Cardio-X 2.0: Protecting Hearts with Intelligence

Tri-Level AI System for Early CVD Detection—fusing Exposome data, clinical vitals, and 12-lead ECGs using Explainable Gradient Boosting & CNN-LSTMs (0.93 AUC).

Comment

Inspiration

Cardiovascular disease kills ~17.9 million people per year, yet early detection is bottlenecked by expensive hospital systems. I noticed a disconnect: community health workers collect vitals and environmental data daily, but have no tool to triage who actually needs a cardiologist. I wanted to build a system that bridges this gap — from village screening all the way to specialist-grade ECG analysis — using nothing but existing public datasets and open-source ML.

What it does

Cardio-X 2.0 is a hierarchical triage system that funnels patients through three levels of increasingly specialized AI:

Level 1 — Community Screening:

  • Processes 27 non-invasive features (age, BMI, smoking, sleep hours, income ratio, waist circumference, etc.)
  • Integrates real-time air quality data via WAQI API (PM2.5 as an environmental CVD risk factor)
  • Ensemble of XGBoost + Random Forest + LightGBM, sensitivity-optimized at threshold 0.51
  • Result: 85.23% sensitivity — catches the vast majority of at-risk individuals

Level 2 — Clinical Assessment:

  • XGBoost classifier (200 estimators, dynamic scale_pos_weight = n_neg/n_pos × 1.2) on clinical lab data
  • Full SHAP (Shapley Additive Explanations) integration: waterfall plots show doctors exactly why a patient was flagged
  • Result: 0.932 AUC-ROC

Level 3 — ECG Specialist Confirmation:

  • 1D CNN-LSTM hybrid: 4× Conv1D blocks (32→64→128→256 filters) + 2× Bidirectional LSTM layers + Dense head
  • Trained on PTB-XL (21,837 real clinical 12-lead ECG records, 500Hz, Wagner et al. 2020)
  • 50 epochs, batch_size=16, with early stopping and class-weight balancing
  • Result: 0.925 AUC multi-label arrhythmia classification

How I built it

Data sources:

  • Hack4Health Dataset — cardio_base.csv (primary tabular data)
  • CDC NHANES 2017–2018 — 9,254 samples across 9 SAS tables (.XPT), parsed via pyreadstat, yielding 11 behavioral/metabolic features
  • WAQI API — Real-time PM2.5 air quality index, mapped by ZIP code
  • PTB-XL (PhysioNet) — 21,837 de-identified 12-lead ECG recordings

Stack:

  • Python, XGBoost, scikit-learn, TensorFlow/Keras, SHAP, wfdb, pandas, NumPy, Matplotlib
  • Streamlit (custom dark-themed dashboard with CSS overrides)
  • All models saved as .keras and .pkl for instant inference without retraining

(Note: I conceptualized the architecture individually. I used generative AI as a research assistant for hyperparameter tuning, gradient flow optimization, and deployment boilerplate.)

Challenges I ran into

  • Data alignment: Harmonizing real-time WAQI API data with static NHANES tabular records required custom temporal mapping logic to prevent data leakage
  • Class imbalance: Clinical datasets are inherently skewed. I had to dynamically compute scale_pos_weight and apply 1.2× sensitivity boosting to avoid majority-class collapse
  • Black box problem: Mid-project, I realized high accuracy means nothing if doctors can't trust it. I pivoted to fully integrate SHAP waterfalls into the Level 2 dashboard UI
  • Training time: The Level 3 CNN-LSTM took ~2 full days on a Mac M2 chip (50 epochs, 21,837 records × 12 leads × 5000 timesteps). I had to carefully manage memory with batch_size=16

Limitations

  • cardio_base.csv demographic composition is unknown — potential selection bias.
  • Level 1 specificity is 56.4%, meaning ~44% false-positive referral rate at the community level.
  • PTB-XL is predominantly European — generalization to South Asian populations is unvalidated.

Accomplishments I'm proud of

  • 0.925 AUC on multi-label arrhythmia classification using real PTB-XL clinical ECGs — not synthetic data
  • 85.23% sensitivity at Level 1, exceeding the 85% clinical target (95% CI: 0.8432–0.8609)
  • Successfully fused 4 completely different data modalities (tabular, environmental API, clinical labs, time-series ECG) into one unified pipeline
  • Built a fully interpretable system — every prediction comes with a visual SHAP explanation

What I learned

  • In medical AI, interpretability > accuracy. Doctors need to see why, not just what
  • Weighted loss functions are essential for clinical datasets — without them, the model collapses into majority-class prediction
  • Real-world data alignment (temporal, spatial, modal) is the hardest part of multi-modal ML — harder than the models themselves
  • Training deep networks on raw time-series (12-lead × 5000 timesteps) demands meticulous memory management

What's next for Cardio-X 2.0

Immediate:

  • Deploy a pilot study at a partner hospital for prospective clinical validation
  • Integrate a Conversational AI Health Assistant that takes the model's SHAP outputs and translates them into personalized dietary advice, maintenance routines, and triage recommendations for patients

Technical roadmap:

  • Replace the cascaded models with a single Multi-Modal Late-Fusion Transformer that simultaneously ingests tabular vitals, text surveys, and continuous ECG signals
  • Multi-center dataset expansion (target: 50,000+ records from 10+ institutions)
  • Federated learning for privacy-preserving cross-hospital collaboration (HIPAA/GDPR compliant)

Societal Impact:

  • By filtering 85% of truly at-risk patients at Level 1 using only a smartphone, the system reduces unnecessary specialist referrals. In India's public health system (1 cardiologist per 100,000 people), this hierarchical filtering could save an estimated 40% of specialist consultation time.

Long-term vision:

  • Longitudinal CVD progression modeling (1–5 year risk trajectories)
  • Expand framework to other cardiac conditions beyond arrhythmia.

Built With

Share this project:

Updates