Bioinformatics and Medical Informatics

Our research is focused on advancing machine learning, statistical modeling, natural language processing and data analytics to tackle key challenges in biomedical fields. We work on developing novel deep learning methods for analyzing single-cell genomic data, improving medical informatics and creating medical ontologies. Recent projects include modeling Social and Commercial Determinants of Health, which significantly influence public health and building ontologies to better understand their impacts. Our work also applies deep learning to medical AI challenges, such as tumor identification in brain MRI images and vessel identification in vascular ultrasound and enhances electronic health records (EHR) usability through machine learning. This includes improving comprehension of clinical notes with large language models and other methods to simplify EHR text for patients. In addition, we are leveraging generative models to predict Alzheimer’s disease progression by addressing missing data issues and under-representation in marginalized populations. Our aim is to integrate complex, multimodal data for more accurate prediction models in aging-related diseases. Overall, our efforts focus on integrating biological information and computational techniques to improve both the analytical performance and interpretability of complex biomedical data.

Research Report 2025

 

James Geller
Professor
geller@njit.edu

Research Areas:Medical Informatics, Medical Ontologies/Terminologies

Social and Commercial Determinants of Health

The pivotal impact of Social Determinants of Health (SDoH) on people’s health and well-being has been widely recognized and researched. These include upbringing, living and working environment and access to transportation and healthcare. In simple words, it is easier to stay healthy if you have financial means. However, the effect of Commercial Determinants of Health (CDoH) is only now garnering increased attention. The study of CDoH can offer a systematic approach to identifying and categorizing the diverse commercial factors affecting health, e.g., advertising of tobacco products. These factors, including the production, distribution and marketing of goods and services, may exert a substantial influence on health outcomes. An ontology is a wellorganized collection of the major terms in a subject area. We have developed an ontology for CDoH by utilizing published medical articles and ChatGPT and an ontology for SDoH and integrated them.

 

Yehoshua Perl
Professor
perl@njit.edu

Research Areas: Medical informatics, Semantic web, ontologies, interface terminologies, quality assurance of biomedical terminologies, electronic health records

Highlighting of Electronic Health Records (EHR)

We use Machine Learning to design an interface terminology to highlight the important content of EHR notes for fast skimming.

Simplification of EHR notes

It will enable patients to comprehend their clinical notes in Portals. EHR texts are difficult. We use highlighting, LLM summaries and ML techniques.

 

Usman Roshan
Associate Professor
usman.w.roshan@njit.edu

Research Areas: Machine learning, medical informatics

Deep Learning for Medical AI Problems

We work on several medical AI problems involving images and DNA. In collaboration with neuroscientists, we are creating new models for tumor identification in brain MRI images and are studying them across different data sources. With vascular surgeons we are proposing novel models for identifying vessel and plaque in vascular ultrasound images from real patients. We are studying simple random networks for the classification of histopathology slide images and find them to be highly accurate there.

 

Zhi Wei
Distinguished Professor
zhiwei@njit.edu

Research Areas: Deep learning, statistical modeling, bioinformatics

Deep Learning Methods to Integrate Biological Information for Analysis of Single-cell RNAseq Data

The broad long-term objective of the project concerns the development of novel machine learning methods and computational tools for modeling genomic data, motivated by important biological questions and experiments. The analysis of single-cell RNA sequencing (scRNAseq) data presents substantial computational and bioinformatics challenges. The specific aim of the project is to develop novel modelbased deep learning methods, with prior biological information considered, for modeling scRNAseq data. These problems are driven by collaborations with biomedical investigators and the proposed approaches are designed to integrate biological information to improve both analytical performance and biological interpretability. Central to the project is the novel integration of biological insights with advanced deep learning methods for analyzing the noisy, sparse and over-dispersed scRNAseq data. These methods include the zero-inflated negative binomial model, autoencoder, deep embedding, hyperbolic embedding and reversed graph embedding. Additionally, the project will explore leveraging large language models (LLMs) specifically for single-cell omics data. LLMs will be applied to improve the modeling of complex gene expression patterns, capture biological heterogeneity and assist in interpreting the highdimensional data characteristic of single-cell omics. By combining LLMs with deep learning techniques, we aim to create more robust computational models that enhance biological discovery and provide deeper insights into cellular processes.

 

Chenxi Yuan
Assistant Professor
cy324@njit.edu

Research Areas: Deep learning, generative AI, bioinformatics

Enhancing Alzheimer’s Disease Progression with Deep Learning Models

This project aspires to refine the predictive modeling of Alzheimer’s disease (AD) progression by leveraging multimodal longitudinal data, encompassing biomarkers, neuropsychiatric results and neuroimaging. A significant obstacle in such studies is the lack of comprehensive data due to small sample sizes, limited data elements and the issue of missing data. This is particularly concerning for marginalized populations, either affected by AD or from ethnic minority groups. Existing imputation techniques inadequately address the intricate patterns in high-dimensional data. This project aims to counter these challenges by developing generative AI models to impute missingness and augment data for under represented groups with simulated data to enhance fairness. Long-term, this initiative seeks to develop an end-to-end prediction system that will accelerate prediction modeling for aging-related problems addressing the challenges of missing data, under representation and integration of complex, multimodal data.