Heart Reader

From Deep Learning to Edge AI for Automated 12-Lead ECG Diagnosis

Chaima Massaoudi — March 2026

A production-ready, multimodal deep learning pipeline for multi-label ECG classification of five diagnostic superclasses from 12-lead electrocardiogram recordings — featuring an InceptionTime1D + Squeeze-and-Excitation backbone fused with 1,313 PTB-XL+ structured features, an interactive web dashboard, and a complete edge deployment pipeline.

Quick Start • Demo • Architecture • Results • Dashboard • Edge AI • Full Report

---

Key Results

Macro-AUC 0.9268

Macro F1 0.7239

F-max 0.7574

AUPRC 0.8097

Parameters 977K

Model Size 2.37 MB

Per-Class AUC Breakdown (Test Set — Fold 10)

Class	AUC	90% CI	Description
NORM	0.957	[0.949, 0.964]	Normal ECG
MI	0.927	[0.915, 0.940]	Myocardial Infarction
STTC	0.930	[0.918, 0.941]	ST/T Change
CD	0.913	[0.902, 0.928]	Conduction Disturbance
HYP	0.908	[0.893, 0.924]	Hypertrophy

All classes exceed AUC 0.90, demonstrating robust classification across the full diagnostic spectrum.

---

Live Demo

Interactive dashboard — 12-lead ECG visualization, real-time diagnosis, and confidence scores (5× speed)
Watch full demo video →

---

Why Heart Reader?

Challenge	Our Solution
Manual ECG interpretation requires years of specialist training	Automated AI-powered diagnosis in milliseconds
Single-modality models miss complementary clinical features	Multimodal fusion of raw signals + 1,313 structured PTB-XL+ features
Large models can't run on portable medical devices	Edge-optimized pipeline: pruning + INT8 quantization + ONNX (2.37 MB)
Research models lack practical interfaces	Interactive web dashboard with real-time analysis & export
ECG diagnosis is multi-label (patients have multiple conditions)	Multi-label classification handling co-occurring pathologies

---

Architecture

Multimodal Late Fusion Pipeline

 ┌─────────────────────────────┐       ┌─────────────────────────────┐
 │     12-Lead ECG Signal      │       │     PTB-XL+ Features        │
 │     (1000 × 12 matrix)      │       │     (1313-dim vector)       │
 └──────────────┬──────────────┘       └──────────────┬──────────────┘
                │                                      │
   ┌────────────▼────────────┐            ┌────────────▼────────────┐
   │   InceptionTime1D       │            │     Feature MLP         │
   │   + SE Channel Attention│            │  [1313 → 256 → 128 → 64]│
   │   (6 blocks, multi-     │            │  BN + ReLU + Dropout    │
   │    scale: k=39,19,9)    │            └────────────┬────────────┘
   └────────────┬────────────┘                         │
                │                                 64-dim embed
   AdaptiveConcatPool1d                                │
   + Linear → 256-dim                                  │
                │                                      │
                └────────────┬─────────────────────────┘
                             │
                    Concat [256 ‖ 64] = 320-dim
                             │
                    ┌────────▼────────┐
                    │  Fusion MLP Head │
                    │ [320 → 128 → 5] │
                    │ ReLU + BN + Drop │
                    └────────┬────────┘
                             │
                      5-class sigmoid
                    (multi-label output)

Signal Backbone: InceptionTime1D + SE Attention

Component	Configuration
Inception Blocks	6 deep, residual every 3
Multi-scale Kernels	k=39, k=19, k=9 (captures patterns at multiple temporal resolutions)
SE Attention	Reduction ratio 16 (adaptive channel recalibration)
Pooling	AdaptiveConcatPool1d (avg + max)
Output	256-dim signal embedding

Training Enhancements

Technique	Configuration	Purpose
Focal Loss	gamma=2, alpha=0.25	Handle class imbalance (HYP is 4x rarer)
Label Smoothing	epsilon=0.05	Better calibrated confidence scores
Mixup	alpha=0.3	Data augmentation + regularization
SWA	From epoch 40	Converge to flatter minima
OneCycleLR	max_lr=0.003	Super-convergence scheduling
AdamW	wd=0.01	Decoupled weight regularization
AMP (FP16)	Automatic	2× memory reduction & speed
Gradient Clipping	max_norm=1.0	Training stability

ECG-Specific Data Augmentation

• Gaussian noise (sigma=0.05) — simulates sensor noise
• Random amplitude scaling (0.8–1.2×) — accounts for gain variation
• Lead dropout (up to 2 leads, p=0.1) — robustness to disconnection
• Baseline wander (sinusoidal drift, p=0.3) — common ambulatory artifact
• Temporal warping (interpolation, p=0.2) — heart rate variability robustness

---

Results

Test Set Performance (Fold 10, N=2,198)

All confidence intervals via 100 bootstrap resamples (5th–95th percentile).

