PaperlayerView exact revision on arXiv

Source-linked AI summary

The Athlete as a Scientist of One: Agentic Bayesian Experimental Design for Individualized Physiology from Consumer Wearables

Behnam Moradi

arXiv:2610.40370v1cs.ROcs.AI

TL;DR

Wearable-health agents can recommend actions without maintaining testable, revisable claims about an individual, while N-of-1 coaching must learn from sessions that also change the athlete. EpiCoach combines a prediction ledger, a wrist-derived twin, and safety-constrained experiment selection; its twin performs better on heart-rate forecasts than standard forecasters but not on segment time, and the agentic loop matches a static model.

  • Problem

    Personal-health agents can read wearable data and advise users, but do not maintain explicit, probabilistic claims about an individual that can be tested and revised.

    Source: §I · p. 1

  • Method

    EpiCoach uses an evidence-gated hypothesis registry and pre-ride prediction ledger, with a wearable-derived twin and workout selection constrained by load and safety rules.

    Source: §I · p. 2, C. The design problem · p. 3

  • Results

    The twin beat standard forecasters on heart rate (CRPS 7.03 bpm vs 7.87 for an exponentially weighted average) but not segment time; across 60 origins, agentic and static twins matched on HR.

    Source: E. Results · pp. 8–9

  • Takeaways & Limitations

    EpiCoach provides a protocol for accountable, falsifiable personalization in digital health, with language models proposing and explaining while deterministic tools fit, score, gate, and record.

    Source: §VIII · p. 12

    The prospective evidence covers three weeks with one athlete, and the randomized information-directed versus conventional comparison has not yet begun.

    Source: §VII · p. 12

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper frames personal coaching as accountable learning about one athlete, rather than only turning wearable data into recommendations. It positions this approach against personal-health agents, automated-science systems, training models, and self-experimentation methods.

I. INTRODUCTION

  • EpiCoach formalizes coaching as an evidence-gated registry with pre-ride predictions, a wrist-only mountain-biking twin, and safety-constrained experiment selection.The system separates language-model proposals from deterministic fitting, scoring, and gating.Source: §I · pp. 1–2

II. RELATED WORK

  • Existing personal-health agents answer questions and give advice but do not maintain persistent, probabilistic, falsifiable user models or choose what data to collect next.The paper contrasts their data-in/recommendation-out direction with claims that can be tested and revised for an individual.Source: §I · p. 1, §II · p. 2
  • Automated-science systems close hypothesis–experiment–analysis loops in stationary in-silico or laboratory settings, whereas this setting has a living subject and irreversible state.The Hypothesis Evolution Protocol supplies registry discipline; EpiCoach adds dated predictions about unobserved outcomes.Source: §I · p. 1, §II · p. 2
  • Training-load models relate load to performance through fitness and fatigue filters, but are ill-conditioned; EpiCoach treats parameter uncertainty as something experiment design should reduce.Its twin uses wrist-derived load and segment-based performance rather than power meters or laboratory tests.Source: §II · p. 2
  • The agent proposes candidate N-of-1 experiments from model residuals, selects by expected information gain, executes them through the athlete’s device, and analyzes them with randomization-respecting estimators.This extends the cited LLM assistance for experiment design into an execution-and-analysis loop.Source: §II · p. 2
  • Mountain-bike physiology motivates separating climb-based heart-rate fitness signals from descending skill and treating surface-dependent rolling resistance as a segment-level nuisance.Heart rate can remain elevated on descents despite low power output, while rolling resistance varies two- to three-fold with surface.Source: §II · p. 2

III. COACHING AS ACCOUNTABLE SEQUENTIAL EXPERIMENTATION

The section formulates coaching as sequential design over a stream of rides and wellness records, with a registry that admits only auditable evidence. The next workout balances learning and predicted performance while obeying hard safety and workload constraints.

