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

Behnam Moradi

cs.ROOct 29, 2026

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

Authors: Behnam Moradi

Organizations: Software Department, KavAI California, USA

Abstract

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.

Figures & tables

Panel of Figure 1 — Fig. 1 : The weekly EpiCoach cycle. Purple nodes are LLM agents; teal nodes are deterministic tools; the red node is the safety gate; gold nodes are persistent stores. LLMs may register hypotheses and predict , but only the Statistician’s record_evidence , keyed by a stored fit_id , can move a belief. The Safety Officer’s constraints are code and cannot be argued around.

Fig. 2 : The instrument ride on Burnaby Mountain (one recording, coloured by heart-rate zone; satellite view from the vendor’s app). From the car park at lower left the athlete warms up along the Power Line Trail, climbs the Trans Canada Trail (a wide, steady-gradient gravel track under canopy) to University Drive, traverses the campus ring road, and descends Nicoles Trail, a narrow technical singletrack, back to the start. The climb is held at $160\pm 5$ bpm; the descent is ridden at self-selected effort. Lap presses at the four gates mark the two timed segments.

Fig. 3 : Instrument segments on Burnaby Mountain, all traversals to 22 October 2026. Left :

90

GPS traces on the Trans Canada Trail climb (blue), start gate (circle) and end gate (square) where the gravel meets University Drive. Right :

31

traces on the Nicoles descent via Poplar Trail (red); the start gate is the athlete’s lap-press point, not the OSM junction node, which sat

49

m off the ridden line. Grey dots: OSM corridor geometry used for the corridor check.

Fig. 4 : Justification for the fixed-heart-rate protocol. Left : moving time vs mean heart rate across

79

historical TCT traversals; the

-3.6

s/bpm slope shows effort explains most between-ride spread. The red band marks the

160\pm 5

bpm target. Right : uncontrolled time distribution (CV 9%) and the protocol prediction

\mathcal{N}(17.7,1.25^{2})

min; fixing heart rate is expected to collapse the spread to the model’s residual uncertainty.

	Segment time ( $n{=}81$ )			Conditional climb HR [bpm] ( $n{=}106$ )
Method	CRPS	Cov 90	DM $p$	CRPS	Cov 90	DM $p$
Persistence	0.052 [0.047, 0.060]	0.96	0.93	9.25 [7.50, 11.06]	0.88	$<$ 0.001
EWMA	0.046 [0.040, 0.055]	0.96	0.31	7.87 [6.48, 9.19]	0.89	0.003
Ridge on covariates	0.109 [0.081, 0.134]	0.83	0.004	8.06 [6.69, 10.07]	0.83	0.13
GP on time + covariates	0.097 [0.070, 0.126]	0.75	0.026	8.06 [6.88, 9.46]	0.70	0.004
Banister, population $\tau$	0.056 [0.049, 0.064]	0.91	0.21	—	—	—

TABLE I : Rolling-origin replay,

60

weekly origins, one-week horizon. CRPS (lower is better) with

95\%

moving-block bootstrap intervals;

90\%

-interval coverage; Diebold–Mariano

p

against the twin row of the same target (HAC variance, lag

4

). Segment time is scored in log-seconds. The “no segment instruments” and “no HR quality control” ablations listed in Section VI-B were not run.

Fig. 5 : Rolling-origin replay CRPS (lower is better) with

95\%

moving-block bootstrap CIs; DM test significance vs the static twin reference (green dashed line). The agentic system matches the static twin on both targets (n.s. on HR,

p{=}0.007

on segment time); the ReAct baseline is significantly worse on both.

Fig. 6 : Posterior distributions from the retrospective twin. Left : cardiac drift

\gamma_{d}

is clearly positive (H1, belief

95\%

). Centre : fitness time constant

\tau_{1}

remains wide despite

79

instrument traversals (H4, belief

80\%

); the prospective fresh/loaded pair is designed to resolve this. Right : switching from

6

Mountain Highway rides to

79

TCT rides tightened HR model noise from

6.5

5.3

bpm.

Fig. 7 : Belief trajectories for the eight most-tested hypotheses in the agentic replay (

60

rolling origins). Colour encodes state: grey = proposed, blue = under test, green = supported, red/orange = refuted/undetermined. Flat segments show hypotheses that reached a verdict; invariant I8 (no permanent verdicts) would reopen them if later evidence disagreed.

Fig. 8 : The prospective ledger, 1–23 October 2026. Horizontal bars are sealed predictions (mean; shaded

90\%

interval) drawn from the moment of sealing to the ride that scored them; dotted bars are predictions withdrawn before any outcome (I9). Filled circles are instrument rides (lap protocol), open circles the outcome as scored, crosses training rides through the gates (not scored). Dashed verticals: the descent-gate and forecast correction (4 Oct) and the two live refit-and-cycle events. Top : the climb forecast barely moves while the athlete’s time at

160

bpm falls by

\sim\!80

s. Bottom : descent forecasts before and after the instrument correction.

