-
Cardio Lens
🫀 Cardio-Lens: Democratizing Heart Health — Project Story
💡 Inspiration
Cardiovascular disease (CVD) kills 17.9 million people every year — more than cancer, diabetes, and respiratory diseases combined. Yet early detection remains locked behind expensive clinical tests that most of the world can't access.
We asked ourselves one question:
What if a smartwatch could be the first line of defence?
The idea was simple but ambitious — build a Two-Tier AI System that mirrors how healthcare actually works:
- A mass screening model using data that wearable devices already collect (blood pressure, weight, activity level)
- A clinical model that kicks in only when the screening flags a risk, using data from actual medical tests (ECG readings, chest pain type, exercise angina)
This tiered approach means that billions of people with a wearable device could get an initial risk assessment — and only those flagged would need the clinical follow-up. That's the democratization we were after.
🧠 What We Learned
The Math Behind the Models
Both tiers use a Random Forest Classifier, an ensemble of decision trees that votes on the final prediction. The predicted probability for class $c$ is:
$$ P(c \mid \mathbf{x}) = \frac{1}{T} \sum_{t=1}^{T} P_t(c \mid \mathbf{x}) $$
where $T$ is the number of trees (150 for Tier 1, 200 for Tier 2) and $P_t$ is the probability estimate from tree $t$.
Feature Engineering Matters More Than Fancy Algorithms
One of our biggest "aha" moments was realizing that BMI — a derived feature — significantly boosted Tier 1 accuracy. We computed it as:
$$ \text{BMI} = \frac{\text{weight (kg)}}{\left(\frac{\text{height (cm)}}{100}\right)^2} $$
The raw dataset stored age in days, so we also had to convert:
$$ \text{age_years} = \left\lfloor \frac{\text{age_days}}{365.25} \right\rfloor $$
These small transformations taught us that domain knowledge + simple math often outperforms blindly throwing data at a neural network.
Data Cleaning is Half the Battle
The Tier 1 dataset (cardio_base.csv) contained 70,000 records, but many had physiologically impossible blood pressure values (e.g., systolic BP of 16000!). We applied clinically informed filters:
$$ 90 \leq \text{ap_hi} \leq 200 \quad \text{and} \quad 50 \leq \text{ap_lo} \leq 140 $$
After cleaning, we retained ~68,551 records — a 2% loss that dramatically improved model reliability.
Explainability is Non-Negotiable in Healthcare AI
We learned that in medical AI, showing why a prediction was made is just as important as the prediction itself. Our Tier 2 model exposes its feature importance scores, letting users (and doctors) see exactly which clinical factors drove the result — making it an Explainable AI (XAI) system, not a black box.
🔨 How We Built It
Architecture
┌──────────────────────────────────────────────────────────┐
│ Streamlit Frontend │
│ ┌────────────┐ ┌──────────┐ ┌──────────┐ ┌───────────┐ │
│ │ The Pitch │ │ Tier 1 │ │ Tier 2 │ │Health Twin│ │
│ └────────────┘ └──────────┘ └──────────┘ └───────────┘ │
├──────────────────────────────────────────────────────────┤
│ backend.py │
│ ┌─────────────────────┐ ┌─────────────────────┐ │
│ │ load_and_preprocess │ │ predict_tier1() │ │
│ │ _tier1() / _tier2 │ │ predict_tier2() │ │
│ │ train_tier1_model │ │ simulate_bp_reduc() │ │
│ │ train_tier2_model │ │ │ │
│ └─────────────────────┘ └─────────────────────┘ │
├──────────────────────────────────────────────────────────┤
│ dataset/cardio_base.csv dataset/heart_processed.csv│
│ (70k records) (918 records) │
└──────────────────────────────────────────────────────────┘
Tech Stack
| Layer | Technology | Why We Chose It |
|---|---|---|
| Frontend | Streamlit | Rapid prototyping with gorgeous custom CSS — no React/JS needed |
| ML Models | scikit-learn (Random Forest) | Interpretable, fast to train, works great on tabular data |
| Data | Pandas + NumPy | Industry-standard for data wrangling and numerical ops |
| Visualizations | Altair | Declarative, interactive charts that respond to user input |
The Two-Tier Pipeline
Tier 1 — The "Watch" Model (Mass Screening)
- Trained on 70k+ records from a population cardiovascular dataset
- Uses 12 features: age, gender, height, weight, BMI, systolic/diastolic BP, cholesterol, glucose, smoking, alcohol, physical activity
- Hyperparameters:
n_estimators=150,max_depth=12,min_samples_leaf=10 - Includes a BP Simulator — an interactive slider that shows how reducing your systolic BP changes your risk in real-time, powered by running the model across a range:
$$ \text{risk}(bp) = P(\text{CVD} = 1 \mid \text{ap_hi} = bp, \ldots) \quad \forall \ bp \in [\text{target}, \text{current}] $$
Tier 2 — The Clinical Model (Doctor's Diagnosis)
- Trained on 918 clinical records with 15 one-hot encoded features
- Includes ECG readings, chest pain type, exercise-induced angina, ST slope
- Hyperparameters:
n_estimators=200,max_depth=10,min_samples_leaf=5 - Outputs a feature importance chart for full explainability
🧬 The Health Twin Simulator — Our Unique Feature
This is the feature we're most proud of. No other heart disease app does this.