Metric	Score	90% CI
Macro-AUC	0.9268	[0.9204, 0.9332]
F-max	0.7574	[0.7441, 0.7696]
Macro F1	0.7239	[0.7048, 0.7468]
Macro AUPRC	0.8097	[0.7946, 0.8256]
F-beta (beta=2)	0.7902	[0.7809, 0.8066]

Per-Class Classification Report

Class	Precision	Recall	F1	Support
NORM	0.82	0.95	0.88	954
MI	0.59	0.87	0.70	415
STTC	0.63	0.90	0.74	506
CD	0.67	0.82	0.74	496
HYP	0.38	0.74	0.50	222
Macro Avg	0.62	0.86	0.71	2,593

High recall across all classes (≥0.74) means the model reliably detects pathological patterns — clinically preferred over high precision alone.

Comparison with Published PTB-XL Benchmarks

Model	Macro-AUC	Input	Source
inception1d (benchmark)	0.925	Signal only	Strodthoff et al., 2021
xresnet1d101 (benchmark)	0.925	Signal only	Strodthoff et al., 2021
resnet1d_wang (benchmark)	0.930	Signal only	Strodthoff et al., 2021
Heart Reader (Ours)	0.927	Signal + PTB-XL+	This work

---

Quick Start

1. Clone & Install

git clone https://github.com/chaima-massaoudi/Heart-Reader-From-Deep-Learning-to-Edge-AI-for-Automated-ECG-Diagnosis.git
cd Heart-Reader-From-Deep-Learning-to-Edge-AI-for-Automated-ECG-Diagnosis

python -m venv .venv
.venv\Scripts\Activate.ps1          # Windows
# source .venv/bin/activate         # Linux / macOS

pip install -r heart_reader/requirements.txt

2. Download Datasets

Download from PhysioNet (free, requires credentialed access):

Dataset	Version	Path (relative to repo root)
PTB-XL	1.0.3	data/ptbxl/
PTB-XL+	1.0.1	ptb-xl-a-comprehensive-electrocardiographic-feature-dataset-1.0.1/

Expected directory layout

repo-root/
├── data/
│   └── ptbxl/
│       ├── ptbxl_database.csv
│       ├── scp_statements.csv
│       └── records100/...
├── ptb-xl-a-comprehensive-electrocardiographic-feature-dataset-1.0.1/
│   └── features/
│       ├── 12sl_features.csv
│       └── ecgdeli_features.csv
└── heart_reader/
    └── (this project)

3. Train

cd heart_reader

# Improved config (focal loss, mixup, SWA) — recommended
python train.py --config configs/improved.yaml --backbone inception1d

# Default baseline training
python train.py

4. Evaluate

python run_evaluation.py

5. Launch Web Dashboard

python run_server.py
# Open http://localhost:8000

6. Export for Edge Devices

python export.py --backbone inception1d

---

Web Dashboard

A full-featured, interactive FastAPI-powered medical dashboard for real-time ECG analysis.

Features

Feature	Description
Random Sample	Load a random test-set ECG with full-model predictions
File Upload	Upload your own 12-lead CSV for instant analysis
12-Lead ECG Viewer	Full waveform visualization across all leads (Chart.js)
Probability Chart	Interactive bar chart with per-class confidence scores
Diagnosis Panel	Clear Normal/Abnormal banner with detailed class breakdown
Graph Export	One-click PNG download of any visualization
Dark Medical Theme	Professional dark interface with ECG-green (#00df80) accents

Upload Formats

Format	Description
Signal-only	12 columns × 1000 rows (one column per ECG lead)
Signal + Features	>12 columns × 1000 rows (first 12 = leads, rest = PTB-XL+ features)

Pre-built Test Files

One sample CSV per diagnostic class is included in heart_reader/test_files/:

File	Expected Diagnosis	Confidence
sample_norm.csv	NORM	1.000
sample_mi.csv	MI	0.872
sample_sttc.csv	STTC	0.674
sample_cd.csv	CD	0.962
sample_hyp.csv	HYP	0.710

---

Edge Deployment

A complete compression pipeline for deploying on resource-constrained devices (Raspberry Pi, mobile, embedded medical hardware).

Pipeline

Full Model (3.79 MB)
    │
    ├── L1 Structured Pruning ──→ 50% filter removal
    │
    ├── INT8 Dynamic Quantization ──→ FP32 → INT8 weights
    │
    └── ONNX Export ──→ Cross-platform inference

Compression Results

Variant	Size	Compression	Format
Full Fusion (FP32)	3.79 MB	1.0×	PyTorch
Quantized (INT8)	2.37 MB	1.6×	PyTorch
Signal-Only ONNX	2.16 MB	1.75×	ONNX

Inference Latency (CPU)

Configuration	Mean	Std Dev
Full Fusion (FP32)	717 ms	±398 ms
Quantized (INT8)	827 ms	—

---

Data Protocol

Following the official PTB-XL benchmark (Strodthoff et al., 2021) — strict fold-based stratification with no patient overlap.

