PTB-XL¶

PTBXL (registry name ptbxl) loads PTB-XL (Wagner et al., 2020; PhysioNet ptb-xl 1.0.x): 21,799 ten-second 12-lead ECGs at 500 Hz (records500/) and 100 Hz (records100/), with diagnostic statements, demographics, and reports. Optionally it joins PTB-XL+ (Strodthoff et al., 2023) 12SL measurements.

data:
  dataset:
    name: ptbxl
    params:
      root: /path/to/ptb-xl/1.0.1
      sampling_rate: 500        # or 100
      leads: standard12         # stored order: I, II, III, aVR, aVL, aVF, V1–V6
      scale: zscore
      labels: all               # scp vocabulary, or null
      ptbxl_plus_root: /path/to/ptb-xl-plus/1.0.1   # optional: adds "features"

Splits¶

The recommended stratified folds: strat_fold 1–8 train, 9 val, 10 test (override with folds: {train: [...], val: [...], test: [...]}). All normalizers are fit on the training folds.

Condition fields¶

Field	Value	Encoder to use	Switch
`scp`	multi-hot statements	`multilabel` (`num_labels` = vocabulary size)	`labels`: `all`, `superclass`, `subclass`, `diagnostic`, `form`, `rhythm`, or `null`
`demographics`	normalized `[age, height, weight]` + missing mask	`numeric` (`num_features: 3`)	`demographics: true`
`sex`	class (0 male, 1 female)	`categorical` (`num_classes: 2`)	`sex: true`
`report`	report text (mostly German)	`text` or `text_hash`	`report: true`
`features`	normalized PTB-XL+ 12SL measurements + missing mask	`numeric` / `fourier`	`ptbxl_plus_root`, `feature_columns`

Vocabulary sizes: all 71, superclass 5 (NORM, MI, STTC, CD, HYP), subclass 23, rhythm 12; the order is dataset.label_names. min_likelihood drops statements below a likelihood (PTB-XL stores 0–100; the default 0 keeps every listed statement, as SSSD-ECG does). Height and weight are often missing — the numeric encoder sees the mask.

Default PTB-XL+ features: HR_Ventr_Global, PR_Int_Global, QRS_Dur_Global, QT_Int_Global, QT_IntBazett_Global, P_AxisFront_Global, R_AxisFrontal_Global, T_AxisFront_Global.

Example configs¶

configs/experiments/ptbxl_sssd_ecg.yaml — SSSD-ECG-style: 100 Hz, 8 independent leads, 71 statements, S4D diffusion backbone.
configs/experiments/ptbxl_text_demographics.yaml — flow matching conditioned on the report (frozen multilingual BERT), demographics and sex.