AI for Healthcare

We build trustworthy AI that learns from health records, biosignals, molecules and medical literature to support clinicians, patients and scientists.

Health systems collect vast records, brain signals and research papers, yet clinicians and scientists lack time and tools to use them. We turn this data into accurate, explainable, openly shared AI that fits real clinical and research workflows.

Clinical prediction & decision support

Models that forecast patient outcomes from health records while showing clinicians why, drawing on medical knowledge graphs.

RETAIN architecture: two attention paths weight visits and clinical variables. Figure from the paper.

NeurIPS · 2016

RETAIN: An Interpretable Predictive Model for Healthcare using Reverse Time Attention Mechanism

Choi et al.

The problem: Doctors need risk predictions they can trust, but the most accurate deep learning models were black boxes that could not explain their reasoning.
What we built: A two-level attention model that reads a patient's visits in reverse time order, highlighting which past visits and diagnoses drove each prediction.
Why it matters: Tested on 14 million visits from 263K patients, its accuracy was comparable to RNNs while its interpretability was comparable to traditional models.

Paper Code

GraphCare builds patient knowledge graphs for four clinical prediction tasks. Figure from the paper.

ICLR · 2024

GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs

Jiang et al.

The problem: Health records alone miss much of the medical knowledge doctors use, and building patient-specific knowledge into predictions has been hard.
What we built: GraphCare builds a personalized knowledge graph for each patient using large language models and biomedical knowledge bases, then reasons over it with a graph neural network.
Why it matters: On MIMIC-III it boosted AUROC by 17.6% for mortality and 6.6% for readmission prediction, with the largest gains when training data is limited.

Paper Code

Large language models for medicine

Specialized language models that read clinical notes and medical literature, working alongside experts rather than replacing them.

How LEADS training data was collected, linked and turned into tasks. Figure from the paper.

Nature Communications · 2025

A foundation model for human-AI collaboration in medical literature mining

Wang et al.

The problem: Systematic reviews underpin evidence-based medicine, but searching, screening and extracting data from thousands of trial papers is slow manual work.
What we built: LEADS, a language model trained on 633,759 examples from 21,335 systematic reviews, 453,625 trial publications and 27,015 trial registries.
Why it matters: In a 16-expert user study, LEADS raised screening recall to 0.81 from 0.78 while saving 20.8% of time, and cut data-extraction time 26.9%.

Paper Code

DRG-LLaMA predicts billing codes from discharge summaries, directly or in two steps. Figure from the paper.

npj Digital Medicine · 2024

DRG-LLaMA: tuning LLaMA model to predict diagnosis-related group for hospitalized patients

Wang et al.

The problem: Diagnosis-related groups determine how U.S. hospitals are paid for inpatient stays, yet assigning them from clinical notes is inefficient.
What we built: DRG-LLaMA fine-tunes the LLaMA language model on 236,192 MIMIC-IV discharge summaries to predict each hospital stay's DRG code directly from text.
Why it matters: It reached 52.0% top-1 accuracy and improved macro-F1 by 40.3% over ClinicalBERT and 35.7% over CAML.

Paper Code

AI for drug discovery

Shared datasets, benchmarks and models that help machine learning find drug candidates and predict how molecules act.

TDC organizes datasets, tasks, tools and leaderboards across therapeutic machine learning. Figure from the paper.

NeurIPS Datasets and Benchmarks · 2021

Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development

Huang et al. · co-led with the Zitnik Lab, Harvard

The problem: Machine learning for medicines lacked well-defined tasks and curated data, making methods hard to compare across the drug development pipeline.
What we built: An open platform and Python library unifying 66 AI-ready datasets across 22 learning tasks, with data splits, evaluators, molecule-generation oracles and leaderboards.
Why it matters: Experiments on TDC showed even the strongest algorithms struggle with distribution shift and novel data, giving the field shared, public benchmarks.

Paper Code

MolTrans workflow from drug and protein substructures to interaction prediction. Figure from the paper.

Bioinformatics · 2021

MolTrans: Molecular Interaction Transformer for drug-target interaction prediction

Huang et al.

The problem: Testing which drug compounds bind which protein targets experimentally is costly and slow given the enormous space of possible molecules.
What we built: MolTrans breaks drugs and proteins into substructures, learns from large unlabeled molecular databases, and models their pairwise interactions with a transformer.
Why it matters: It improved drug-target interaction prediction over state-of-the-art baselines on real-world data, while pointing to the substructures behind each prediction.

Paper Code

Biosignals: EEG, ECG and sleep

Models that read brain and heart signals and transfer across hospitals, devices and recording setups.

BIOT tokenizes biosignals and supports supervised and pre-training settings. Figure from the paper.

NeurIPS · 2023

BIOT: Biosignal Transformer for Cross-data Learning in the Wild

Yang et al.

The problem: Biosignal models are usually built for one dataset, because recordings differ in channels, lengths and missing data across hospitals and devices.
What we built: BIOT turns any biosignal into a uniform sequence of tokens, letting one transformer pre-train across EEG, ECG and activity datasets with mismatched formats.
Why it matters: On CHB-MIT seizure detection it improved balanced accuracy 3% over baselines, and pre-training on other datasets added up to 4% more.

Paper Code

Synthetic health data & privacy

Realistic synthetic patient records that let researchers build and share models without exposing real patients.

HALO generates records with visit-level and code-level modules. Figure from the paper.

Nature Communications · 2023

Synthesize high-dimensional longitudinal electronic health records via hierarchical autoregressive language model

Theodorou et al.

The problem: Privacy rules rightly restrict access to patient records, slowing research; synthetic records help only if they faithfully mirror real data.
What we built: HALO, a hierarchical language model that generates full patient histories visit by visit and code by code, without simplifying the data.
Why it matters: Its synthetic records matched real disease-code statistics with R2 above 0.9, and models trained on them reached similar accuracy to real-data models.

Paper Code

Open-source tools

Software from this line of work that others can use and build on.

PyHealth

Python toolkit for building and benchmarking healthcare deep learning models on EHR, signals and more; includes HALO.

Therapeutics Data Commons (TDC)

Open datasets, tasks, benchmarks and leaderboards for therapeutic machine learning (GitHub: mims-harvard/TDC; co-developed with Harvard).

DeepPurpose

Deep learning toolkit for drug-target interaction, drug property, and drug repurposing prediction.

BIOT

Biosignal transformer with pre-trained EEG models for cross-dataset learning.

LEADS model weights

Open LEADS model for medical literature search, screening and data extraction.

Want to work on this with us?

We work with students, clinicians, and industry partners. Email jimeng@illinois.edu.

Get in touch All publications