Large Language Models for Trustworthy Healthcare and Clinical Decision Support

Project Description

Large Language Models (LLMs) are rapidly reshaping how clinical knowledge is accessed, explained, and acted upon. Yet healthcare is precisely the domain where an LLM’s familiar failure modes—hallucination, opacity, and unsafe recommendations—are least acceptable. The challenge is not simply to apply LLMs, but to make them accurate, explainable, and safe enough for real clinical and pharmacy workflows.

At InfoLab, Sungkyunkwan University (SKKU), we develop and evaluate LLM-powered systems that pair the fluency and reasoning of modern language models with knowledge grounding, explainability, and responsible-AI safeguards, spanning medication safety, patient-facing chatbots, and clinical outcome prediction.

Core Research Themes and Contributions

1. LLMs for Smart Pharmacy and Drug Safety

We study how LLMs can strengthen pharmacy systems—flagging drug–drug interactions, supporting safe dispensing, and improving operational efficiency—while keeping a human pharmacist firmly in the loop and minimizing unsafe automated advice.

2. Knowledge-Augmented, Explainable Clinical Chatbots

To curb hallucination, we build knowledge-augmented chatbots that ground their answers in trusted medical sources rather than free-form generation, producing explainable, evidence-linked responses that clinicians and patients can verify.

3. LLMs for Clinical Outcome Prediction

We investigate LLMs as predictive engines over heterogeneous clinical records—for example, mortality prediction for ICU patients with mental disorders—where language models can fuse unstructured notes with structured signals that traditional pipelines struggle to combine.

4. Responsible AI for Sensitive Domains

Across mental-health and other high-stakes settings, we examine the responsible-AI requirements—fairness, transparency, privacy, and accountability—that any deployed clinical LLM must satisfy.

Project Objectives

• Apply LLMs to pharmacy, clinical NLP, and outcome prediction without sacrificing safety.
• Ground and explain model outputs so recommendations are verifiable, not opaque.
• Mitigate hallucination and bias in patient- and clinician-facing systems.
• Establish responsible-AI guidelines for deploying language models in healthcare.

Research Impact

By treating trustworthiness as a first-class design goal, InfoLab (SKKU) advances LLM applications that clinicians can actually rely on—knowledge-grounded, explainable, and safety-aware tools that augment medication management, patient communication, and decision support rather than introducing new risks into care.