Split	Folds	Samples	Purpose
Train	1–8	16,289	Model training
Validation	9	2,034	Hyperparameter tuning, threshold optimization
Test	10	2,050	Final evaluation only (never used during development)

Diagnostic Superclasses

Class	Abbreviation	Test Samples	Multi-label?
Normal ECG	NORM	954	—
Myocardial Infarction	MI	415	Often + STTC
ST/T Change	STTC	506	Often + MI, HYP
Conduction Disturbance	CD	496	Can co-occur
Hypertrophy	HYP	222	Often + STTC

64% of records have exactly one label, 28% have two, 8% have three or more.

---

Project Structure

heart_reader/
├── configs/
│   ├── default.yaml              # Base configuration
│   └── improved.yaml             # Enhanced: focal loss + mixup + SWA
│
├── data/
│   ├── preprocessing.py          # PTB-XL loading, label mapping, PTB-XL+ features
│   ├── dataset.py                # PyTorch Dataset & DataLoader
│   └── augmentation.py           # ECG-specific augmentations
│
├── models/
│   ├── inception1d.py            # InceptionTime1D + SE attention
│   ├── xresnet1d.py              # XResNet1d family (18–152 layers)
│   ├── se_resnet1d.py            # SE-ResNet1d (Wang-style)
│   ├── fusion_model.py           # Multimodal signal + feature fusion
│   ├── feature_branch.py         # MLP for structured features
│   ├── ensemble.py               # Weighted ensemble (Nelder-Mead)
│   └── heads.py                  # Pooling, SE blocks, weight init
│
├── training/
│   ├── trainer.py                # Training loop (AMP, SWA, early stopping)
│   ├── losses.py                 # Focal loss, label smoothing BCE, mixup
│   ├── metrics.py                # AUC, Fmax, G_beta, AUPRC
│   └── callbacks.py              # LR scheduling callbacks
│
├── evaluation/
│   ├── evaluate.py               # Bootstrap evaluation, CI, CSV export
│   └── visualization.py          # ROC curves, confusion matrix, plots
│
├── edge/
│   ├── prune.py                  # L1 structured pruning
│   ├── quantize.py               # INT8 dynamic quantization
│   ├── export_tflite.py          # ONNX export
│   └── benchmark.py              # Inference benchmarking
│
├── frontend/
│   ├── app.py                    # FastAPI backend (REST API)
│   └── static/                   # Dashboard (HTML, CSS, JS)
│
├── results/
│   ├── evaluation_report.json    # Full metrics with bootstrap CIs
│   └── edge/edge_stats.json      # Compression & latency stats
│
├── test_files/                   # Pre-selected test CSVs per class
├── train.py                      # Main entry point
├── evaluate.py                   # Standalone evaluation
├── export.py                     # Edge deployment pipeline
├── run_server.py                 # Web dashboard launcher
├── demo.mp4 / demo.gif          # Video demonstration
├── REPORT.md                     # Detailed technical report (658 lines)
└── requirements.txt              # Python dependencies

---

Hardware & Environment

Component	Specification
GPU	NVIDIA GeForce RTX 3050 6GB
CUDA	12.6
PyTorch	2.10.0+cu126
Python	3.13
Training Time	~27 minutes (41 epochs, early stop at 26)
Seed	42 (full deterministic reproducibility)

---

References

1. Wagner, P., et al. "PTB-XL, a large publicly available electrocardiography dataset." Scientific Data 7 (2020). DOI
2. Strodthoff, N., et al. "Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL." IEEE JBHI 25.5 (2021). DOI
3. Strodthoff, N., et al. "PTB-XL+, a comprehensive electrocardiographic feature dataset." Scientific Data 10 (2023). DOI
4. Fawaz, H.I., et al. "InceptionTime: Finding AlexNet for time series classification." DMKD 34.6 (2020). DOI
5. Hu, J., et al. "Squeeze-and-Excitation Networks." CVPR (2018).
6. Lin, T.-Y., et al. "Focal Loss for Dense Object Detection." IEEE TPAMI 42.2 (2020).
7. Zhang, H., et al. "mixup: Beyond Empirical Risk Minimization." ICLR (2018).
8. Izmailov, P., et al. "Averaging Weights Leads to Wider Optima and Better Generalization." UAI (2018).

---

License

This project is for research and educational purposes. Please refer to the individual dataset licenses for data usage terms:

• PTB-XL License
• PTB-XL+ License

---

Heart Reader — Automated ECG Diagnosis from Signal to Edge
_{Built with PyTorch, FastAPI, and clinical rigor}