CiPA ECG Validation Study

5,749 ten-second 12-lead ECGs at 1 kHz from 60 healthy volunteers in a double-blind randomised Phase I trial (study SCR-004, ClinicalTrials.gov NCT03070470) designed by the US FDA’s Center for Drug Evaluation and Research.

The study exists to validate the CiPA hypothesis: that measuring QTc together with J-Tpeakc in an ordinary Phase I study can tell a “balanced” ion-channel blocker — one that blocks hERG but also blocks late sodium or L-type calcium current at comparable potency, and so carries little proarrhythmic risk — apart from a “predominant hERG” blocker, which does not. QTc alone cannot: both prolong it.

• Part 1 — 50 subjects (1001–1050), parallel design, ten per group, three consecutive days of ranolazine, verapamil, lopinavir+ritonavir, chloroquine or placebo.
• Part 2 — 10 subjects (2001–2010), two-period crossover of dofetilide against diltiazem+dofetilide, dosed on days 1–3 and 8–10.

Records were extracted in triplicate at up to 28 nominal timepoints from one hour before dosing to 73 hours after, 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 drug in effect.

Two levels of clustering sit between the record count and anything you could call an independent sample, and both matter more here than in most datasets.

Triplicates. Three ten-second segments were extracted per subject per nominal timepoint — 1,916 of the 1,917 timepoint groups hold exactly 3 records (one holds 1). 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 1,917 than to 5,749.

Subjects. Those 1,917 groups come from 60 people, 24 to 168 records each (median 84), because subjects completed different numbers of timepoints — most 28 per period, but one only 8. So every per-record statistic is weighted by trial compliance rather than by person.

Folds are grouped on patient_id, which handles both at once: verified on the shipped release, no subject spans two folds and no timepoint triplicate 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 X, Y, Z 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 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.

Their headers look corrupt — 6276255.687397709(-1227133513)/uV — and are not: each channel is scaled to fill the int32 range, and wfdb.rdrecord recovers physiologic microvolts. Spot-checked against the raw records, median lead II peak-to-peak comes out at 0.87–0.93× the raw value, which is what median-beat averaging predicts.

from ecgbench import load_config
from ecgbench.labels.ecgcipa import load_median_beat_fiducials

config = load_config("ecgcipa")
fid = load_median_beat_fiducials("/path/to/ecgcipa/1.0.0/", config)
fid.loc["00689D31-8491-4643-B3C8-45241FBBD47C"]
# p_onset_ms 165, qrs_onset_ms 353, qrs_peak_ms 396,
# qrs_offset_ms 431, t_peak_ms 663, t_offset_ms 724

The fiducials and the interval table are not independent measurements. The annotations are what the published intervals were measured from, and they reproduce adeg.csv exactly: across all 5,749 records, every one of PR, QRS, QT, J-Tpeak and Tpeak-Tend recomputed from the fiducials equals the published value to the millisecond (for the record above: 353−165 = PR 188, 431−353 = QRS 78, 724−353 = QT 371, 663−431 = J-Tpeak 232, 724−663 = Tpeak-Tend 61). Treat agreement between them as a format check, never as corroboration.

adeg.csv holds 69,556 rows, and only 51,686 belong to a record. The other 17,870 carry DTYPE=AVERAGE and a blank EGREFID — they are the mean over each triplicate, and they are the only rows where the study’s actual endpoints exist:

The published exposure-response analysis — ΔQTcF and ΔJ-Tpeakc against concentration — therefore lives one level of aggregation above the signals, and no join can bring it down. ECGBench exposes the per-record absolute intervals, and the averages separately, keyed by (patient_id, period, timepoint_n):

from ecgbench import load_config
from ecgbench.labels.ecgcipa import load_triplicate_averages

avg = load_triplicate_averages("/path/to/ecgcipa/1.0.0/", load_config("ecgcipa"))
len(avg)                                           # 17870
qtcf = avg[(avg["parameter"] == "QTCF") & avg["CHG"].notna()]
len(qtcf), round(qtcf["CHG"].max(), 1)             # (1776, 171.5)

If you want a per-record endpoint, build it yourself from used_for_baseline (627 pre-dose records) and say so — it will not equal the published figures, which average the triplicate first.

All values in ms, recomputed over all 5,749 records from ecgbench.labels.ecgcipa. “Pre-dose” is the mean over each arm’s records at nominal_hours_from_period_start <= 0; “Cmax” is the mean over its top decile of plasma concentration (for placebo, which has no analyte, every post-dose record).

This is a descriptive summary, not the study’s analysis. The published result fits a concentration–response model to placebo-corrected change from baseline on the triplicate averages, for the reason given in the previous section, and its effect sizes differ from these.

Even so the pattern the study was built to detect is visible: dofetilide, the predominant-hERG positive control, prolongs QTcF by ~32 ms while barely moving J-Tpeakc; verapamil and ranolazine, the balanced blockers, do not prolong QTcF at all and shorten J-Tpeakc substantially. Adding calcium block to dofetilide (Diltiazem+Dofetilide) cuts its ΔQTcF and pushes J-Tpeakc negative. That divergence between the two intervals is the CiPA biomarker.

Read every arm against placebo, whose −9.8 ms sets the diurnal drift: absolute changes here are not drug effects until that is subtracted.

1. The signals are microvolts. Every one of the 5,749 headers declares gain 0.26595744680851063(0)/uV, so wfdb.rdrecord returns µV and signal_unit_scale is 0.001. Left at 1.0, every record’s peak reads as ~2000 “mV” and amplitude_outlier fires on the whole dataset. units="uV" gets the source scale back.