A. Setting

  • A ride supplies heart-rate, GPS, altitude, and temperature data; the twin predicts outcomes for executable workout designs using the posterior over athlete parameters.Each workout specifies an archetype, duration, heart-rate target, segment, and day.Source: A. Setting · p. 2

B. Hypothesis registry and prediction ledger

  • A hypothesis combines a natural-language claim with a formal predicate or sub-model, probabilistic belief, lifecycle state, and lineage pointers.Beliefs change only through a logged statistical fit, and a hypothesis enters testing only after a future-ride prediction is registered before the ride begins.Source: B. Hypothesis registry and prediction… · pp. 2–3

C. The design problem

  • Workout choice trades expected information gain against predicted performance change, while code enforces workload, rest, heart-rate, readiness, illness, and calendar constraints.The information weight decays over a training block, shifting from exploration toward exploitation; infeasible designs cannot be admitted by the language model.Source: C. The design problem · p. 3

IV. A WRIST-ONLY MECHANISTIC TWIN FOR MOUNTAIN BIKING

The twin derives climb-response inputs from wrist and route data, uses repeated trail segments as measurement units, and fits a climb heart-rate response model. The reported retrospective fit estimates cardiac drift but finds no supported heat–drift interaction in the available observations.

A. Signal quality and derived quantities

  • The device’s wrist heart rate is quality-flagged for cadence lock and implausible jumps; climb virtual power uses grade, speed, and acceleration, with rolling resistance absorbed by segment.Virtual power is used only on climbs as an input to the heart-rate response and is never reported as mechanical power.Source: A. Signal quality and derived quantities · p. 3

B. Repeated segments as instruments

  • Consensus gates define trail segments from the longest stretch covered by at least 60% of rides touching a chain and at least five rides, enabling repeated traversal timing.Traversals record moving time as the primary timing outcome, along with heart-rate summary and quality.Source: B. Repeated segments as instruments · pp. 3–4

C. Heart-rate response on climbs

  • On 188 climb traversals across 9 segments, cardiac drift was 6.0 bpm/h, while the heat–drift interaction was indistinguishable from zero.Only 17 rides exceeded 25°C; the model estimated β = 19 ± 3 bpm per 50 W and a heat offset γT = 0.34±0.23 bpm/°C.Source: C. Heart-rate response on climbs · p. 4

D. Technical descending as a latent skill

The paper models technical descent time as a measure of latent skill, while treating fitness–fatigue parameters as difficult to identify from the observed training record. Information-directed experiments are proposed to create contrasts that can distinguish those parameters.

D. Technical descending as a latent skill

  • For descents, the model treats segment time at matched conditions as a latent technical-skill measure because heart rate decouples from mechanical work.Segment loading represents how strongly a trail informs skill; covariates can include prior-night sleep and elapsed ride time, while skill follows a weekly random walk.Source: D. Technical descending as a latent… · p. 4
  • The fitted relationship indicates that one hour less sleep slows descent time by 2.2%, with a 90% interval of −0.2 to +4.7%; wet conditions show no effect.Source: D. Technical descending as a latent… · p. 4

E. Fitness–fatigue with learned time constants

  • The fitness–fatigue model remains weakly identified: its posterior is essentially the prior, and the full model cannot be distinguished from a fitness-only model.Without a seasonal term, it absorbs a 14% seasonal swing into fitness because riding volume is higher in summer.Source: E. Fitness–fatigue with learned time… · p. 4
  • For information-directed design, a planned deload followed by a standardized test climb most reduces uncertainty about τ2, whereas a sustained block most reduces uncertainty about τ1; conventional schedules may miss both contrasts.No real posterior is interpreted before recovery is verified on synthetic athletes.Source: E. Fitness–fatigue with learned time… · p. 4

V. THE EPICOACH AGENT

EPICOACH allocates proposals and explanations to language-model agents while deterministic tools fit models, score evidence, and enforce safety constraints. In practice, the safety gate overrode the highest-ranked plans when workload limits were exceeded.

