ECG Effects of Ranolazine, Dofetilide, Verapamil, Quinidine (ECGRDVQ)

5,232 ten-second 12-lead ECGs at 1 kHz from 22 healthy volunteers in a randomised, double-blind, placebo-controlled 5-period crossover Phase I trial (study SCR-002, ClinicalTrials.gov NCT01873950) run by the US FDA’s Center for Drug Evaluation and Research.

This is the first of the three FDA studies in the catalogue — ECGRDVQ is SCR-002, ECGDMMLD is SCR-003 and the CiPA validation study is SCR-004 — and it is the one that established the measurement the other two were built to test.

The question: can J-Tpeak tell two kinds of QT prolongation apart? A drug that blocks only the hERG potassium channel is proarrhythmic. A drug that blocks hERG and an inward current — late sodium, or L-type calcium — prolongs QTc just as much but carries far less risk. QTc alone cannot distinguish them. The claim under test is that J-Tpeak, the early-repolarisation half of the QT interval, can.

Every subject received all five treatments, as a single oral dose per period with washout in between:

Records were extracted in triplicate at 16 nominal timepoints from half an hour before dosing to 24 hours after, each paired with a plasma concentration draw, and each is published twice: as the raw 10 s segment and as a derived 16-channel median beat carrying semi-automatic fiducial annotations.

This is a pharmacology dataset, not a diagnosis dataset. Every participant was screened to exclude cardiac disease. There is no rhythm, morphology or arrhythmia label anywhere in it, and none of the usual ECGBench habits built around a diagnostic class transfers — starting with stratification, which uses the treatment.

In ECGDMMLD the treatment arms are staged combinations — the second agent was dosed hours after the first, so a record labelled Mexiletine + Dofetilide at the 2-hour timepoint contains no dofetilide at all, and roughly a third of that dataset’s labels do not describe what is in the blood.

Nothing like that applies here. Each of the five periods administered a single agent, so treatment names both the randomisation arm and the drug. Measured against the plasma column, 93–94% of every active arm’s records carry a measured concentration of exactly the drug the label names.

One much smaller caveat survives, and it is the same one every crossover has: 327 records are pre-dose. They sit at timepoint_hours = −0.5, are flagged is_baseline, and carry their period’s drug name while containing no drug — a “Dofetilide” record that is a drug-free ECG. At timepoint_hours = 0.5 absorption is still incomplete. Filter on is_baseline, or use plasma_concentration_ng_ml, before treating treatment as exposure.

Because this is a near-complete crossover: 21 of the 22 subjects passed through all five arms, so nearly every subject appears in every row of the table above. Two consequences a parallel-group dataset would not have:

• No split of this dataset separates the arms. Folds are grouped on patient_id (they must be — see the clustering section), and each patient carries four or five treatments, so every fold receives all five automatically. Per-arm balance across folds is a property of the design, not of the stratifier. Treat “all 5 arms in all 10 folds” as expected rather than as evidence of anything.
• Every treatment comparison rests on the same 22 people. There is no independent control group. That is a strength for the study — a within-subject comparison is far more sensitive — and a limit on generalisation.

Subject 1002 is the single exception: it withdrew after 4 of the 5 periods, so it has 192 records instead of 240 and never received quinidine. That is why quinidine has 21 subjects and 1,008 records where every other arm has 22 and 1,056 — and placing that one lopsided subject is essentially the only work stratifying on treatment does here.

Two levels of clustering sit between the record count and anything you could call an independent sample.

Triplicates, and here they are exact. Three ten-second segments were extracted per subject per nominal timepoint, and all 1,744 timepoint groups hold precisely 3 records — 5,232 = 1,744 × 3 with no exceptions. (ECGDMMLD manages 1,403 of 1,404; CiPA is raggeder still.) They are the same person, in the same posture, at the same plasma concentration, seconds apart: near-duplicates, not repeats of an experiment. The effective sample size is closer to 1,744 than to 5,232.

Subjects. Those 1,744 groups come from 22 people, 240 records each except subject 1002’s 192. So the whole dataset is 22 individuals, and any subject-level quantity has n = 22.

Folds are grouped on patient_id, which handles both at once: verified on the shipped release, no subject spans two folds and none of the 1,744 timepoint triplicates is ever split. Group your own analyses the same way, and weight by subject before quoting any rate.

