# Multi-label ECG Classification: An Automated Cardiac Disease Diagnosis System Based on the PTB-XL Dataset

> This article introduces a multi-label cardiac disease classification system based on the large-scale PTB-XL ECG dataset, integrating signal preprocessing, feature engineering, deep learning model fusion, and other technologies to achieve automated diagnosis of six common cardiac diseases.

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-06-16T00:45:56.000Z
- 最近活动: 2026-06-16T00:55:02.021Z
- 热度: 135.8
- 关键词: 心电图分类, 多标签学习, 心脏疾病, 深度学习, 医疗AI
- 页面链接: https://www.zingnex.cn/en/forum/thread/ptb-xl
- Canonical: https://www.zingnex.cn/forum/thread/ptb-xl
- Markdown 来源: floors_fallback

---

## Introduction: Core Overview of the Multi-label ECG Classification System Based on the PTB-XL Dataset

This article introduces a multi-label cardiac disease classification system developed by Ashritha Kotagiri from the University of North Florida, based on the large-scale PTB-XL ECG dataset. It integrates signal preprocessing, feature engineering, deep learning model fusion, and other technologies to achieve automated diagnosis of six common cardiac diseases. The system provides an end-to-end solution for automated cardiac disease diagnosis, aiming to improve diagnostic efficiency and accuracy.

## Research Background and Introduction to the PTB-XL Dataset

Cardiac diseases often involve multiple coexisting pathological states. Traditional single-label classification cannot capture comorbidity phenomena, while multi-label classification is more in line with clinical reality. The PTB-XL dataset is one of the largest publicly available ECG datasets: it contains 21,799 12-lead records, 18,869 patients, dual sampling frequencies (100Hz/500Hz), and multi-label diagnostic annotations, providing a solid foundation for model training. Data source: https://physionet.org/content/ptb-xl/

## System Architecture and Processing Flow

The end-to-end system flow includes: 1. Signal preprocessing (noise filtering, baseline drift elimination, normalization); 2. PQRST waveform detection (Christov algorithm); 3. Feature engineering (extracting 132 time-domain, frequency-domain, and morphological features such as PR interval, QRS energy, etc.); 4. Multi-label classification models (traditional ML: Logistic regression, Random Forest, etc.; deep learning: 1D CNN, LSTM, CNN+BiLSTM, etc.).

## Model Performance Comparison Results

Among traditional ML models, Random Forest performed the best (accuracy 0.893, F1 score 0.660); among deep learning models, 1D CNN was optimal (accuracy 0.924); the M4 fusion model integrates multi-source information and achieves an AUROC of 0.9510, showing excellent discriminative ability. For detailed performance metrics, refer to the tables in the original article.

## Clinical Application Value and Limitations

Potential value of the system: improving screening efficiency, reducing missed diagnosis rates, assisting primary care, and standardizing diagnosis. Note: AI systems are currently auxiliary tools, and the final diagnosis requires decisions by professional physicians.

## Future Research Directions

Future research directions include: introducing Transformer architecture, explainable AI (SHAP/LIME), real-time monitoring (deployment on wearable devices), integration with clinical decision support systems, and cross-dataset generalization validation.