What tools do

  • Deterministic tools fit models, compute evidence scores, and apply feasibility constraints, while the LLM proposes, selects, and explains within the weekly cycle.The Safety Officer can veto a plan, but the LLM cannot lift the veto.Source: §V · p. 5

What the LLM does

  • The Hypothesizer proposes testable hypotheses and the Modeler converts accepted ones into unit-tested model templates; a Critic checks identifiability and registry lineage.The Modeler cannot emit free-form model code, and the Hypothesizer–Critic pair iterates at most three times.Source: §V · p. 5

Information-directed design in practice

  • With acute load at 1.26× chronic, the Safety Officer vetoed ten top-ranked plans for ACWR, and the Designer settled on a conventional week with one test.The retrospective menu ranked an overload week followed by full rest and three repeated efforts highest at 0.17 nats, versus 0.16 for a three-week block-and-taper and at most 0.06 for conventional weeks.Source: §V · p. 5

Registry protocol

  • Registry updates require evidence from a logged statistical fit, and hypotheses can enter testing only after a prediction is registered before the specified ride.The registry logs each call and restricts refinement or merging to hypotheses with a verdict or at least one piece of evidence.Source: §V · p. 5

What tools do

  • The instrument ride separates a climb held at 160 ± 5 bpm from a descent ridden at self-selected effort, with lap presses marking both timed segments.The route warms up on Power Line Trail, climbs the Trans Canada Trail, traverses the campus ring road, and descends Nicoles Trail.Source: Figure 2 · p. 6

VI. EVALUATION

The evaluation uses eighteen months of one mountain biker’s wearable record and repeated trail segments, then tests next-week forecasts in a rolling-origin replay. The supplied passages specify the data, benchmark design, and comparison methods, but do not report the replay outcomes.

A. Data

  • The dataset comprises 223 rides over eighteen months, recorded with wrist heart rate, GPS, and barometric altitude, with no power meter or chest strap.Segment discovery yielded 24 segments and 458 traversals; non-ride activities contributed to daily load but not segment outcomes.Source: A. Data · p. 6
  • The instrument had 79 historical traversals, versus six on a rejected longer climb (5.9 km, +425 m, 47 min); the twice-weekly instrument also had to be near home.Across those rides, each +1 bpm in mean heart rate corresponded to about 3.6 s, so the fixed-heart-rate protocol removes effort-driven spread.Source: A. Data · p. 6

B. Retrospective replay benchmark

  • The rolling-origin replay uses only data available at each weekly origin to predict outcomes for the next seven days, after a four-month warm-up.Forecast targets include segment time, climb heart rate conditional on realized virtual power and temperature, within-ride drift, and the sign of weekly device VO2max change.Source: B. Retrospective replay benchmark · p. 6
  • Forecasts are assessed with proper scoring and calibration measures, while dependent-series comparisons use Diebold–Mariano, moving-block bootstrap, and Giacomini–White tests.The comparison tests use per-origin score differences with a four-week HAC lag; bootstrap intervals use 14-day blocks.Source: B. Retrospective replay benchmark · pp. 6–7
  • Replay baselines include persistence and moving averages, device outputs, a population-constant Banister model, Gaussian-process and sequence models, a PHIA-style agent, and a static twin.The stated ablations remove the identifiability gate or registry; active design is evaluated prospectively because the replay history is fixed.Source: B. Retrospective replay benchmark · p. 6

C. Prospective phase

The prospective phase tests the twin and agentic loop through a weekly cycle built around instrument rides and safety checks. Replay results favor the twin for heart-rate forecasting but not segment time, while live forecasts show a calibrated descent and a climb model that lags improvement.

