Skip to content

modelcard.md

Latest commit

 

History

History
102 lines (68 loc) · 7.4 KB

modelcard.md

File metadata and controls

102 lines (68 loc) · 7.4 KB
Raw

Model Card: ALADIN

Model summary

ALADIN is a neuro-symbolic AI model for ECG preprocessing, delineation (segmentation), and diagnosis. It combines a deep-learning segmentation backbone (nnU-Net based, C++/PyTorch hybrid backend) with a symbolic logic engine that reasons over the segmented waveform to produce rhythm diagnoses and explanations. It supports 1-, 3-, and 12-lead ECG input and has been validated on ECGs ranging from 6-second clinical recordings to 24-hour ambulatory sessions.

Model architecture

ALADIN is a pipeline of four stages (see aladin/src/aladin/__init__.py):

  1. Preprocessing — C++ backend (aladin/src/*.cpp) filtering and signal conditioning.
  2. Segmentation (UNetSegmenter, aladin/src/aladin/backend) — an nnU-Net-based model (built on the vendored nnUNet/nnunetv2 and dynamic-network-architectures) that delineates P, QRS, and T waves, and flags abnormal QRS, atrial fibrillation, and noise regions. Two pretrained checkpoints are shipped:
    • 1-lead model — used when only lead II is available.
    • 3-lead model — used when leads II, V1, and V6 are all available (falls back to the 1-lead model otherwise). Rhythm/logic reasoning always uses lead II regardless of which segmentation model is selected (aladin/src/aladin/configuration.py).
  3. Self-reflection (Reflection, aladin/src/aladin/selfreflection) — beat clustering and morphology-based correction of the raw segmentation output.
  4. Symbolic diagnosis (LogicEngine, aladin/src/aladin/logicengine/logic.py) — a rule-based logic engine that reasons over the corrected delineation to produce rhythm diagnoses with human-readable explanations (see aladin_explanation.txt for example outputs).

Diagnoses currently supported by the logic engine: normal sinus rhythm (NSR), atrial fibrillation (AFIB), first/second-degree AV block including Wenckebach (AVB), complete heart block (CHB), supraventricular tachycardia (SVT), ventricular tachycardia (VT), idioventricular rhythm (IVR), ectopic atrial rhythm (EAR), junctional rhythm, bigeminy/trigeminy, premature atrial/ventricular contractions (PAC/PVC), and noise detection. Each is individually togglable via the customarrhythmia argument.

Intended use

  • Primary use case: Research use for automated ECG delineation (P/QRS/T segmentation, quality assessment, beat classification, beat clustering, feature extraction, median beat extraction, etc) and rhythm-level diagnosis support across clinical (MUSE), ambulatory (ZioPatch), and handheld (KardiaMobile) ECG recordings.
  • Primary users: Researchers and developers working on ECG signal processing, arrhythmia detection, and cardiology AI tooling.
  • Out of scope: ALADIN is a research tool and has not been cleared or approved as a medical device (e.g. no FDA/CE clearance). It is not intended for standalone clinical diagnosis or treatment decisions, and outputs should not be used as the sole basis for patient care without review by a qualified clinician.

Training data

The segmentation model was trained using an active-learning selection process across nine public ECG datasets, with separate internal and external hold-out sets used for delineation validation, and three additional cohorts used for diagnosis validation.

Training set (𝒰)

Dataset Records Patients Duration Frequency (Hz) Leads Type Selected (records)
BUT-PDB 50 50 2 min 360 2 Ambulatory 61
CHAPMAN 10,247 10,247 10 sec 500 12 Clinical 392
CPSC2018 6,877 6,877 6–60 sec 500 12 Clinical 398
CPSC2018 extra 3,453 3,453 10 sec 500 12 Clinical 341
GEORGIA 10,344 10,344 10 sec 500 12 Clinical 371
INCART 72 32 30 min 257 2 Clinical 90
LANCET 828 828 10 sec 500 12 Both 622
NINGBO 34,905 34,905 10 sec 500 12 Clinical 150
PTB-XL 22,353 18,869 6–60 sec 500 12 Clinical 637
Subtotal 89,129 85,605 3,962

Validation data

Internal delineation validation (𝒱del,intern)

Held-out subsets (no overlap with the training set) from CPSC2018 (210 records), CPSC2018 extra (186), ICENTIA (81, 24h ambulatory, 1-lead, 250 Hz), LANCET (72), and NINGBO (248).

External delineation validation (𝒱del,extern)

Dataset Records Patients Duration Frequency Leads Type
RDB 2,399 2,399 10 sec 500 12 Clinical

Combined internal + external delineation validation total: 3,196 patients.

Internal diagnosis validation

Not applicable (no training was involved, all validation is external)

External diagnosis validation (𝒱diag)

Dataset Records Patients Duration Frequency Leads Type
iRhythm Zio™ Monitor 328 328 30 sec 200 1 Ambulatory
AliveCor KardiaMobile (CinC 2017 Challenge) 8,528 8,528 6–60 sec 300 1 Ambulatory
CardioSTAT Long-term ECG (ICENTIA, cleaned) 157,508 4,924 70 min 250 1 Ambulatory

Diagnosis validation total: 13,780 patients.

Reproduction

Benchmark reproduction instructions and scripts are in benchmark.md and paper/. Please request a Huggingface token from the authors to run the benchmark reproduction as the reproduction uses datasets that are not openly available anymore. Published comparisons include:

  • Delineation performance against Jimenez-Perez et al.'s DelineatorSwitchAndCompose model on the RDB and internal validation sets.
  • Diagnosis performance against ECGFounder, a ResNet baseline, and average cardiologist performance, on the Stanford (iRhythm) and CinC (AliveCor) diagnosis validation cohorts (see paper/boxplot-stanford.py, paper/boxplot-cinc.py).

Note: the data/STANFORD and data/CINC folders present in this working copy (used by benchmark_diagnosis_STANFORD.sh / benchmark_diagnosis_CINC.sh) are evaluation/benchmark-reproduction data, not training data.

Limitations

  • Rhythm/logic diagnosis always reasons over lead II only; providing more leads improves the segmentation model choice (1-lead vs 3-lead) but not which lead drives diagnosis.
  • The 3-lead model requires leads II, V1, and V6 specifically; any other lead combination beyond lead II alone falls back to the 1-lead model with a warning.
  • Diagnosis validation cohorts are single-lead, ambulatory-leaning (Zio, KardiaMobile, ICENTIA); performance on multi-lead clinical-only rhythms outside the validated diagnosis set is less characterized.

Ethical considerations & clinical status

ALADIN is a research tool and is not a certified medical device — it has not received FDA, CE, or other regulatory clearance. It is intended for research and clinical decision-support exploration, not for standalone diagnosis or treatment decisions. Any clinical use must involve review by a qualified healthcare professional.

How to use

See README.md for installation and usage instructions.