Selected work

Projects

A sample spanning clinical prediction, biomedical NLP, and core bioinformatics. Full code is on GitHub.

ICU Readmission Prediction from Clinical EHR Data (MIMIC-IV)

2025–26

A 30-day ICU readmission model built on MIMIC-IV (300K+ critical-care admissions from Beth Israel Deaconess). Complex PostgreSQL queries join demographics, ICD diagnoses, labs, vitals, and medications into 40+ engineered clinical features. Logistic regression, random forest, and XGBoost are compared with cross-validated tuning (AUC 0.78 on held-out test), class imbalance handled with SMOTE and threshold tuning, and SHAP used to surface the top predictive drivers for clinical interpretability.

PythonPostgreSQLXGBoostscikit-learnSHAPSMOTEMIMIC-IV

Spatial Transcriptomics Analysis of Human Brain Tissue

2025–26

An end-to-end 10x Visium analysis of human dorsolateral prefrontal cortex (spatialLIBD): QC, normalization, and dimensionality reduction with Scanpy and Squidpy; spatially variable genes via Moran's I and SpatialDE; unsupervised clustering to delineate cortical-layer expression domains validated against histology; and a cell2location deconvolution pipeline estimating cell-type proportions per spot from paired scRNA-seq reference.

ScanpySquidpycell2locationSpatialDE10x VisiumscRNA-seq

Bulk RNA-seq Differential Expression Pipeline

2025–26

A reproducible Nextflow pipeline for bulk RNA-seq: FastQC, Trimmomatic trimming, STAR alignment to GRCh38, and featureCounts quantification, followed by DESeq2 differential expression on a tumor-vs-normal GEO dataset (500+ significant DEGs at FDR < 0.05). GO and KEGG enrichment via clusterProfiler, publication-quality volcano/MA plots and heatmaps, all containerized with Docker for full reproducibility.

NextflowSTARDESeq2clusterProfilerDockerR

Multi-Class Text Classification with Traditional ML and Transformers

2025–26

An NLP pipeline for news-topic classification on AG News (120K articles), comparing TF-IDF + logistic regression, embeddings + LSTM, and a fine-tuned DistilBERT (94% accuracy vs. 91% baseline). Includes a full preprocessing stack, misclassification analysis with confusion matrices and per-class F1, and deployment as a FastAPI REST service with a simple web interface, containerized with Docker.

HuggingFaceDistilBERTPyTorchFastAPIDockerNLP