The simulator lets a user create a "Future Healthy Self" by setting health goals (lower BP, lose weight, quit smoking, exercise more). The AI then runs 22 parallel predictions — 11 years × 2 scenarios (current trajectory vs. healthy trajectory) — to build a side-by-side 10-Year Risk Trajectory.
The risk at each future year $y$ is modeled as:
$$ R_{\text{current}}(y) = P!\left(\text{CVD} \mid \text{age} + y,\; \text{current features}\right) $$
$$ R_{\text{healthy}}(y) = P!\left(\text{CVD} \mid \text{age} + y,\; \text{goal features}\right) $$
The "Years of Aging Reversed" metric converts the risk reduction into something intuitive — telling a user, for example, "Your heart will be 7 years younger after these changes."
🧗 Challenges We Faced
1. Noisy, Real-World Data
The Tier 1 dataset had extreme outliers — blood pressure values of 16,000 mmHg, negative ages, and heights of 250+ cm. We spent significant time designing robust filters that removed noise without accidentally discarding edge-case patients.
2. Class Imbalance
Both datasets had near-balanced classes (~50/50), which is unusual in medical datasets. While this simplified training, it made us question data collection methodology and iterate carefully with stratify=y in our train-test splits to preserve the ratio.
3. The "Two Dataset" Problem
Our two tiers used completely different datasets with zero overlapping features. This meant we couldn't build a single unified model — we had to design a system where the two models work sequentially but independently: the screening model acts as a gate, and the clinical model provides the definitive diagnosis.
4. Making AI Output Feel Human
Raw probability scores like 0.7341 mean nothing to a non-technical user. We spent a lot of time on UX — converting probabilities to percentage-based risk scores, color-coded cards (🟢 Low / 🟡 Moderate / 🔴 High), and the Health Twin's "Years of Aging Reversed" metaphor.
5. Performance at Scale
The BP Simulator runs the model across potentially 100+ blood pressure values in real-time as the user moves a slider. We used @st.cache_resource to avoid retraining models on every page load and vectorized predictions where possible to keep the experience snappy.
📊 Results
| Metric | Tier 1 (Screening) | Tier 2 (Clinical) |
|---|---|---|
| Dataset Size | ~68,551 records | 918 records |
| Features | 12 (wearable-level) | 15 (clinical-level) |
| Accuracy | ~72% | ~87% |
| Use Case | Mass population screening | Hospital-grade diagnosis |
While 72% accuracy for Tier 1 may seem modest, it's important to note that this model uses only wearable-level data — no blood tests, no ECGs, no clinical visits. For a tool that could run on a smartwatch and screen millions of people, a 72% early warning signal is a powerful first step.
🚀 What's Next
- Deep Learning Upgrade: Replace Random Forest with a gradient-boosted ensemble (XGBoost/LightGBM) for potential accuracy gains
- Real Wearable Integration: Connect with Apple HealthKit / Google Fit APIs for live data
- Federated Learning: Train across hospital datasets without sharing patient data
- Longitudinal Tracking: Let users log predictions over time and track their health journey
"The best time to check your heart health was 10 years ago. The second best time is now."
