AI Health Models
Structured physiology. The tokenizer for heart rate.
Foundation models in health train on raw sensor traces—noisy, unlabeled, unstructured. TrueZone converts the same stream into labeled, physiologically-grounded features that actually teach the model something.
per-step
session
Tokenization
Same timeline. More columns.
TrueZone runs alongside the live stream. Every incoming sample becomes a token: the same (hr, speed) row, augmented with instantaneous zone, fatigue, substrate split, target pace, and time-to-exhaustion—plus a session-level fit when the activity ends.
# session_raw.csv — 1 Hz sampling, 48 min timestamp,hr,speed 0, 142, 9.2 1, 148, 10.1 2, 151, 10.8 3, 155, 11.4 4, 159, 11.9 5, 162, 12.3 6, 164, 12.5 7, 166, 12.6 8, 168, 12.8 9, 169, 12.9 10, 170, 13.0 # ... 2,870 more rows ... # no labels. no thresholds. no physiology.
# tokens.csv — per-step physiology (1 Hz) timestamp,hr,speed,zone,fatigue,fat_ox,carb_ox,target_v,tte_s 0, 142, 9.2, Z2, 1.2, 62, 38, 11.1, 3891 1, 148, 10.1, Z2, 1.8, 58, 42, 11.1, 3402 2, 151, 10.8, Z3, 2.4, 54, 46, 13.2, 2998 3, 155, 11.4, Z3, 3.1, 51, 49, 13.2, 2601 4, 159, 11.9, Z3, 3.9, 48, 52, 13.2, 2312 5, 162, 12.3, Z3, 4.8, 45, 55, 13.2, 2089 6, 164, 12.5, Z3, 5.7, 43, 57, 13.2, 1923 7, 166, 12.6, Z4, 6.6, 41, 59, 15.9, 1781 8, 168, 12.8, Z4, 7.5, 39, 61, 15.9, 1658 9, 169, 12.9, Z4, 8.4, 37, 63, 15.9, 1552 10, 170, 13.0, Z4, 9.3, 35, 65, 15.9, 1461 # ... 2,870 more rows ... # + session_summary (post-session fit) { "E": 0.74, "Vmax": 19.2, "P": 192, "LT1": 12.4, "LT2": 15.1, "Fatmax": 9.8, "MFI": 0.82, "drift_bpm": 8.4, "session_kcal": 487 }
Each per-step token is produced by computeRealtime(hr, speed, t, E, Vmax, P)—every field derived from the same ODE model, not learned from data.
The analogy
Raw HR is to physiology what characters are to language. TrueZone features are the tokens—each one a unit of physiological meaning the model can compose with.
A tokenizer doesn't do the reasoning. It prepares the input so the model can. Everything downstream—training, prediction, audit—improves when the input units carry meaning instead of noise.
Every dimension has a physiological meaning.
E = 0.75 means something specific about aerobic capacity. A neural embedding dimension does not. Clinicians, researchers, and regulators can audit exactly what the model is learning—and trace any prediction back to the physiological features it rested on.
Trained once. Works across devices, populations, and sports.
Features derived from an ODE mean the same thing whether the data came from an Apple Watch, a Garmin chest strap, or a clinical ECG—from a 25-year-old runner or a 65-year-old walker. The feature layer absorbs the device-specific noise. The model sees physiology, not hardware.
Any existing HR archive becomes a labeled dataset.
Sessions already sitting in databases—at Garmin, Apple, Strava, research institutions—can be processed through TrueZone to generate physiological labels after the fact. You do not need to collect new data. You need to label the data you already have.
The unlock
Your HR archive is already a labeled dataset.
Billions of sessions sit on a billion wrists. The data is collected. The labels have been missing. Process the archive through TrueZone and every session gains 60+ physiological features—no new collection, no ground-truth campaign, no device change.
The Feature Layer
60+ physiological features per session. All from heart rate.
TrueZone fits a physiological ODE to each individual's heart rate response, extracting three base parameters. From those, it derives a complete physiological profile—every feature grounded in the model's physics, not learned from data.
Base parameters
Endurance (E), Max Speed (Vmax), HRmax (P), confidence intervals, variance matrix
Thresholds
LT1, LT2, Fatmax, VO₂max speed, speed thresholds across the V-scale, heart rate at each threshold
Energy & metabolism
RMR, session calories, fat/carb oxidation, fat max rate, carb footprint, substrate balance, TEF, TDEE
Zones & load
6 HR zones, 3 metabolic zones, 12 fiber zones, training load (oxidative/glycolytic split), glycolytic ratio
Recovery
HR recovery curve, tau constants, EPOC (fast/slow), muscular recovery hours, time to HR recovery
Cardiac drift
Drift magnitude, drift rate, metabolic debt, drift percentage
Performance
Race predictions (100m–marathon), PrimeScore, Cooper test, VO₂max, training paces
Use Cases
Where structured physiology unlocks new models.
Metabolic health prediction
Train models to predict glycemic response, insulin sensitivity, or metabolic syndrome risk using MFI, RZI, and substrate balance as input features instead of raw HR. MFI already correlates R² = 0.99 with glycemic markers.
Personalized training AI
Recommendation engines that prescribe training based on individual endurance, threshold positions, and recovery state—not population averages. Features update with every session.
Clinical digital biomarkers
Physiological features as endpoints in clinical trials. Track metabolic flexibility, aerobic decline, or intervention response continuously from wearable data instead of periodic lab visits.
Population health surveillance
Process millions of existing HR records into structured physiological profiles. Identify metabolic risk trends at population scale. Every wearable user becomes a data point with physiological meaning.
Integration
SDK or API. HR in, physiology out.
Process sessions through the REST API or run the SDK on your own infrastructure. Input: heart rate + speed/power time series. Output: 60+ structured physiological features per session, Bayesian parameter history, and confidence intervals. No training data, no GPU, no population model.
Send HR data
Heart rate + speed/power time series. Any device, any format. Historical archives work too.
ODE fitting
Bayesian parameter extraction. E, Vmax, P converge per individual. Confidence intervals included.
Get features
60+ physiological features returned per session. Structured JSON. Ready for your ML pipeline, clinical dashboard, or recommendation engine.
Stop training on noise. Start training on physiology.
TrueZone is available as a C++ SDK, REST API, or NPM package. Patent-protected to 2045. Compatible with any HR data source.