Research Demonstration

Pulse
Sage

A comprehensive deep learning framework for diagnosing hypertrophic cardiomyopathy using multi-modal data with explainable AI integration

Try Demo Read Paper

Sampurn Kumar · Divya Raj · Sarvagya Singh · Aditya Kumar

M.G.M Higher Secondary School · Delhi Public School · VIBGYOR High

Addressing Critical Diagnostic Gaps

Hypertrophic cardiomyopathy affects 1 in 500 people globally, yet 80% of cases remain undiagnosed. PulseSage combines multi-modal deep learning with explainable AI to provide accurate, interpretable cardiac diagnostics.

Multi-Modal Integration

Late fusion strategy combining ECG and cMRI data with recall-weighted optimization

Explainable AI

LIME integration provides transparent, interpretable predictions for clinical trust

Clinical-Grade Accuracy

98.53% accuracy with 98.70% sensitivity in cMRI, 90.19% accuracy in ECG

Try with Sample Data

No medical data? Use our pre-loaded sample from anonymized research datasets. Multimodal analysis provides the most comprehensive diagnosis.

Try the Demo

Upload ECG signals or cardiac MRI scans to see the framework in action. The system will analyze your data using both classification models and provide interpretable results.

ECG Data (.dat files)

Drop ECG file here or click to upload

Supports .dat, .csv, .txt files

cMRI Scan

Drop cMRI file here or click to upload

Supports .jpg, .png, .dicom files

How it works

Upload ECG data (12-lead recordings) and/or cardiac MRI scans

Models process data through CNN architectures with optimized preprocessing

Receive classification results with LIME-based explainability visualizations

98.53%

cMRI Model Accuracy

98.70% Sensitivity

90.19%

ECG Model Accuracy

89.35% Recall

80%

Undiagnosed Cases

Current healthcare gap

Framework Architecture

Two specialized deep learning models working in tandem through an optimized late fusion strategy

ECG Classification Model

Custom CNN Architecture

Three-block convolutional neural network trained on 21,000+ 12-lead ECG recordings from the PTB-XL dataset. Features custom loss function with 2.5× penalty on false negatives.

—90.19% accuracy on validation set

—89.35% recall for HCM detection

—Ensemble with XGBoost for robustness

—Data augmentation and class balancing

ECG Signal Processing Pipeline

cMRI Analysis with LIME Explainability

cMRI Classification Model

Deep CNN with Explainability

Six-layer convolutional architecture trained on 59,267 cardiac MRI scans from Omid Hospital dataset. Integrates HSV-based color filtering and LIME for interpretable feature highlighting.

—98.53% accuracy, 98.70% sensitivity

—LIME highlights diagnostic regions

—HSV color filtering preprocessing

—Optimized for clinical interpretability

Late Fusion Integration

The framework employs a recall-weighted late fusion strategy, combining predictions from both models using weights derived from their individual recall scores (89.35% for ECG, 98.70% for cMRI). This approach maximizes diagnostic sensitivity while maintaining high specificity.

Research Contributions

Clinical Impact

1 in 500

people worldwide affected by HCM

80%

of HCM cases currently undiagnosed

annual risk of sudden cardiac death

Technical Innovations

—

Late fusion with recall-based weighting

—

Custom loss function penalizing false negatives 2.5×

—

LIME integration for clinical trust

—

Multi-modal data preprocessing pipeline

Read Full Research Paper

Research Team

Sampurn Kumar

M.G.M Higher Secondary School, Bokaro

Divya Raj

M.G.M Higher Secondary School, Bokaro

Sarvagya Singh

Delhi Public School, Bokaro

Aditya Kumar

VIBGYOR High, Mumbai