A Guideline-Grounded Multi-Agent Framework for Personalized Exercise Prescription from Clinical and Wearable Data
Authors: Xintong Li, Yufen Lin, Yuhua Wu, Fengran Wang, Minxiao Wang, Rebecca A. Gary, Jacob Kariuki, James Smith, Sandra A. Billinger, Canhua Xiao, Xiao Hu, Runze Yan
Organizations: Center for Data Science, Nell Hodgson Woodruff School of Nursing, Emory University · Department of Computer Science, Emory University · Nell Hodgson Woodruff School of Nursing, Emory University · Cancer Prevention and Control Research Program, Winship Cancer Institute, Emory University · Lillian Carter Center for Global Health and Social Responsibility, Nell Hodgson Woodruff School of Nursing, Emory University · Lancaster Medical School, Lancaster University · Department of Neurology, University of Kansas Medical Center · Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology and Emory University
Translating population-level exercise guidance into individualized prescriptions requires integrating clinical and wearable data, including diagnoses, medications, functional limitations, and baseline activity, with condition-specific recommendations, a process difficult to scale beyond specialist rehabilitation settings. We developed FITTER, a multi-agent framework that integrates these data with condition-specific guidance to generate personalized FITT-VP prescriptions and applies deterministic safety audit and repair to correct rule-checkable violations. We evaluated FITTER using data from 400 All of Us participants (300 cancer, 100 stroke) against five LLM comparators through computational evaluation, ablation, and structured clinician review. FITTER achieved 98.5% safety-constraint compliance versus 75.8-85.8% for comparators and showed greater quantitative dosing precision and guideline alignment. Cancer-cohort ablation identified safety audit and repair as the largest contributor to safety-constraint compliance. Clinicians reviewed 80 prescriptions across safety, personalization, instruction concreteness, anticipated effectiveness, and clinical acceptability. FITTER provides a clinician-facing approach integrating clinical and wearable data, guidance, and explicit safety constraints.
Figures & tables
Cancer cohort
Stroke cohort
Characteristic
Full
Evaluation
SMD
Full
Evaluation
SMD
( n=3,764 )
( n=300 )
( n=790 )
( n=100 )
Age, years
67.0 [58.0, 73.0]
66.0 [57.0, 72.2]
-0.052
67.0 [57.0, 73.0]
66.5 [56.0, 72.0]
-0.062
Body mass index, kg/m 2
28.1 [24.7, 32.5]
28.3 [25.0, 32.2]
0.027
28.6 [25.0, 33.8]
30.5 [25.4, 34.1]
0.133
Systolic blood pressure, mmHg
127.0 [119.0, 136.3]
127.6 [118.5, 138.0]
0.024
126.3 [118.1, 134.0]
124.5 [120.5, 134.5]
0.071
Diastolic blood pressure, mmHg
74.0 [69.0, 79.0]
73.8 [68.0, 78.9]
-0.044
73.1 [68.0, 78.0]
73.2 [69.4, 77.5]
0.061
Table 1: Characteristics of the eligible cancer and stroke cohorts and the two randomly selected evaluation samples comprising 400 participants. Continuous variables are reported as median [Q1, Q3] and categorical variables as n (%). Pain, fatigue, general health and sleep variables are self-reported. Standardized mean differences (SMDs) compare each diagnosis-specific evaluation sample with its eligible source cohort; absolute values below 0.10 indicate very small differences and values below 0.20 indicate small differences.
Figure 1: Computational evaluation across the 400 samples. a , Proportion of prescriptions with no detected safety-constraint violation by an independent model assessor. b , Instruction-concreteness composite (0–1). c , Guideline alignment, the embedding (cosine) similarity between each FITT-VP component and the guideline passages, averaged across components to give one value per prescription. FITTER is compared with five LLM comparators. Error bars show 95% confidence intervals. Asterisks indicate Holm-adjusted significance versus FITTER ( ∗P<0.05;∗∗∗P<0.001)
Table 2: Cross-system comparison across 400 matched participant profiles. Each metric was first tested with a Friedman or Cochran omnibus test across the six systems, then FITTER was compared with each comparator, and the five resulting P values were Holm-adjusted. Differences are FITTER minus comparator, so positive values favor FITTER. The pairwise measure is the matched-pairs rank-biserial correlation for the continuous metrics and the discordant-pair counts (FITTER safe/comparator safe) for the binary safety outcome. Confidence intervals are from 10,000 bootstrap resamples of the paired differences.
