MAE, bias and error spread AI
Every signed error is prediction minus reference, in mmHg. A negative error is an underestimate. Three quantities summarize different aspects of the prediction:
| Measure | Calculation | Interpretation |
|---|---|---|
| Mean participant MAE | Average absolute error within each person, then average the 40 person-level values | Typical absolute error with equal weight per person |
| Bias | Mean signed error across all valid test windows | Average tendency to overestimate or underestimate |
| Error SD | Sample SD of signed errors around their pooled bias, using denominator n−1 | Error spread, including large failures |
For example, errors of −10 and +10 have zero bias despite both being substantial mistakes. Low bias is therefore insufficient on its own. Mean participant MAE gives a person with 16 valid windows the same weight as a person with 26; pooled bias and SD give each window equal weight.
Uncertainty and supplementary metrics
The report’s MAE intervals use 4,000 bootstrap resamples of whole participants, with fixed seed 20261024. Keeping each person’s beats together respects within-person dependence. The intervals are conditional on this cohort and the fitted models; they do not establish new-cohort or long-term accuracy. Paired comparisons likewise resample participants together across models.
RMSE is the square root of average squared error. It combines average bias and spread, and is retained in the detailed numerical outputs rather than duplicating the main table. Error coverage reports the fraction of pairs within 5, 10 or 15 mmHg.
For visual interpretation, see agreement plots. The 5 mmHg value in validation guidance refers to bias, not an MAE acceptance criterion.