Heartbeats-To-Heatmaps: An Intelligent Heart Disease Diagnosis System Integrating Clustering, Ensemble Models, and Neural Networks

The Heartbeats-To-Heatmaps project developed by MuhammadNoor7 integrates unsupervised clustering, ensemble learning, and lightweight CNN to build an efficient and interpretable intelligent heart disease diagnosis system. The system uses a lightweight architecture that can run on ordinary CPUs, lowering the deployment threshold. It also includes MNIST recognition verification and a Streamlit visualization dashboard, providing auxiliary diagnostic support for areas with limited medical resources.

Cardiovascular disease is the number one health killer globally. Traditional diagnosis relies on subjective experience-based judgment, which has problems such as inconsistent standards and high missed diagnosis rates. This project builds an end-to-end data mining system that integrates multiple machine learning technologies, with both high accuracy and strong interpretability. Its lightweight design is suitable for promotion in resource-limited areas, helping with early diagnosis and intervention.

The system includes three modules: data preprocessing, multi-model fusion, and visualization. Data preprocessing uses clustering methods like K-Means to identify patterns and anomalies; the model fusion layer adopts a Stacking ensemble strategy, integrating base learners such as Random Forest and XGBoost with meta-learners; the deep learning layer converts electrocardiograms into heatmaps, uses lightweight CNN to capture deep features, and after optimization, it can perform real-time inference on CPUs.

The system provides feature-level explanations through SHAP value analysis, displaying SHAP force plots and global feature importance; decision path tracking makes the decision process of the ensemble model transparent. The Streamlit dashboard supports real-time prediction, batch analysis, and model monitoring, helping doctors understand model decisions and track performance changes.

The project focuses on CPU optimization, achieving real-time inference on ordinary CPUs/edge devices through model quantization, OpenVINO acceleration, etc. Application scenarios include outpatient screening, physical examination report interpretation, telemedicine support, and medical education training, adapting to different clinical needs.

The open-source project provides a full-process reference, helping medical AI researchers and engineers; limitations include data dependency and the need for clinical validation. Future plans include integrating multi-modal data, developing mobile applications, and establishing a federated learning framework to continuously improve the system's capabilities.

This project proves that the design of multi-technology integration, interpretability priority, and deployment orientation can enable AI to play a value in resource-constrained environments. Its modular structure and engineering practices provide references for the medical AI field, promoting technology popularization and application.