C. Prospective phase

  • Each prospective week uses fixed fresh and loaded climb and descent instrument rides; the agent chooses intervening training from load patterns, with predictions sealed before rides and weekly refits after scoring.A ride scores only if it is the first traversal after sealing; a gate traversal without the lap protocol remains training data.Source: C. Prospective phase · p. 7

D. Synthetic recovery check

  • Synthetic recovery recovered daily performance state (r=0.996) and load-scale quantities within their 90% intervals, but not the fitness–fatigue time constants, whose posteriors stayed close to their priors.A multi-athlete simulator for information-accrual experiments is identified as future work.Source: D. Synthetic recovery check · p. 8

E. Results

  • The twin beat standard forecasters on conditional heart rate, but no method beat the exponentially weighted average on segment time; the agentic loop matched the static twin on heart rate and was marginally worse on segment time.The agentic system scored CRPS 7.00 vs 7.03 on heart rate and 0.050 vs 0.048 on segment time; the twin scored 7.03 bpm versus 7.87 for an exponentially weighted average on conditional heart rate, while segment-time CRPS was 0.048 versus 0.046 for the average.Source: E. Results · pp. 8–9
  • The live cycle sealed 20 predictions and scored 14; after a gate correction, five later descent predictions scored CRPS 0.040 log-s (4.0%), with residuals inside ±0.9.The first two descent forecasts were scored against a gate 49 m from the athlete’s lap-press line, then the gate and forecast method were corrected and logged.Source: E. Results · p. 9
  • Across three weeks, climb time at matched heart rate fell from 1046 s to 1003 s, while the twin forecast shifted only from 1060 to 1054 s and missed the pre-specified moving-average standard.The athlete was faster than the climb forecast in six of seven predictions; the moving average scored 31.5 s.Source: E. Results · p. 9
  • Fresh/loaded tests falsified the modeled fatigue penalty twice, but two weeks of designed contrast left the fitness time-constant belief unchanged because prior–posterior overlap remained 0.84.The refit estimated τ1 = 27 d [17, 59], P(τ1 < 42) = 0.80, and τ2 = 3.7 d [0.7, 15.9], while H4 remained at belief 0.5.Source: E. Results · pp. 9–10
  • The highest-information overload plan was vetoed in both live cycles because it exceeded workload limits, so the agent used the second-ranked plan.The overload was 130% of the four-week mean; the safety officer cited ACWR 1.35 > 1.3 and weekly load outside ±15%.Source: E. Results · p. 10
  • The registry preserves four score corrections, including reversed wrong-ride matches, and records the rules requiring a post-sealing traversal, one observation per prediction, and the lap protocol.The errors included date-level rather than timestamp matching, one traversal scoring two predictions, and a training ride claiming a test prediction.Source: E. Results · p. 11

VII. LIMITATIONS AND ETHICS

The study’s evidence is limited to one athlete, noisy wearable measurements, and a short prospective phase; the authors also caution that fitness–fatigue parameters remain poorly identified.

VII. LIMITATIONS AND ETHICS

  • The empirical study is one author-athlete, so it cannot establish generality from that athlete’s physiology.The authors base their generality argument on the method, benchmark, simulator, and a pipeline that can admit another athlete without code changes.Source: §VII · p. 11
  • Replay tests only the passive system; the active component’s short prospective evaluation has limited power for small effects.The paper states in advance which effects are detectable.Source: §VII · p. 11
  • Wrist optical heart rate and forest GPS are noisy, and quality controls and Student-t likelihoods mitigate but do not remove measurement error.Virtual power is a modeling input rather than a measurement; several baseline device metrics are proprietary.Source: §VII · p. 11
  • Fitness–fatigue parameters are poorly identified after eighteen months of observational data and may remain so after a short prospective block.The system is designed to report this uncertainty.Source: §VII · p. 11
  • The prospective work is self-experimentation by a healthy adult under hard safety gates, and the prototype is not medical advice.The authors say the design should not be read as safe for people with cardiovascular or other conditions without clinical oversight.Source: §VII · p. 11
  • The first three prospective weeks are a one-athlete shakedown, not a randomized comparison, and protocol changes were dated and logged.The phase had 14 scored predictions and two fresh/loaded contrasts; every live week so far was information-directed. The authors caution that the randomized comparison had not begun and that the short phase has limited power.Source: §VII · p. 12

VIII. CONCLUSION

The paper presents EpiCoach as an accountable learning loop for one mountain biker, with language models proposing and explaining while deterministic tools fit, score, gate, and record. Its evidence so far includes unresolved fitness time constants in eighteen months of wrist data and a twin that was too slow to follow the athlete during three weeks of designed rides.

