🎯 Overview

System Arch

Cardiovascular diseases are the leading global cause of death. ECGs are essential for diagnosis, but manual interpretation is slow and expert-dependent. This project — my final master’s thesis in Data Science and AI — introduces a real-time, AI-powered ECG analysis system.

The system:

Uses SPADE-based multi-agent architecture
Automates segmentation (U-Net) and classification (Random Forest)
Supports real-time monitoring through a React-based web interface
Handles noisy signals and full-signal diagnosis
Achieves high accuracy

It’s modular, scalable, and ready for future clinical integration and explainable AI.

🧠 making the project

🧠 System Architecture

System Arch

Component	Technology
Multi-Agent System	SPADE
Segmentation Model	1D U-Net
R-peak Detection	1D CNN
Classification Model	Random Forest
Backend	Django REST + WebSockets
Frontend	React

🔁 SPADE Multi-Agent Architecture

Multi-agents system

Modular, Flexible and Distrubuted
Asynchronous communication between agents using XMPP protocol and json messages
A supervisor Agent (BPMN-style orchestrator)

⚙️ Acquisition & Preprocessing

Bandpass filtering (noise and drift removal)
Smoothing and resampling to 250 Hz
Z-score normalization
Output: Clean signals ready for ML models

🩻 Signal Segmentation

Goal: Detect P, QRS, and T waves
segmentaion

📊 Datasets

QTDB: 105 signals, 15 min each
LUDB: 200 signals, 10 sec each

🧠 Models Compared

1. CNN-LSTM

CNN-LSTM

Accuracy: 94.39%
Loss: 0.1444
Issues in noisy/overlapping signals

2. TCN (Temporal Convolutional Network)

TCN

Accuracy: 95.63%
Loss: 0.1091
Better results, still not perfect in noisy data

3. U-Net (Selected Model)

UNet

1D architecture with encoder-decoder
Weighted focal loss
Accuracy: 94.13%
Loss: 0.0107

4. Results

CNN-LSTM:
TCN:
UNet:

🔧 Post-processing

by applying some function we ( remove_uncompleted_first_last_wave, merge_close_waves, remove_irrelevant_waves, check_repeated_waves, fix_P, fix QRS)
post-process