Eye Tracking Dataset for 12-Lead ECG Interpretation

This release measures how clinicians read an ECG, not what an ECG contains. Sixty-three practitioners — from junior medical students to board-certified consultants — were each shown the same ten printed 12-lead ECGs for about 30 seconds apiece while an eye tracker recorded where they looked. What ships is the aggregated gaze behaviour:

• Datasets/Grid_Anonymized.csv — 14,742 rows, one per reader x image x area of interest, where the areas are the individual lead boxes, the quarters of each rhythm strip, and the printed settings footer.
• Datasets/Long_Short_Anonymized.csv — the same 630 sessions scored against a two-region split instead: Long (the rhythm strips) versus Short (the twelve short lead traces).
• ECGs/ECG_Images/ — the ten stimulus images (JPG/PNG).
• AOI_Distributions/ — two figures showing how the AOI grids were laid out over the normal-sinus-rhythm image.

Each row carries 17 gaze and fixation measures: time to first hit, total and percentage dwell time, revisit counts, fixation counts and durations. The research use is expertise modelling and attention prediction — which lead draws the eye first, and how that changes between a medical student and a consultant.

There is deliberately no eye_tracking_ecg config, and no ecgbench splits --dataset eye_tracking_ecg. Nothing in this release is an ECG recording in ECGBench’s sense. There is no sampled signal, no sampling rate, no signal_format that could name a JPEG, and no patient behind a record — the ten ECGs are printed exemplars shipped without waveforms or demographics. The unit of observation is a reader session, so a ten-fold partition over “records” would be partitioning ten pictures.

So ECGBench exposes it as a table provider instead, and leaves splitting to whoever defines a task on it — which is the right call anyway, since a reader study can be split by reader or by image and those are two different experiments.

from ecgbench.labels.eye_tracking_ecg import load_eye_tracking_ecg

df = load_eye_tracking_ecg("/path/to/eye-tracking-ecg/1.0.0/")
df.groupby(["Group", "aoi_lead"])["Time_spent_G_Percentage"].mean()

This is the same treatment PTB-XL+ gets, for a related reason: a dataset carrying no recordings of its own gets no fold assignment of its own.

Ventricular tachycardia is starred above because its 12 strip quarters carry only 11 distinct names — see the naming defects further down.

The grid AOI vocabulary is not self-explanatory, and AOI_Distributions/Grid_AOIs.png is what decodes it. On a full 25-AOI image:

• 1, 2, 3 are leads I, II and III. They are numbered rather than named, so they read like indices and the three limb leads appear to be missing entirely.
• aVR, aVL, aVF, V1–V6 are the remaining lead boxes.
• II-1…II-4, V1-1…V1-4, V5-1…V5-4 are quarters of the three rhythm strips along the bottom.
• Information is the printed footer (25mm/s 10mm/mV 150Hz 12SL …), which is not a trace at all — exclude it before averaging over leads.

Labels are scoped to their image: the same lead is V1 NSR on one and V1 AFib on another, giving 233 distinct labels for what are really ~25 regions. Grouping by Label therefore compares nothing across images. load_aoi_metrics() derives aoi_area, aoi_kind (lead / rhythm_strip / information) and aoi_lead (standard spelling, so 1 becomes I and V5-3 becomes V5) to make cross-image grouping correct.

Note the marked row in the table above: the hyperkalemia image is a 16-lead trace, adding V3R, V4R, V7 and V8. The catalogue’s leads: 12 describes the other nine images.

All 19 shipped files were verified against the release’s own SHA256SUMS.txt before anything below was computed, so the discrepancies are upstream rather than download damage.

The published demographics do not match the shipped CSVs. The paper’s Table 1 and the Group column agree on 63 participants and on eight of the ten categories, but two nurse categories are transposed and one participant’s recorded gender differs:

Derivation: load_respondents(), which takes the first Group/Gender/Age per Respondent_Name (all three are constant across each reader’s 244 rows). The totals agree, so this is a categorisation difference between the manuscript and the released file rather than missing participants. Use the CSV’s Group, since that is what the gaze rows are actually keyed to.

Age is unusable as shipped. Table 1 reports five age bands covering all 63 participants, but the CSV records a real age for only 9 of 63 — the other 54 carry 0, an anonymisation artefact rather than a value. Because it is 0 and not a blank, notna() reports the column 100% populated and a mean age comes out around 4 years. load_respondents() and load_aoi_metrics() convert it to NaN.

Two more sentinels, in the gaze columns. -1 means “this never happened”: Hit_time_G is -1 for the 2,086 AOIs (14.2%) a reader never looked at, and First_Fixation_Duration / Average_Fixations_Duration are -1 for the 3,654 with no fixation at all. Left in place they drag every mean down. Converted to NaN by default; pass sentinels_to_nan=False for the raw values. One related inconsistency is not auto-corrected, because fixing it would be inference rather than decoding: TTFF_F carries a plausible-looking time even on those 3,654 rows where there was no first fixation to time. Mask it on Fixations_Count > 0 yourself.

Sessions are not all 30 seconds. Duration is nominally ~30,000 ms but ranges 2,839–30,392 ms, and 139 of 630 ran under 29 s — the shortest is a consultant spending 2.8 s on Wolff-Parkinson-White. Compare the Time_spent_*_Percentage columns rather than absolute milliseconds, since the percentages are relative to each session’s own duration.

Two AOI naming defects. II-3 VTach names two different regions with different metrics — ventricular tachycardia has 25 AOI rows per reader but only 24 distinct labels, for all 63 readers — so aggregating by name silently merges them, and (reader, image, label) is not a unique key. load_aoi_metrics() adds aoi_occurrence to disambiguate. Separately, three complete-heart-block labels carry accumulated authoring-tool suffixes (II-2 CompleteHeartBlock copy, … copy copy, … copy copy copy), and the normal-sinus-rhythm image labels one strip quarter V3-3 where the sequence is otherwise V1-1, V1-2, V1-4 — that image has no V1-3 and no other V3 strip, so it is a typo upstream. ECGBench strips the copy suffixes, which is unambiguous, and reports V3-3 as written rather than rewriting data on a guess.