Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous syndrome, and social, environmental, and behavioral exposures, collectively the exposome, shape how it presents and progresses. This project links where patients live to what happens biologically and to what can be changed.

• Developing interpretable multimodal data fusion models that integrate EHR, cardiac imaging (echocardiography, CMR), multi-omics, and exposome/geospatial data to link social and environmental exposures to HFpEF subphenotypes and outcomes.
• Analyzing a multimodal HFpEF cohort (NHLBI HeartShare: imaging, ECG, CPET, omics) on the BioData Catalyst cloud platform to identify biological pathways (inflammation, fibrosis, metabolic dysfunction) and modifiable targets for precision prevention.

• Multimodal Fusion
• Interpretable ML
• EHR
• Cardiac Imaging
• Multi-omics
• Social Determinants of Health
• BioData Catalyst

• Designed an intermediate-fusion pipeline with modality-specific MLP encoders that integrates clinical, intracardiac electrogram (EGM), and metabolomics data, and handles missing modalities without imputation.
• Benchmarked classical ML (Random Forest, XGBoost) and deep models (TabPFN, Transformers) for AF-type classification and post-ablation recurrence prediction.

• 0.87 AUPRC · AF-type classification

• Multimodal Fusion
• Missing Modalities
• TabPFN
• Transformers
• PyTorch

• Built a locally deployable pipeline that turns open-source VLMs into mammography report generators and multi-task classifiers, with an optional multimodal RAG mode over image–text exemplars.
• Benchmarked MedGemma, LLaVA-Med, and Qwen2.5-VL on VinDr-Mammo and DMID under zero-shot, few-shot, chain-of-thought, and RAG prompting, plus QLoRA fine-tuning.
• Few-shot prompting and RAG improved report quality (BERTScore, ROUGE-L); QLoRA-tuned MedGemma gave the best classification results.

• 0.88 breast density acc.
• 0.93 calcification acc.
• 0.75 BI-RADS acc.

• VLMs
• Multimodal RAG
• QLoRA
• Ollama

• Built an LLM agent-based RAG recommender (Llama, Mistral, DeepSeek; LangChain, ChromaDB) that matches researchers' clinical and technical needs to 22 public mammography datasets.
• Designed a multichannel interface that pairs conversational search with an interactive network visualization (PyVis) of dataset relationships.

• LLM Agents
• RAG
• LangChain
• ChromaDB