The median beat adds the vector-magnitude lead VCGMAG and the Frank vx, vy, vz components to the 12, and carries a .atr annotation file. Both directories are 1 kHz and use the same record ID, so they are two representations of one acquisition, not two records.

The median beat’s length is not fixed here. It runs 968 to 1,876 samples across 667 distinct values, where ECGDMMLD’s is a constant 1,200 — so never assume a shape, and never read a fiducial as a fraction of the beat.

The medians get no fold of their own, deliberately. Every median beat is a derivation of a raw record ECGBench already partitions; generating a second ten-fold split over the same recordings would let someone train on one and evaluate on the other. ECGDataset therefore always reads raw/, and the median beats reach you as paths plus a fiducial loader.

from ecgbench import load_config
from ecgbench.labels.ecgrdvq import load_median_beat_fiducials

config = load_config("ecgrdvq")
fid = load_median_beat_fiducials("/path/to/ecgrdvq/1.0.0/", config)
fid.loc["0003bfaf-41e9-4411-a1c2-a216330cbdf6"]
# p_onset_ms 133, qrs_onset_ms 310, qrs_offset_ms 404,
# t_peak_ms 607, t_peak_secondary_ms NaN, t_offset_ms 688, n_annotations 5

The fiducials and the interval table are not independent measurements. The annotations are what the published intervals were measured from, and they reproduce the clinical table exactly: recomputed over every median beat in the release, PR matches for all 5,221 un-wrapped records, QRS and J-Tpeak for all 5,223, QT and Tpeak-Tend for the 5,219 with a T offset, and Tpeak-Tpeak′ for all 42 with a secondary T peak — every one to the millisecond (for the record above: 310−133 = PR 177, 404−310 = QRS 94, 688−310 = QT 378, 607−404 = J-Tpeak 203, 688−607 = Tpeak-Tend 81). Treat agreement between them as a format check, never as corroboration.

All 26,137 files match the release’s own SHA256SUMS.txt, so none of these is download damage. Every one is a missing value rather than a wrong one, once the third is repaired.

1. Nine records have no median beat. medians/ holds 5,223 of the 5,232. Because every interval in the clinical table was measured from the median beat, those nine rows have no PR, QRS, QT, J-Tpeak or Tpeak-Tend. rr_ms is present for all nine (it comes from the raw rhythm strip) and — inconsistently — two of them still carry erd_30_ms/lrd_30_ms/twave_amplitude_uv and all nine carry twave_asymmetry/twave_flatness. So the median beat clearly was computed upstream and simply was not published. Their raw/ records are intact, so this costs the split nothing; filter on median_beat_available.

2. Four records have no T-offset annotation, so qt_ms and tpeak_tend_ms are NA and their .atr carries 4 marks instead of 5. All four are subject 1004 on quinidine at 2.5–3.5 h, and they hold the four flattest T waves in the release — twave_amplitude_uv 69.7–78.5 µV against a median of 490.5. Quinidine flattened the T wave until its end could not be marked. With defect 1, qt_ms is NA for 13 records.

3. Two records store PR as a 32-bit arithmetic wrap, -4294966951 and -4294966972. Both are subject 1007 on verapamil at 1.0 h, and both are exactly the two records whose .atr has no P-onset mark — the P onset fell before the start of the median-beat window, so an unsigned subtraction wrapped. Adding 2³² recovers 345 ms and 324 ms:

It is the only physiologic residue, it is corroborated by the triplicate’s third record (PR 293 ms — the highest un-wrapped PR in the release) and by verapamil’s expected AV-nodal PR prolongation, and 293 → 324/345 is exactly the trend subject 1007’s other verapamil timepoints trace. ECGBench repairs both and flags each with pr_ms_repaired, so the repair is auditable rather than silent. It is also why pr_ms has a maximum of 345 in the table below.

4. Two records have a dead V4 lead, held at a constant −0.00625 mV for all 10,000 samples. Both are subject 1019 on quinidine at 14.0 h and both are among the nine of defect 1 — the flat lead is presumably why no median beat was derived. These two are the only records excluded from clean/.

Not a defect, but unexplained: twave_asymmetry and twave_flatness are NA for 129 records, spread across all five treatments and all 22 subjects with no pattern, and not concentrated on the flattest T waves (their amplitudes run 81.7–960.7 µV). The release does not say why.