VIII. CONCLUSION

  • EpiCoach commits probabilistic claims about one athlete to predictions before outcomes, then chooses observations and judges claims by their scores.The registry retains corrected and incorrect scores, making the loop auditable.Source: §VIII · p. 12
  • Language models propose mechanisms, choose instruments, negotiate with the athlete, and explain; deterministic tools fit, score, gate, and record.Source: §VIII · p. 12
  • Eighteen months of wrist data did not resolve a fitness time constant, and three weeks of designed rides showed the twin lagging the athlete’s improvement.The conclusion presents recording that uncertainty and retaining earlier scores as part of the kind of coach the authors seek.Source: §VIII · p. 12

APPENDIX A SEGMENT DISCOVERY

The appendices describe how repeated trail segments are discovered and filtered, then specify the twin, weekly updates, and constrained experiment-design procedures used in the study.

APPENDIX A SEGMENT DISCOVERY

  • Named OpenStreetMap ways are joined into trail chains, densified to 5 m, and chains shorter than 100 m are dropped.Paved roads are retained for discovery but excluded from heart-rate and fitness–fatigue models because virtual power is unreliable at road speeds.Source: Appendix A · p. 13
  • GPS samples snap to the nearest chain within 20 m, and runs require at least 30 samples, 20 snapped samples, and robust progress of at least 0.3 m/s.Gaps up to 30 seconds on a chain are bridged.Source: Appendix A · p. 13
  • Segments use the longest contiguous range of 10 m bins meeting a ride-coverage threshold and a 250 m minimum; median altitude change classifies climbs, descents, and rolling terrain.A bin count must reach max(5, 0.6 × rides touching the chain); climb and descent thresholds are above 15 m and below −15 m, respectively.Source: Appendix A · p. 13
  • Traversals run from the first entry-gate crossing to the subsequent exit-gate crossing and must meet coverage, gap, monotonic-progress, and duration checks.Accepted traversals require at least 85% of samples within the gates ±15 m, no gap over 90 seconds, monotone progress for at least 90% of steps, and duration of at least 60 seconds; laps can yield multiple traversals per run.Source: Appendix A · p. 13
  • Canopy GPS error can snap samples to parallel trails, and trails absent from OpenStreetMap or unnamed produce no segments.In the main area, 20% of ride samples were unmatched, mostly unsanctioned connectors; entry/exit jitter of about one second motivates preferring several-minute segments for hypothesis testing.Source: Appendix A · p. 13

APPENDIX B MODEL SPECIFICATIONS AND PRIORS

  • The twin separates descent skill from climb fitness–fatigue: descent time uses a latent skill state, while climb performance is modeled at matched heart-rate cost.Technical segments receive higher skill loadings; the fitness–fatigue component uses exponential load filters and learned time constants, with bridge-sampling Bayes factors and PSIS-LOO for comparing time-constant hypotheses.Source: Appendix B · p. 14, Figure 9 · p. 14
  • The twin is refit weekly after that week’s predictions are scored; climb and descent ledger forecasts draw from their respective model posteriors.The descent forecast advances the skill state one random-walk step and uses protocol-condition covariates.Source: Appendix B · p. 14
  • The experiment menu spans ride archetypes and scheduling options, then applies workload, rest-day, heart-rate, readiness, illness, injury, and availability constraints.The designer prunes the menu to 20–40 candidates before scoring; the feasibility set bounds ACWR and weekly load relative to target.Source: Appendix B · p. 14

APPENDIX C PROSPECTIVE N-OF-1 DESIGN AND POWER

The prospective design pairs information-directed and conventional weeks and uses two standardized instrument rides to create a within-week fatigue contrast. The planned analysis estimates within-person treatment effects and declares effects below 3% untestable in the prospective window.

