Clinical AI

AI for Healthcare and Clinical Decision Support

The program moves from predictive models toward trustworthy clinical AI: models that explain themselves, respect physiology, improve equity, and connect to real clinical decisions.

01EHR and clinical context

02Patient-specific representation

03Prediction with uncertainty

04Clinical decision support

What this program is building

Selected recent and foundational papers, summarized around the task, why it matters, and the main technical result.

2026arXiv

PyHealth 2.0: A Comprehensive Open-Source Toolkit for Accessible and Reproducible Clinical Deep Learning

1Clinical datasets

2Reusable tasks and models

3Reproducible AI studies

Task: Build a reusable software foundation for clinical deep learning across datasets, tasks, and modalities.
Why it matters: Students and collaborators can move faster because benchmark setup, model comparison, uncertainty, and interpretability live in one reproducible toolkit.
Main result: PyHealth 2.0 turns many clinical AI experiments into repeatable pipelines rather than one-off codebases.

Paper details

2025npj Digital Medicine

Improving Medical Machine Learning Models with Generative Balancing for Equity and Excellence

1Imbalanced cohorts

2Generative balancing

3Equitable prediction

Task: Use generative data balancing to train medical models when patient subgroups are underrepresented.
Why it matters: The work targets a practical failure mode in clinical AI: strong average performance that hides poor performance for smaller populations.
Main result: Generative balancing improves the equity-performance tradeoff by strengthening subgroup representation during training.

Paper details

2025ICLR

Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval

1Patient history

2KG community retrieval

3Reasoned prediction

Task: Retrieve clinically relevant communities from biomedical knowledge graphs to support patient-level prediction.
Why it matters: It makes prediction models less isolated from medical knowledge and easier to inspect for clinical reasoning.
Main result: Knowledge graph community retrieval gives models structured context that improves reasoning around patient risk.

Paper details

2025ICLR

Small Models are LLM Knowledge Triggers for Medical Tabular Prediction

1Medical table

2Compact predictor + LLM knowledge

3Clinical risk score

Task: Combine compact tabular models with LLM knowledge for structured medical prediction.
Why it matters: The approach keeps prediction efficient while drawing on broader medical knowledge when it matters.
Main result: Small predictive models can trigger useful LLM-derived knowledge without turning the whole workflow into a large-model system.

Paper details

2024ICLR

GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs

1EHR events

2Personalized graph

3Prediction + explanation

Task: Create patient-specific knowledge graphs for clinical prediction from EHR context and biomedical knowledge.
Why it matters: Personalized graphs help connect model outputs to patient-specific evidence instead of generic feature importance.
Main result: GraphCare improves prediction by tailoring graph structure to each patient and exposing relevant clinical concepts.

Paper details

2024npj Digital Medicine

DRG-LLaMA: Tuning LLaMA Model to Predict Diagnosis-Related Group for Hospitalized Patients

1Hospital record

2Tuned medical LLM

3DRG prediction

Task: Adapt an LLM to predict hospital diagnosis-related groups from clinical records.
Why it matters: DRG prediction is tied to hospital operations, coding, and reimbursement, making it a useful testbed for deployable medical AI.
Main result: Domain tuning turns general language modeling into a targeted hospital workflow model.

Paper details

Representative publication links

ICLR · 2024

GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs

Personalized graph structure for clinical prediction.

Paper details npj Digital Medicine · 2025

Improving Medical Machine Learning Models with Generative Balancing for Equity and Excellence

Balances performance with equity across patient groups.

Paper details arXiv · 2026

PyHealth 2.0: A Comprehensive Open-Source Toolkit for Accessible and Reproducible Clinical Deep Learning

Unifies clinical datasets, tasks, models, interpretability, and uncertainty tools.

Paper details

Interested in this program?

Send a concise note with the program name, your role, the problem you want to work on, and any relevant data, code, clinical setting, or research experience.

Contact Sunlab

Medical LLMs Drug Discovery Synthetic Data