- for age binning, use 20 year bins [0, 20, 40, 60, 80, 300] <-- right edge 300 given variable age censoring > 90 for adults, dataset-specific quartiles otherwise (e.g. ZZU pediatric dataset)
- for stratification (by binned age, sex, labels), when there are multiple ECGs per patient, need to define patient-level variables. take the max for each class when grouping by the patient ID. the implication of this for age is that a patient may have more than one age assignment, semantically meaning that the stratifier tries to preserve the prevalence of each age bin.
- if no author provided initial train/val/test → use ptbxl stratifier to make splits stratified by binned age, sex, labels
- ensure train/val/test have at least 1 positive instance of all labels
- applied to CinC Georgia and ZZU datasets
- for train subsets, use ptbxl stratifier to make nested subsets stratified by binned age, sex, labels
- ensure each subset has at least 1 positive instance of all labels
| dataset | defines splits | has patient IDs |
|---|---|---|
| echonext | train/val/test | yes |
| ptbxl | folds | yes |
| mimic | train/val/test | yes |
| cinc | no | no |
| zzu | no | yes |