APPENDIX C PROSPECTIVE N-OF-1 DESIGN AND POWER

  • Weeks are randomized in pairs so each fortnight contains one information-directed and one conventional week, with instrument-ride order counterbalanced.The two rides are scheduled on non-adjacent days; washout and baseline rules apply to sleep-extension and caffeine manipulations.Source: Appendix C · p. 14
  • Two weekly instrument rides use a climb target of 160 ± 5 bpm and a self-paced descent, with a fresh test after rest and a loaded test on day 5.The pair supplies the within-week fatigue contrast needed to separate τ2 from τ1; each ride carries about 60 exp-TRIMP, roughly 12% of weekly load.Source: Appendix C · pp. 14–15
  • The primary analysis is Bayesian regression of within-person log-speed effects; effects below 3% are designated untestable in the prospective window.A randomization test and a time-varying g-formula are secondary and sensitivity analyses; under the stated assumptions, a 4% effect has about 0.9 power and a 2% effect about 0.4.Source: Appendix C · p. 15

APPENDIX D REGISTRY SCHEMA, TOOL PROTOCOL AND PROMPTS

The registry stores hypotheses, predictions, evidence, fits, and events, while enforced invariants constrain belief updates, prediction timing, verdicts, and logging.

APPENDIX D REGISTRY SCHEMA, TOOL PROTOCOL AND PROMPTS

  • The schema tracks hypotheses and their lineage alongside predictions, scored outcomes, evidence, fitted models, and logged events.Source: Appendix D · p. 15
  • Beliefs change only through recorded evidence from a valid fit, with verdict thresholds, prediction-timing rules, and logged actors enforced by the store.Evidence is capped at |log BF| ≤ 10, and parameter-posterior evidence must pass an identifiability gate.Source: Appendix D · p. 15
  • The hypothesizer proposes at most three athlete-specific, falsifiable claims with mechanisms, formal specifications, and bounded prior log-odds.The allowed formal changes include adding covariates, interactions, changepoints, random walks, or condition splits.Source: Appendix D · pp. 15–16
  • The critic rejects changes with unidentifiable parameters, descent heart rate used as fitness, post-outcome conditioning, or an equivalent hypothesis in the lineage.Source: Appendix D · p. 16

APPENDIX E WORKOUT WRITE-BACK

The appendix describes workout write-back, a synthetic recovery check, and the activity and wellness data inventory.

APPENDIX E WORKOUT WRITE-BACK

  • The workout pusher serializes ordered cycling steps, schedules them through the vendor service, and keeps a FIT-file fallback for device delivery.The unofficial endpoint is validated against workouts created in the vendor interface; requests are paced, retried, and archived.Source: Appendix E · p. 16

APPENDIX F SYNTHETIC RECOVERY CHECK

  • The synthetic check simulates parameter, trail-network, ride-frequency, sensor-quality, and weather variation to assess recovery and prospective policy performance.It evaluates parameter-recovery coverage, information accrual, and held-out-week CRPS/ECE.Source: Appendix F · p. 16
  • Two of three 90% intervals covered the true τ1; τ2 and k2/k1 were recovered in all three runs, and latent performance correlated 0.95–0.99 with truth.The authors conclude that the design identifies performance but not the time constants.Source: Appendix F · p. 16

APPENDIX G DATA INVENTORY

  • The inventory contains 653 FIT-recorded activities, including 222 mountain-bike rides, with ride counts ranging from 1 to 22 per month.Across 556 days, sleep and overnight HRV completeness was 77%, while readiness and acute/chronic load completeness were 98%.Source: Appendix G · p. 16

Related papers

Loading 2610.40370v1…