Recomputed from ecgbench.labels.ecgrdvq over all 5,232 records of the shipped v1.0.0, not copied from the paper.

† The PR maximum of 345 ms is a repaired value — one of the two 32-bit wraps described above. The highest PR the release stores un-wrapped is 293 ms.

tpeak_tpeakp_ms is missing for 5,190 records because a secondary T peak genuinely is not present in them, not because the measurement failed. That is the opposite of ECGDMMLD, where the same column is empty in all 4,211 rows and no annotation marks a second peak anywhere in the release. A pipeline written against that sibling will silently drop a real measurement here.

Every other “missing” count traces to defect 1, 2 or the unexplained 129 — see the defects section for which is which.

All values in ms, recomputed over all 5,232 records from ecgbench.labels.ecgrdvq. “Pre-dose” is the mean over each arm’s is_baseline records; “Cmax” is the mean over its top decile of plasma concentration of the analyte named in the last column (96–102 records per arm; for placebo, which has no analyte, all 990 post-dose records).

This is a descriptive summary, not the study’s analysis. The published result fits a concentration–response model to placebo-corrected change from baseline, and its effect sizes differ from these.

Read every arm against placebo before concluding anything. Placebo is not flat: it drifts −7.6 ms on QTcF and −16.8 ms on J-Tpeak over the same hours, so the raw J-Tpeak column above understates every drug and the naive reading (“ranolazine and verapamil shorten J-Tpeak”) is partly artefact. Subtracting the placebo row is what the study does, and it is what makes the pattern appear:

Now the result is legible, and it is the finding the whole CiPA programme rests on. Every one of the four drugs prolongs QTcF — by +17 to +95 ms, a range in which the ordering tells you nothing about risk. J-Tpeak splits them in two: the predominant-hERG blockers (dofetilide +37, quinidine +24) prolong it, and the multichannel blockers (ranolazine +6, verapamil −8) do not. Tpeak-Tend separates them the same way, and even more sharply.

That is the point of the dataset: QTc says all four drugs are dangerous; J-Tpeak says two of them are not. This is a descriptive reconstruction rather than the published model, but the split it recovers is the study’s finding — and it is the measurement ECGDMMLD (SCR-003) and CiPA (SCR-004) were then built to confirm.

The study’s result is change from baseline, not the absolute interval. In the sibling CiPA release those change values exist only on adeg.csv’s triplicate-average rows, which carry a blank record ID and therefore cannot be attached to any waveform. Here nothing is lost.

BASELINE = Y flags the three pre-dose (timepoint_hours = −0.5) records of each (subject, period) pair, and all 109 pairs have one — 327 baseline records over 109 pairs. So a baseline is just their mean, and the delta is a per-record quantity:

from ecgbench import load_config
from ecgbench.labels.ecgrdvq import load_baseline_deltas

df = load_baseline_deltas("/path/to/ecgrdvq/1.0.0/", load_config("ecgrdvq"))
post = df[~df["is_baseline"]]
post.groupby("treatment")["delta_jtpeak_ms"].mean().round(1)
# Dofetilide          +2.4
# Placebo            -17.7
# Quinidine Sulph    -11.6
# Ranolazine         -20.4
# Verapamil HCL      -22.3

Note the placebo row: J-Tpeak falls by 18 ms on placebo alone. These raw deltas are dominated by diurnal drift, which is precisely why the published analysis is placebo-corrected and why you must be too. (These figures average over all post-dose timepoints, so they are smaller than the Cmax-restricted ones in the table above; both are in the example script.)

Baselines are per period, not per subject — each crossover period has its own pre-dose triplicate, and sharing one across periods would attribute a washout drift to the drug.

What is not done for you is placebo-correction, which needs the placebo arm’s mean across subjects at the same nominal timepoint. That is an analysis decision rather than a label, so it is left to you — and it is why the figures here do not equal the published ones.

1. The signals are millivolts. Every channel of every record declares its own gain against unit /mV — per-lead and per-record, e.g. 149792(-19192)/mV for lead I next to 40562.6(-22753)/mV for lead II of the same record, each fitted to that channel’s range. wfdb.rdrecord applies them, so signal_unit_scale is 1.0, matching ECGDMMLD. CiPA declares /uV and needs 0.001; copying that value here divides every sample by 1000. units="uV" multiplies by 1000 if you want the microvolt scale.

