Medical Imaging AI

• CARDINAL, Clinically Grounded CT Representation Learning, developed a novel representation learning framework that extracts clinically grounded latent features directly from non-contrast CT scans, enabling prediction of 10-year major adverse cardiac events beyond traditional imaging biomarkers and risk scores.
• Multimodal Cardiovascular Risk Prediction, led development of a multimodal AI framework integrating imaging-derived representations and structured clinical data for long-horizon cardiac risk prediction, achieving AUROC up to 0.88 on rare outcomes and improving performance by ~30% over clinical and imaging baselines.
• Probability-Level Multimodal Fusion, proposed a probability-level multimodal fusion method for rare downstream outcome prediction, improving AUROC by 32.4%, AUPRC by 60.0%, and F1 by 46.3% while outperforming standard fusion techniques on imbalanced datasets.
• Large-Scale Tabular Clinical Modeling, built a weighted-loss XGBoost pipeline with sequential feature integration, achieving AUC 0.88 for 10-year cardiac event prediction across 27,000+ patients, exceeding established clinical risk scores by 30%.
• Active Learning for Label Efficiency, first author on a peer-reviewed framework reducing labeling requirements by over 75%, matching full-model performance using only 15.4% of data in binary classification and 23.1% in multi-class classification.
• Cross-Modal Cardiovascular Screening, developed models that predict coronary artery calcium scores from chest X-rays, achieving up to 80% AUROC and enabling lower-cost cardiovascular screening.
• Women's Health AI, conducting research on multimodal models for heavy menstrual bleeding using imaging and structured clinical data, achieving sub-1 mL RMSE with internally developed systems.
• Human-in-the-Loop Eye-Tracking AI, co-authored a system leveraging radiologists' eye-tracking data to guide model attention, improving interpretability and alignment with clinical decision-making.
• Fairness in Medical AI, investigating methods to mitigate racial and gender bias in medical AI models, with a focus on fairness and generalization across diverse patient populations.
• Clinical Collaboration, working with leading Emory clinicians including Drs. Carlo De Cecco, Marly van Assen, Fatemeh Ardeshir-Larijani, Arshed Quyyumi, and Elizabeth Krupinski to ensure clinical validity, robustness, and translational relevance.

Publications and Working Papers

• Deep Active Learning for Lung Disease Severity Classification from Chest X-rays: Learning with Less Data in the Presence of Class Imbalance
  Journal of Imaging Informatics in Medicine, 2025.
  Demonstrates a deep active learning framework reducing labeling requirements by over 75% while maintaining full-model performance under class imbalance.
• Observer Performance and Eye-Tracking Variations as a Function of AI Output Format
  SPIE Medical Imaging 2025: Image Perception, Observer Performance, and Technology Assessment.
  Evaluates how different AI output formats influence radiologist decision-making and diagnostic behavior.
• Explainable Machine Learning for Risk Stratification of Major Adverse Cardiac Events Using Clinical and Imaging Data
  Circulation, 2025 Abstract.
  Demonstrates interpretable multimodal machine learning models outperforming traditional cardiovascular risk scores.
• Predicting 10-year Major Adverse Cardiac Events Using Multi-Source Modalities with XGBoost: Establishing a Baseline for Multimodal Fusion in Cardiac Risk Assessment
  medRxiv, 2025. Under review at Journal of Cardiovascular Computed Tomography.
  Establishes a large-scale multimodal baseline achieving AUC 0.883 on 27,000+ patients and improving performance by ~30% over clinical risk models.
• CARDINAL: Clinically Grounded Representation Learning from CT for Long-Term Cardiovascular Risk Prediction
  In preparation, target submission 2026.
  Introduces a representation learning framework that learns prognostic features directly from CT imaging for long-horizon cardiovascular risk prediction.
• Multimodal Fusion for Cardiovascular Risk Prediction Using Imaging and Clinical Data
  In preparation, target submission 2026.
  Develops a principled multimodal fusion framework for rare outcome prediction with improved robustness and interpretability.
• GazeRIB-CXR: Gaze-guided Radiologist-Informed Blackout AI for Chest X-ray Analysis
  In preparation, target submission 2026.
  Introduces a radiologist-informed training paradigm using multi-reader eye-tracking data, with up to 14% improvement across classification metrics.

View my Google Scholar profile

Integrated AI-Based Discovery & Innovation

• Ensemble Portfolio Optimization, pioneered an ensemble portfolio optimization algorithm incorporating LSTM neural networks and Deep Q-Learning, surpassing equal-weighted portfolios by 200%, SPX by 100%, and QQQ by 50%.
• Reinforcement Learning for Rebalancing, replicated and advanced research in institutional portfolio rebalancing, achieving fast convergence using Linux, PyTorch, and GPU acceleration at scales up to 512 assets.
• High-Dimensional Sampling and Volatility Analysis, conducted research on 100-dimensional sampling with the Ellipsoid method in under 20 minutes while analyzing stock-universe volatility across look-back periods.
• Implied Volatility Curve Modeling, created robust implied volatility curves for derivatives and developed profitable trading strategies.
• Generative AI for Trading Decisions, developed and optimized LLM-based systems for financial reasoning and trading decision support.