Kokoro - Intelligent Cardiac Health Analysis

Predicting the Future of Every Pulse.

Comment

Inspiration

The Global Cardiac Health Crisis

Every year, cardiovascular diseases (CVDs) claim approximately 17.9 million lives – making them the #1 leading cause of death worldwide, accounting for 31% of all global deaths. But the statistics get even more alarming:

  • 1.12+ Billion People currently suffer from cardiovascular diseases globally.

  • One person dies every 2-3 seconds from a heart-related condition somewhere in the world.

  • 20.5Million annual deaths from CVDs, with 85% occurring in low and middle-income countries.

  • 80% of heart attacks and strokes are preventable through early detection and lifestyle changes.

  • Only 36% of cardiac patients get diagnosed in the early, treatable stages.

  • Healthcare costs: CVDs cost the global economy $1 trillion annually in healthcare services and lost productivity.

  • Economic barriers: A single cardiac diagnostic scan costs $300-$5,000, out of reach for billions in poverty.

In a world where 1.12 billion people suffer from cardiovascular disease, and 80% of these deaths are preventable, the greatest tragedy isn't heart disease itself – it's the barriers that prevent people from knowing they have it. Kokoro exists to tear down those barriers.

What it does

Kokoro is a dual-model AI system that provides comprehensive cardiac health analysis through two complementary approaches:

Model 1: Heart Disease Risk Assessment

  • Input: 15 patient health parameters (Age, Blood Pressure, Cholesterol, Heart Rate, etc.)
  • Output:Cardiac risk percentage with binary classification (Disease/No Disease)
  • Use Case: Quick preliminary screening in clinical settings

Model 2: ECG Condition Classification

-Input: 12-lead ECG image (standard medical format)

-Output: Classification into 4 cardiac conditions:

  1. Normal (No) - Healthy ECG
  2. HB - Heart Block
  3. MI - Myocardial Infarction (Heart Attack)
  4. PM - Pacemaker
  • Technology: PCA + Deep Neural Network

How we built it

Tech Stack

 -Deep Learning: PyTorch (Neural Networks)
 -Data Processing: scikit-learn, pandas, numpy
 -Image Processing: scikit-image (ECG image analysis)
 -Frontend: Streamlit (Interactive web interface)
 -Deployment: Google Colab + ngrok (Public URL access)
 -Model Persistence: joblib (Model serialization)

Model Training

  • Dataset: 7,500 samples
  • Features: 15 (after one-hot encoding)
  • Train-Test Split: 80-20
  • Model: Binary Classifier (PyTorch)
  • Optimization: Adam optimizer with BCE Loss
  • Result: 80%+ accuracy on test set

ECG Classification Model

  • Dataset: 1000+ 12-lead ECG images
    • Image Processing:
      1. Grayscale conversion
      2. Gaussian filtering
      3. Otsu's thresholding
      4. Contour detection
      5. Feature extraction (255 features per lead × 12 leads)
    • Dimensionality Reduction: PCA (3060 → 400 components)
    • Result: Multi-class classification with confidence scores

Challenges we ran into

1. Environment & Installation Issues

  -WFDB library installation kept failing with cryptic errors.

 -Dependency conflicts with numpy/scipy versions    

 -Pip installation timeouts on slow connections

2. ECG Image Processing

-ECG images had non-standard coordinates across different datasets

-Lead extraction coordinates were inconsistent (300:600, 646:1135 hardcoded)

3. Feature Dimensionality

-Raw ECG features: 3060 dimensions (12 leads × 255 features each)

-Risk of overfitting with small dataset

Accomplishments that we're proud of

1.Dual-Model System

-Successfully integrated two completely different ML architectures:

  -Tabular data classification (Binary Classifier)

  -Image-based classification (PCA + Neural Network)

-Both models work seamlessly in one interface

2.Overcoming Technical Hurdles

-Solved WFDB installation issues by pivoting to alternative libraries.

-Reduced dimensionality from 3060 to 400 features without losing accuracy

-Achieved production-ready code despite multiple obstacles

3.Model Serialization & Loading

-Properly saved and loaded PyTorch models without compatibility issues.

-Implemented efficient caching for 10x faster load times

What we learned

1. Deep Learning Fundamentals

  • PyTorch architecture design: Building efficient neural networks

-Dropout & regularization: Preventing overfitting with limited data

  • Activation functions: Understanding ReLU, Sigmoid, Softmax

  • Loss functions: Binary Cross-Entropy vs Multi-class classification

2. Problem-Solving -When one approach fails, pivot quickly (WFDB → scikit-image) -Start simple, then optimize (basic model → fine-tuned version) -Test assumptions early (validate data before complex processing)

What's next for Kokoro - Intelligent Cardiac Health Analysis

1. Advanced Models

-LSTM/RNN:Analyze ECG as time-series data instead of images.

-Attention Mechanisms: Highlight important ECG regions.

-Ensemble Methods: Combine multiple models for better accuracy.

2. Clinical Integration

-Doctor Portal: Physician dashboard for reviewing results.

-HIPAA Compliance: Secure patient data handling.

-Risk Alerts: Automatic notifications for high-risk patients.

The goal: Transform cardiac healthcare from reactive treatment to proactive prevention through intelligent AI assistance.

Share this project:

Updates