinical and wearable data. Xintong Li and colleagues evaluated it using records from 400 All of Us Research Program participants. The results support improved compliance with specified safety constraints—not proof that the prescriptions are safe or effective in practice.

The authors position FITTER as clinician-facing decision support. Its intended output is a structured draft for professional review and approval, rather than an autonomous exercise recommendation delivered directly to a patient.

Building a prescription in stages

FITTER organizes prescriptions around FITT-VP: frequency, intensity, time, type, volume and progression. These dimensions describe what exercise to perform, how often and how hard, for how long, and how the overall dose should change.

Five language-model agents work sequentially. They choose activities, set frequency, determine intensity and session duration, calculate weekly volume, and organize progression. Each receives the earlier agents’ decisions, allowing later steps to account for choices already made.

The system retrieves relevant passages from four guidance documents covering general exercise, stroke, cancer prevention and diabetes. These include a guideline, a scientific statement and position documents; they are not all formally clinical practice guidelines.

Before generation, FITTER screens the participant profile and medications for potential concerns, supplemented by a language-model safety review. These concerns become explicit constraints, such as an intensity ceiling or prohibited activity. After generation, deterministic software checks the prescription, corrects rule-checkable violations and updates dependent quantities. Changing session duration, for example, requires recalculating weekly volume.

What the evaluation measured

The evaluation used 300 participant profiles from the study’s cancer cohort and 100 from its stroke cohort. Observed inputs included diagnoses, medications, functional information and Fitbit activity measurements. Exercise goals, readiness, stress, social setting and equipment access were simulated scenario variables, not observed participant reports.

FITTER and five comparison systems used the same Qwen3.5-9B language-model backbone, decoding settings, participant profiles and output schema. The alternatives included direct prompting, step-by-step reasoning, self-refinement, reasoning-and-action prompting and multi-agent collaboration. Participant records were evaluation inputs, not training or fine-tuning data.

A separate GPT-5-family model assessed each prescription against a participant-specific checklist derived from FITTER’s safety-screening process. It received the prescription and checklist, not the underlying clinical and wearable record.

FITTER had no detected checklist violation in 394 of 400 prescriptions, or 98.5%. Comparator results ranged from 75.8% to 85.8%. This measures conformance to the supplied safety specification: risks missing from that specification could escape detection.

An ablation analysis on the 300 cancer profiles identified deterministic audit and repair as the largest contributor to this endpoint. Removing it reduced compliance from 99.0% to 62.0%. Removing guidance retrieval or model-assisted safety review also reduced compliance. These findings concern the tested cancer profiles and computational assessment, not measured reductions in patient harm.

More precise doses, with qualifications

FITTER scored highest on the study’s primary instruction-concreteness composite and on textual and semantic similarity to the guidance corpus. Its strongest concreteness components were quantitative dose specifications and exact session timing.

However, changing the composite’s weights changed the ranking: direct prompting ranked first when all four components received equal weight. The evidence therefore supports greater dosing precision, not superiority in every aspect of instruction clarity. Likewise, similarity to guidance language does not establish that a prescription is clinically appropriate for an individual.

Clinicians found reasons to revise

Four reviewers assessed 80 FITTER prescriptions, with one reviewer per prescription. They accepted 37 as written, accepted 35 with suggested changes and rejected eight. Because clinicians reviewed only FITTER outputs, these ratings cannot establish clinical superiority over the comparison systems.

For safety, reviewers gave 51 unqualified yes ratings, 27 partial ratings and two no ratings. The reported 80.6% safety score assigns half credit to partial responses; it is not the percentage judged fully safe and is not directly comparable with the binary 98.5% computational result.

Comments identified missing exercise components, insufficient attention to equipment or goals, starting doses poorly matched to baseline activity, and inadequate functional information.

The study tested one generation backbone, required linked wearable data and did not measure adherence, functional improvement or clinical outcomes. The authors call for external and prospective evaluation. For now, FITTER illustrates how explicit rules can complement language-model drafting while leaving clinical judgment and approval essential.