Hello! I’m Abhijeet Sahdev, a graduate student in Artificial Intelligence at New Jersey Institute of Technology. I study how latent geometry and transport constraints determine whether generative models preserve clinical hierarchy in real-world EHR data. I work on geometric and generative machine learning for structured clinical data, with a focus on hyperbolic diffusion and flow-based models for EHR- and ontology-grounded risk prediction. My research examines how latent geometry and transport dynamics affect hierarchy preservation, interpretability, and robustness in clinical models. I design methods end-to-end and validate them at MIMIC scale, aiming to build research-grade systems that are mathematically principled, clinically meaningful, and reliable at scale.
Education
Research Experience
Currently working on HyperMedDiff-Risk under Prof. Mengjia Xu.
- Designed a geometry-aware generative framework to study how latent curvature and transport govern hierarchy preservation in clinical risk prediction.
- Developed a synthetic ICD trajectory benchmark using rooted-tree ontologies to isolate geometry–accuracy trade-offs between Euclidean and hyperbolic diffusion and rectified-flow models.
- Introduced HyperMedDiff-Risk, integrating hyperbolic ICD embeddings, graph diffusion encoders, and rectified-flow transport to prevent hierarchical collapse during generation.
- Achieved state-of-the-art performance on MIMIC-III heart-failure prediction (AUPRC 0.79–0.83 vs. 0.706 MedDiffusion; Cohen’s κ ≈ 0.50), with ablations quantifying curvature, diffusion depth, and decoder expressivity.
- Trained hyperbolic ICD embeddings with HDD-supervised alignment (Corr = 0.84–0.91), substantially improving hierarchical fidelity over Euclidean baselines; manuscript in preparation.
Working on a web-based implementation of 3D Copilot under Prof. Lei Zhang.
- Building a browser-native 3D copilot for controllable scene generation with React and react-three-fiber, backed by an LLM-driven staging and commit workflow.
- Developed GLB optimization and streaming (gltf-transform, Draco, texture compression), cutting 3D asset load times by 13% for interactive editing.
- Prototyping GNN-based scene representations and structured tool-calling specs (1.6K+ lines) to encode spatial constraints, collision checks, and safe animation rules.
Publications & Manuscripts
Paper Presentations & Reviews
- Presentation on DEPLOYR, a framework for deploying custom real-time ML models into EMR, with focus on big-data grounding and deployment.
Selected Systems & Research Projects
- Implemented a minimal GPT-2 in C++ with custom autograd, CUDA kernels, tokenizer/dataloader, and Adam optimizer to study transformer internals.
- Extended to distributed training with MPI data parallelism, gradient allreduce, and rank-aware batching across multi-GPU nodes (67× train, 14× inference speedups vs CPU baseline).
- Engineered and preprocessed MIMIC-CXR-JPG v2.1.0 (377K+ chest X-rays, 227K reports) via GCP BigQuery and CheXpert, generating structured mini-reports and building a balanced 66K+ study dataset with robust stratified splits.
- Implemented and benchmarked GIT-base vision–language architecture under Full Fine-Tuning vs. LoRA (vision-only, text-only), designing a resource-optimized training strategy that reduced GPU memory usage by 50% and training time by 67%.
- Demonstrated 5.8× improvement in BLEU with LoRA Text-Only over full fine-tuning, validating LoRA as a parameter-efficient and domain-adaptive strategy for medical vision–language tasks.
- Implemented WD3QNE for ICU sepsis treatment, replicating published results across 8 experiments with 43 vs. 45 feature sets on MIMIC-III.
- Trained agents over 100 epochs, producing structured 5×5 fluid/vasopressor policies aligned with clinician behavior.
- Designed and implemented a distributed ML pipeline for healthcare cost prediction on a multi-node AWS cluster using Hadoop MapReduce, training Ridge Regression with K-fold CV via distributed sufficient-statistics aggregation.
- Reduced network overhead with combiners, solved regularized normal equations in reducers, and computed per-fold R² in a dedicated evaluation job (mean R² ≈ 0.786 across 5 folds).