Biological Age Prediction from Medical Imaging

The Challenge

Chronological age doesn’t reflect an individual’s true physiological state. Biological age—how “old” a person’s body functions—can differ significantly and is a better predictor of health outcomes. Building accurate predictive models requires integrating multiple data modalities: medical imaging, clinical variables, and patient demographics.

The Solution

I developed a multi-modal healthcare ML system that combines imaging and clinical data:

1. Multi-Modal Input Integration:

   • CT Imaging Features: Extracted from Opportunistic Cardiometabolic Screening scans
   • Clinical Variables: Patient demographics, lab values, and health metrics
   • Fusion Architecture: Neural networks that learn joint representations from both modalities

2. Multiple Prediction Tasks:

   • Biological Age Regression: Predicting physiological age from multi-modal inputs
   • Mortality Risk Classification: Forecasting death outcomes for risk stratification
   • Additional Medical Outcomes: Extensible framework for other healthcare predictions

3. Production ML Practices:

   • Proper data preprocessing and normalization for medical data
   • Train/validation/test splits with appropriate stratification
   • Model checkpointing and evaluation metrics

The Impact

This project demonstrates the full ML lifecycle for healthcare applications: handling sensitive medical data, integrating multiple modalities, and building models that can inform clinical decision-making. The multi-modal approach significantly outperforms single-modality baselines, showcasing the importance of combining imaging and clinical data for accurate health predictions—essential skills for ML engineers working in healthcare technology.