Figure 2: Clinician evaluation of 80 FITTER prescriptions. Each prescription was rated by one clinician reviewer on five aspects using a yes/partial/no scale. Values are mean normalized scores on a 0–100 scale, with yes, partial, and no assigned values of 100, 50, and 0, respectively.
“Only includes cardiorespiratory activity. Other types of exercise (e.g., strength) should be included as per guidelines.”
Equipment access or stated goal not fully reflected in the plan
19
“The patient does not have access to a recumbent cycle.”
Starting dose could be better aligned with measured baseline activity
14
“The patient is already fairly active, the starting intensity and dose is too low.”
Less detail provided for non-aerobic than aerobic components
13
“No specific exercises, sets, or reps for strength exercise. RPE may not be appropriate for strength exercises.”
Additional functional detail requested to assess the plan
13
“I would need more information on the patient’s level of disability and functional ability.”
Clinical considerations outside the structured input
12
“Would be concerned about fentanyl use around exercise time, could be a fall risk.”
Table 3: Themes raised in at least ten of the 55 reviews with free-text comments. A review could contribute to more than one theme. Representative comments are reproduced verbatim apart from minor spelling corrections.
Condition
Judged safe
Difference
Holm p
Full FITTER
99.0%
reference
—
No guideline retrieval
82.0%
-17.0 pp
8.6×10−14
No model-assisted safety review
76.3%
-22.7 pp
2.2×10−18
No deterministic audit and repair
62.0%
-37.0 pp
9.6×10−31
Table 4: Ablation analysis using 300 cancer participant profiles. Full FITTER and three ablation variants were assessed by the percentage of prescriptions meeting all specified safety constraints. Each variant removed one component, and all four conditions used the same participant profiles and model assessor. Differences are ablation variant minus full FITTER, in percentage points (pp). Cochran Q=137.9 on 3 degrees of freedom, p=1.1×10−29 ; exact McNemar tests against full FITTER are Holm-adjusted.
Figure 3: FITTER workflow. Clinical, functional, medication, and wearable data form a structured participant profile. Pre-generation screening produces safety constraints; five sequential agents then generate Type, Frequency, Intensity and Time, Volume, and Progression using retrieved guideline passages. After generation, the complete FITT-VP prescription is checked against those constraints; detected violations are corrected, producing the final structured prescription.
Agent
Decision and downstream output
Type
Selects aerobic, resistance, flexibility, and neuromotor components and names the prescribed activities.
Frequency
Sets initial and target sessions per week for each included component.
Intensity and Time
Sets initial and target RPE ranges, talk-test descriptions, session durations, and stop criteria.
Volume
Calculates initial and target weekly minutes for each component and for the full prescription.
Progression
Converts the prescription into six two-week blocks with progression, safety, and regression rules.
Table 5: Sequential FITTER agents and their structured outputs.
Constraint class
Actionable content
FITT-VP fields checked
Intensity limit
Permitted RPE or heart-rate range, including a prohibition on vigorous exercise
Intensity; Progression
Activity restriction
Prohibited exercise component or named activity
Type; Progression
Frequency, time, or volume limit
Bounds on sessions per week, minutes per session, or weekly minutes
Frequency; Time; Volume; Progression
Progression restriction
Permitted increases, required holds, or regression conditions
Progression; Volume
Stop criterion
Symptom or vital-sign condition requiring exercise to stop
Intensity and Time; Progression
Table 6: Standardized safety-constraint records stored in S and the FITT-VP fields governed by each class.