Hypothesis	Formal object	Belief	State	Ledger
H1 Cardiac drift $>0$	$\gamma_{d}>0$	0.95	under test	3/3 scored, CRPS 21.9 s
H2 Drift larger in heat	$\gamma_{dT}>0$	0.21	proposed	—
H3 Less sleep slows descents	$\delta_{\mathrm{sleep}}<0$	0.94	under test	7/7 scored, CRPS 0.082 (0.040 post-correction)
H4 Fitness constant $<42$ d	$\tau_{1}<42$	0.50	under test (gated)	4/4 scored, CRPS 54.1 s; falsifier fired $2/2$
H5 Aerobic response improved	$c_{w_{\mathrm{last}}}-c_{w_{\mathrm{first}}}<0$	0.14	proposed	—
H6 Climb HR rises with temperature	$\gamma_{T}>0$	0.43	proposed	—

TABLE II : Registry state on 22 October 2026 after three live cycles. Beliefs start at

0.5

and move only through Savage–Dickey evidence from stored fits, subject to the identifiability gate. H1–H5 were pre-specified; H6–H11 were registered by the agent during the live cycles. Ledger: predictions scored / sealed per hypothesis, with mean CRPS (climb in seconds, descent in log-seconds).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 11 — Fig. 9 : The two instruments in relief (satellite view, one ride, track coloured by heart-rate zone). Left : the Trans Canada Trail climb, $2.06$ km at $7.2\%$ on wide compacted gravel with no technical features: heart rate is a clean function of mechanical work here, which is why it serves as the fitness instrument. Right : Nicoles Trail, $1.16$ km and $-158$ m of tight, rooted singletrack: heart rate decouples from work and time is governed by skill, which is why it is modelled separately (Section IV ).

Trail (OSM name)	Area	Kind	Length [m]	$\Delta$ alt [m]	Rides	Moving [min]	CV [%]	HR
TCT, lower gate $\to$ summit road (declared)	A	climb	2061	+148	90	17.2	9.1	150
Nicoles via Poplar (declared)	A	descent	1159	-158	31	4.8	14.8	123
Dead Moped Trail	A	climb	692	+28	86	4.6	11.5	140
Upper Cut Trail	A	descent	353	-51	38	1.2	25.8	120
Arts Road	A	rolling	339	+12	31	1.6	18.5	141
North Road Trail	A	rolling	430	-13	30	1.9	18.3	127

Appendix

TABLE III : Discovered segments. Area A is the home hill (instrument network), B and C two mountain networks. Declared routes are gate-to-gate instruments; the rest are discovered. Moving time is the median over traversals; CV is the coefficient of variation of moving time.

Digital Twins

Digital twins are virtual models synchronized with real systems to support monitoring, prediction, simulation, and decision-making.

See all 257 papers

WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning
Jun 16, 2026Yuwei Zhang, Tong Xia, Bianca Emmerich +5Biomedical QA Wearable Sensing
University of Cambridge · Tsinghua University · University College London +2
Towards a General Intelligence and Interface for Wearable Health Data
May 21, 2026Girish Narayanswamy, Maxwell A. Xu, A. Ali Heydari +37Self-Supervised Pre-Training Clinical Outcome Prediction
1Google Research · University of Washington · University of Illinois Urbana-Champaign +2
WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data
Sep 4, 2026Ji Soo Lee, Xilun Chen, Pierce Chuang +5Medical QA Wearable Health Monitoring
Meta · KAIST · Korea University

Jun 16, 2026J/K move · Enter open · S save

WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning

Yuwei Zhang, Tong Xia, Bianca Emmerich, Yu Yvonne Wu+4

University of Cambridge · Tsinghua University · University College London+2

Biomedical QA Wearable Sensing LLM Agents Wearable Health Monitoring

Open paper AskPDF arXiv

Language models are remarkably capable at medical question answering, in some cases surpassing the accuracy of general physicians. However, answering questions about wearable health data remains challenging and understudied, as these ubiquitous sensors produce continuous, high-dimensional, and longitudinal data, which is non-trivial to align with text-centric distributions in LLM pretraining. The diversity of sensor modalities and user intents cannot be effectively handled by a fixed reasoning workflow or a single pretrained foundation model. To address these challenges, we propose WEQA, a query-adaptive agent framework that unifies LLM reasoning with specialized wearable analytical and modeling tools. An LLM controller is employed to synthesize execution plans and dynamically route each query to the appropriate combination of sensor analysis and pretrained models, and perform grounded response auditing with external knowledge. We also curate a benchmark spanning four open wearable datasets comprising analytic and predictive tasks in three different health domains. Experiments show that our framework is 24% more accurate than LLM and agentic baselines, and a blinded study with 12 medical experts and 8 users shows substantial gains in usefulness and clinical soundness.

Abstract

Digital Twins

Explore similar work

WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning

Towards a General Intelligence and Interface for Wearable Health Data

WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data