2. The 1 kHz is interpolated. Acquisition was at 500 Hz on a Mortara Surveyor; the depositors up-sampled the extracted segments to 1000 Hz. Every record is 10,000 samples and the headers say 1000 Hz, but there are only 5,000 measurements in there. Anything sensitive to true bandwidth — high-frequency QRS content, derivative-based delineation — should treat this as 500 Hz data.

3. twave_amplitude_uv is microvolts while the waveforms are millivolts. The source measures it on the median beat’s vector-magnitude lead; the range is 66.6–1,021.5 µV.

4. The plasma column mixes two units, and so does the dose. The PK table is long rather than wide, because one agent per period means one measurement per record: plasma_analyte names it, plasma_concentration is the value, plasma_concentration_unit its unit. Dofetilide is reported in pg/mL and quinidine, ranolazine and verapamil in ng/mL, so groupby("treatment").plasma_concentration.mean() compares numbers 1000× apart in scale. plasma_concentration_ng_ml is the same quantity in one unit, provided so that mistake is avoidable rather than merely documented. dose carries the same hazard: 500 for dofetilide is µg, the other three are mg — never sort or compare dose without dose_unit.

A missing concentration is NA and means either “not dosed / not drawn” or “below the limit of quantification” — this release has no censoring flag and no zeros to distinguish the two. Nine records name an analyte but carry no value (three triplicates at the 0.5 h timepoint).

Standard order, spelled with an uppercase A — AVR/AVL/AVF, like PTB-XL and its sibling ECGDMMLD. Verified identical in all 5,232 raw/ headers, and the medians/ headers agree.

This is the opposite of CiPA, whose raw/ headers write aVR/aVL/aVF while its medians/ headers write AVR/AVL/AVF — so CiPA disagrees with both siblings and with itself. Matching in ECGDataset(leads=...) is case-insensitive, so leads=["aVL"] resolves in all three; lead_names records what the files actually say. Never index leads positionally across datasets — signal[4] is aVL here and in both siblings, but aVF in MIMIC-IV-ECG.

flat_line is the only check that fires, and this is the mirror image of ECGDMMLD, where amplitude_outlier is the only one that does.

amplitude_outlier never fires here because the release is unusually clean on amplitude: scanned over all 5,232 records × 12 leads × 10,000 samples the observed range is only −4.86 to +6.98 mV, the 99.9th percentile of per-record peak absolute amplitude is 3.75 mV, and just 2 records exceed 5 mV. The house default [-10, 10] mV therefore excludes none of them.

flat_line excludes exactly two, and they tell a coherent story:

Both are two of the three records of a single triplicate, and both are among the nine records with no published median beat — a dead V4 is presumably why no median beat could be derived from them. The third record of that triplicate is fine and stays in clean/.

Nothing else is wrong with the release’s signals: no record has a NaN sample, a missing lead or a wrong sample count (all 5,232 headers declare exactly 10,000 samples at 1000 Hz), and every one of the 5,232 raw/ headers parses. The minimum per-lead variance across the release is 9.8 × 10⁻³¹ mV² — that is the dead V4 — while every other lead-record sits far above the check’s 10⁻⁶ threshold.

Fold sizes are uneven, and patient grouping is why. 22 subjects over 10 folds is 2–3 per fold while records per subject are 240 (or 192 for subject 1002), so folds run 480–720 records rather than a uniform 523. That is the correct trade — the alternative leaks triplicates — but it means a per-fold metric is computed over a fixed and very small number of people. Eight of the ten folds are two subjects. Fold 10, the test set, is two subjects: 1003 and 1005. Any result quoted on one fold of this dataset is a result about two or three individuals.

Subject 1002, the early withdrawal, lands in fold 1 alongside two completers, which is why that fold is 672 rather than 720. That placement is the one thing stratifying on treatment accomplished here.

All five arms appear in all ten folds — 0 empty cells of 50. Eight folds hold 96 records per arm; folds 1 and 4, the three-subject folds, hold 144 — except quinidine in fold 1, which holds 96 because subject 1002 never received it. As explained above, that completeness follows from the crossover design and not from the stratifier.

Fold membership is identical between original/ and clean/; clean/ is a row subset, differing only in fold 4 where the two dead-V4 records sat.