Personalized nutrition intervention for patients with multimorbidity is critical for improving health outcomes, yet remains challenging because it requires the simultaneous integration of heterogeneous clinical conditions, medications, and dietary guidelines. Single-agent large language models (LLMs) often suffer from context overload and attention dilution when processing such high-dimensional patient profiles. We introduce NutriOrion, a hierarchical multi-agent framework with a parallel-then-sequential reasoning topology. NutriOrion decomposes nutrition planning into specialized domain agents with isolated contexts to mitigate anchoring bias, followed by a conditional refinement stage. The framework includes a multi-objective prioritization algorithm to resolve conflicting dietary requirements and a safety constraint mechanism that injects pharmacological contraindications as hard negative constraints during synthesis, ensuring clinical validity by construction rather than post-hoc filtering. For clinical interoperability, NutriOrion maps synthesized insights into the ADIME standard and FHIR R4 resources. Evaluated on 330 stroke patients with multimorbidity, NutriOrion outperforms multiple baselines, including GPT-4.1 and alternative multi-agent architectures. It achieves a 12.1 percent drug-food interaction violation rate, demonstrates strong personalization with negative correlations (-0.26 to -0.35) between patient biomarkers and recommended risk nutrients, and yields clinically meaningful dietary improvements, including a 167 percent increase in fiber and a 27 percent increase in potassium, alongside reductions in sodium (9 percent) and sugars (12 percent).
Junwei Wu, Runze Yan, Hanqi Luo +7
Emory University, Atlanta, GA, USA · Texas Woman’s University, Denton, TX, USA · Cecelia Health, New York, NY, USA
Large language model (LLM) agents for personal health read wearable data and recommend, but do not learn about the individual in a way that can be audited or falsified. We present EpiCoach, an agentic system that treats one athlete, a cross-country mountain biker, as the subject of accountable N-of-1 science. Claims about the athlete live in an evidence-gated hypothesis registry: each must issue a probabilistic prediction about a future session before it occurs, and its belief moves only through computed evidence from a logged statistical fit. Predictions come from a mechanistic digital twin inferred from heart rate, GPS and barometric altitude: a lagged heart-rate response model, a fitness-fatigue model with learned time constants, and a latent technical-descending skill state, measured on discovered trail segments. The agent prescribes the next training week as a constrained experiment maximizing information gain subject to load and safety constraints, around two instrument rides. We evaluate the system two ways over eighteen months of data. A rolling-origin replay benchmark shows the twin beats standard forecasters on heart rate but not segment time, and the agentic loop matches a static model. A live prospective phase runs the weekly cycle: in its first three weeks the ledger sealed 20 predictions, scored 14, the descent forecast calibrated after one correction, the climb forecast pessimistic as the athlete improved, and the safety officer vetoed the agent twice. Every correction is preserved in the registry, a template for auditable, falsifiable personalization in digital health. We release the registry, benchmark and protocol.
Personalized health coaching requires moving beyond retrospective wearable summaries toward quantitative, actionable behavioral prescriptions. We introduce Coachable Health Age (CHA), an interpretable health-age metric built from modifiable physiological and behavioral signals passively captured by consumer wearables. Across the WEAR-ME (N=2,780), All of Us (N=22,748), and longitudinal Fitbit (N=1,464) cohorts, CHA aligns with established cardiometabolic health constructs and improves prognostic discrimination for incident cardiovascular and metabolic disease beyond age, sex, and body mass index. To enable prospective behavioral optimization, we developed Sensor World Model (SensorWM), an action-conditioned wearable world model pretrained on over 2.2 million person-weeks of longitudinal sensor data from approximately 100,000 participants. SensorWM accurately forecasts multi-week physiological responses to candidate lifestyle changes (R2=0.664), outperforming generic time-series foundation models while preserving downstream clinical signal for diabetes, hypertension, and other health outcomes. Finally, we integrate CHA and SensorWM into an agentic counterfactual coaching system to synthesize personalized 12-week coaching plans. In blinded paired reviews, certified lifestyle coaches preferred counterfactual-grounded plans over plans grounded in wearable history alone for overall quality (71.4%, P < 0.001) and anticipated adherence (71.0%, P < 0.001), and board-certified clinicians judged 99.2% of counterfactual-grounded plans to meet clinical safety criteria (vs. 97.6% of comparison plans). Automated evaluations across the full cohort (N=2,780) also favored counterfactual-grounded plans. By closing the loop between longitudinal tracking, trajectory prediction, and counterfactual coaching, our framework provides a scalable path to systematically improve cardiometabolic health.
Simon A. Lee, Zeinab Esmaeilpour, Menglian Zhou +17
Google Research · University of California, Los Angeles