Lobachevsky University ECG Database (LUDB)

Format	12-lead · 10 s · 500 Hz · manually annotated waves
Patients	200
Records	200
Leads	12
License	ODC-By 1.0
Origin	Nizhny Novgorod City Hospital No. 5 / Lobachevsky University — Russia

Overview

200 twelve-lead ECGs from 200 patients at Nizhny Novgorod City Hospital No. 5, each 10 seconds at 500 Hz. One record per patient, so folds need no grouping.

What sets LUDB apart is its manual delineation: every record has 12 annotation files, one per lead, marking the onset, peak and offset of each P, QRS and T wave — 58,429 annotated waves in total. This is the reference dataset for delineation algorithms rather than classification.

ECGBench splits and loads the signals; it does not model the annotations. Read them directly where they sit, for example wfdb.rdann(f"{data_path}/data/1", "ii"), whose symbols are ( for a wave onset, ) for an offset, and p / N / t for the P, QRS and T peaks.

Lead names are lowercase in every header (i, ii, avr, v1 …). Select leads by name and the case is handled for you.

Rhythm distribution (all 200 records)

Rhythm	Records
Sinus rhythm	142
Sinus bradycardia	24
Atrial fibrillation	14
Sinus arrhythmia	7
Sinus tachycardia	4
Atrial flutter, typical	3
Irregular sinus rhythm	2
four further rhythms	1 each

Diagnosis categories — how much of the data each covers

Category	Records with a finding	Distinct labels
Rhythms	200	9
Electric axis of the heart	190	5
Hypertrophies	142	7
Conduction abnormalities	66	8
Ischemia	51	22
Non-specific repolarization abnormalities	49	6
Extrasystolies	14	16
Cardiac pacing	10	6
Other states	9	1

About those counts

Recomputed from ludb.csv, the structured source. The per-record .hea comments carry the same content flattened into free text and lose the category grouping, so the CSV is what ECGBench reads.

Two properties of that file matter, and both are handled for you:

Every string cell has a trailing newline, and multi-value cells join their values with newlines rather than a delimiter. Read raw, Ischemia looks like 40 classes with at most 4 records each; split into atomic labels it is 22 findings that co-occur. The counts above are atomised.
One age is not a number — record 34 is recorded as >89. The loader exposes age_raw verbatim alongside a numeric age that is NaN there.

Only Rhythms covers all 200 records. Everything below Conduction abnormalities is sparse enough that a classifier has very little to learn from: Other states has a single label across 9 records, and Extrasystolies has 16 labels across 14 records — more labels than records. primary_rhythm pools rhythms under 10 records into OTHER so 10-fold stratification is defined; train on rhythms, which is verbatim.

Validation summary (500 Hz)

Version	Records	Note
original	200	all records, with is_valid + quality_issues
clean	200	100% pass rate — no check fired on any record
excluded	0

Building the splits

ecgbench splits --dataset ludb --data-path /path/to/ludb/1.0.1/

Loading with ECGBench

from ecgbench import ECGDataset

ds = ECGDataset(
    "ludb",
    split="train",
    data_path="/path/to/ludb/1.0.1/",
    labels=True,
)

sample = ds[0]
sample["signal"].shape              # (12, 5000), millivolts
sample["labels"]["rhythms"]         # ['Atrial fibrillation']  — multi-label
sample["labels"]["hypertrophies"]   # ['Left atrial hypertrophy', ...]
sample["labels"]["electric_axis"]   # 'left axis deviation'
sample["labels"]["age_raw"]         # '57'  (one record is '>89')

# Lead names are lowercase here; selecting by name is case-insensitive.
two = ECGDataset("ludb", split="train", data_path="/path/to/ludb/1.0.1/",
                 leads=["II", "V5"])
two.lead_names                      # ('ii', 'v5')