SomviaEstimate or measurement
A laboratory measures VO2 max with a gas analyser and lactate threshold with a blood sample. A watch on a wrist measures heart rate, movement and, on some models, blood oxygen. Everything else, in every app including this one, is a model fitted to those measurements.
Somvia says "estimated" wherever it means estimated. The reason is not modesty: an estimate has assumptions, and knowing them tells you when to trust the number. Each section below names the assumption.
The models and their sources
| What you see | Model | Published in |
|---|---|---|
| Session load | Training impulse, duration weighted by heart-rate intensity | Banister's training-impulse work on modelling athletic performance |
| Stress score from pace or power | Effort normalised against your own threshold | The training-stress formulations that came out of Coggan's testing work |
| Long-term and recent training load | Exponentially weighted averages of daily load, six weeks and one week | The impulse-response tradition following Banister |
| Threshold pace | Best sustained thirty-minute effort, or a twenty-minute effort scaled | Friel's thirty-minute field test and the twenty-minute protocol from Coggan's testing |
| Threshold heart rate | Average of the final twenty minutes of the best effort | Friel |
| Critical power | Hyperbolic fit across best power at several durations | Monod and Scherrer, Ergonomics, 1965 |
| VO2 max, submaximal route | Heart rate against speed, extrapolated to maximum heart rate | The submaximal extrapolation tradition after Åstrand and Ryhming, with the ACSM running equations for conversion |
| VO2 max, effort route, and race predictions | VDOT from a hard sustained effort, solved back for each distance | Daniels and Gilbert, Oxygen Power, 1979 |
| Percentile and fitness age | Your value placed in reference standards by age and sex | FRIEND reference standards, Kaminsky et al., Mayo Clinic Proceedings 97(2), 2022 |
| Recovery, sleep score, HRV status | Separate analyses of overnight readings, sleep duration and stages, and recent HRV against personal history | The heart-rate-variability monitoring literature; the combination and its weighting are Somvia's own |
Other Apple Watch recovery apps read different inputs and offer different training metrics. Five recovery apps compared lists what Athlytic, Bevel, Training Today, Gentler Streak and Somvia each read and calculate.
Where the computation happens
On your iPhone. Somvia reads authorized Apple Health data and computes scores locally. If you enable health data upload, selected records and scores are shared through your account so a connected AI Agent can read them over MCP. No language model is asked to produce a score, and none is asked to guess a number that the data does not support. When a value cannot be computed, the app leaves the row empty and says what is missing.
What each score needs
| Score | Needs | Appears after |
|---|---|---|
| Recovery | Overnight HRV, resting heart rate and personal history, with sleep and training load when available | Enough valid readings to establish the required personal comparisons |
| Sleep score | Sleep duration, stages and overnight readings | Each part has its own requirements; duration can use a recorded night, while personal comparisons need more history |
| HRV status | Recent overnight HRV and a longer personal history | Enough valid nights in both the recent and reference periods |
| Training load and status | Workouts with heart rate | Session load can use an individual workout; training status needs enough history to assess changes |
| Threshold pace and heart rate | A hard sustained effort with heart rate and speed | One qualifying effort |
| Critical power | Rides with a power meter at several durations | Three usable durations |
| VO2 max estimate | Runs with heart rate and speed | A few sessions |
| Race predictions | A qualifying hard effort and a maximum heart rate | One qualifying effort |
Limits worth stating plainly
- Wrist sensors are less reliable during high-intensity and irregular movement, and every downstream estimate inherits that.
- Reference standards for fitness percentile come from an American laboratory sample, so the percentile is an anchor, not a verdict.
- Estimates built from a twenty to thirty minute effort say less the further the target sits from that duration, which is why the marathon prediction assumes the most.
- Somvia is not a medical device. The scores are training information, not a clinical assessment, and nothing here diagnoses, treats or replaces a professional.
Questions
Does an AI model compute my scores?
No. Every score is computed on your iPhone by the models named on this page. A connected AI Agent reads the finished numbers and the data behind them, and can explain or compare them, but it never produces them.
Why do Somvia's numbers differ from Garmin's or Coros's?
Because every vendor builds and tunes its own version of these models, and each one scales the result differently. Compare a number with its own history rather than with a figure from another device.
Can I see which measurements produced a score?
Yes. Every score opens to the list of measurements it used, each with your own baseline next to it, and each metric has a method page in the app describing the approach and its inputs.
Is Somvia a medical device?
No. Somvia does not provide medical advice, diagnosis or treatment. Consult a qualified healthcare professional in addition to using Somvia and before making any medical decision.