2. Dofetilide plasma concentration is pg/mL; the other six analytes are ng/mL. The source says so in PCSTRESU and nothing rescales it, which is why the columns are named for their own unit (plasma_dofetilide_pg_ml against plasma_ranolazine_ng_ml). Pooling them numerically is a 1000× error.

3. A plasma concentration of 0 means “below the limit of quantification”, not “no drug”. 263 of adpc.csv’s 1,934 rows carry LLOQFL=Y with AVAL=0; every unflagged value is ≥ 1.24, and nothing else in the file is 0. plasma_below_lloq names the censored analytes — but filter on the boolean plasma_any_below_lloq, because the list column is "" for uncensored records and pandas reads that empty string back from a CSV as NaN, so != "" matches every row of a re-read frame.

4. There are two time axes and they disagree by design. timepoint (“54 hrs”, numeric in nominal_hours_from_period_start) counts from the period’s first dose; nominal_hours_from_reference (NRRLT) counts from that day’s reference dose. A record on study day 3 reads 54 and 6 for the same instant, because 48 + 6 = 54. Neither is wrong; picking the wrong one stacks three dosing days onto one time axis.

Two smaller ones, both handled and regression-tested:

• race ships with leading spaces — " WHITE", " ASIAN" — so a groupby on the raw column produces categories that look identical and are not. ECGBench strips them.
• EGREPNUM disagrees with itself in 4 records, where the interval parameters are numbered one apart from HR/RR/QRS. ADTM is identical within each, so the record is unambiguous — only its index within the triplicate is not. replicate_number is anchored on the HR row (present for all 5,749) and replicate_number_inconsistent flags those four.

All 28,752 shipped files were verified against the release’s own SHA256SUMS.txt before any figure on this page was computed — all OK, so everything here is an upstream property rather than download damage. The adeg.csv checksum recorded in manifest.json (3ab7b3cd…) is the provider’s own published value.

Record and subject counts match the PhysioNet landing page exactly at 5,749 and 60, as does the 5,749-record median-beat set. There is nothing to reconcile — unusually for this catalogue.

Every other figure is recomputed from the release’s four analysis datasets (adeg.csv, adpc.csv, adsl.csv, addm.csv) and the WFDB headers. The release ships no table mapping records to signal files, so ECGBench builds one (ecgbench_metadata.csv) on the first ecgbench splits run; see ecgbench/labels/ecgcipa.py.

Cohort, recomputed over the 60 subjects: age 19–50 (mean 31.7, SD 8.7), 38 male / 22 female, race 32 Black or African American / 24 White / 4 Asian, BMI 19.0–30.0 (mean 25.8), weight 56.4–99.8 kg, height 156.5–188.0 cm. Recordings span 2017-03-27 to 2017-06-26.

Interval measurements are incomplete for 19 records, for two distinct reasons. 10 have no pr_ms because no P onset could be annotated; 9 have no qt_ms, qtcf_ms, jtpeak_ms, jtpeakc_ms or tpeak_tend_ms because no T annotation could be placed. hr_bpm, rr_ms and qrs_ms are complete for all 5,749. Both groups are NaN, never 0, and the counts agree exactly with the missing fiducials in the .atr files.

Nothing on this page is a diagnostic class. treatment is a fold-balancing label about a randomised assignment, and it applies to a subject’s whole period — so a pre-dose record in the dofetilide arm is a drug-free ECG that treatment calls "Dofetilide". Filter on nominal_hours_from_period_start or on the plasma concentration before treating the label as exposure.

The amplitude distribution is tight — the 99.9th percentile of per-record peak absolute amplitude is 4.15 mV, and only 4 records exceed 5 mV — so the default amplitude_range_mv of ±10 does real work here rather than passing everything:

Both are confined to one electrode group, which is what an electrode artefact looks like; neither is physiology. Two further records reach 5–7 mV and pass. No record in the release has a NaN sample, a missing lead or a flat lead — the minimum per-lead variance across all 5,749 × 12 channels is 5.6 × 10⁻⁴ mV² — so amplitude_outlier is the only check that fires at all.

StratifiedGroupKFold on patient_id, stratified on treatment. Verified on the shipped release: no subject spans two folds, and none of the 1,917 timepoint triplicates is ever split. The default 8/1/1 mapping gives 4,604 train / 588 val / 555 test (clean; 4,606 / 588 / 555 for original).

The arithmetic works out well but is thin in places. 60 subjects over 10 folds is exactly 6 subjects per fold, and each treatment has exactly 10 subjects — so all 7 treatments reach all 10 folds, at 84 records per treatment per fold in 61 of the 70 cells:

The thin cells — Verapamil in fold 5 (24 records) and Diltiazem+Dofetilide in folds 8 and 10 (51 each) — are why fold sizes run 528–588 rather than a uniform 575, and they come from the 24-record subject and the crossover arm’s shorter second period. A per-arm result read off one fold can rest on as few as 24 records from a single subject, so check the row before quoting it. StratifiedGroupKFold emits none of the “least populated class” warnings StratifiedKFold does, so the silence is not evidence of balance.

One structural limit worth stating: no split of this dataset separates Dofetilide from Diltiazem+Dofetilide by subject. The ten crossover subjects contribute a block of each, and grouping by subject necessarily keeps both in the same fold.

For a stricter harness, rotate the held-out fold rather than using the default 8/1/1 mapping:

from ecgbench import ECGDataset

# Hold out fold 3, train on the rest. With split=None each sample's
# ["split"] reports that record's own default split.
train = ECGDataset("ecgcipa", split=None,
                   fold_numbers=[1, 2, 4, 5, 6, 7, 8, 9, 10],
                   data_path="/path/to/ecgcipa/1.0.0/")
held  = ECGDataset("ecgcipa", split=None, fold_numbers=[3],
                   data_path="/path/to/ecgcipa/1.0.0/")