Projects & Research Areas

The problem

People increasingly turn to chatbots for medical advice, including which antidepressant might work for them. Choosing an antidepressant depends heavily on the individual patient, and a confident but wrong answer can cause real harm. Before these systems can be trusted, we need rigorous ways to test whether their guidance is safe and grounded in evidence.

Our approach

• Simulated patients. LLM-based patient simulations that put conversational agents through realistic antidepressant-selection conversations and assess the risk in what they recommend.
• Causal reasoning. Causal-AI methods that connect patient characteristics to treatment outcomes, used to guide and check model recommendations.
• Beyond depression. The same causal-network approach applied to steering language models in COVID-19 care.

The problem

Chronic pain is shaped by biology, psychology, and life circumstances, yet clinical records capture little of how patients actually experience it. Motivational interviews let patients describe their pain in their own words, but the transcripts are long and unstructured, which makes them hard to analyze at scale.

Our approach

• Annotation schemes. Frameworks for labeling the biopsychosocial factors patients raise in motivational-interview transcripts, including content that is ambiguous by nature.
• Extraction models. Information-extraction systems that pull those factors out of transcripts automatically, so patterns can be studied across many patients.
• Student research. Undergraduate researchers contributing analyses of the interview data.

The problem

Health information is often written well above the reading level of the patients who need it most. And when patients ask language models health questions, the quality of the answer can shift with how the question is worded, which puts people with lower health literacy at a disadvantage.

Our approach

• Simplification corpora. An annotated corpus for simplifying digestive-cancer education materials, with new reinforcement-learning strategies for the simplification models.
• Robustness testing. Measuring how health question answering holds up when questions are rewritten at different reading levels.
• Efficient extraction. LLM-based extraction with knowledge distillation for analyzing HPV health-communication materials at low cost.

The problem

Probation, parole, and incident records hold information that matters for people's health and rehabilitation, but much of it lives in free-text notes or handwritten forms that are rarely analyzed systematically.

Our approach

• Supervision records. NLP analysis of supervisory condition notes in probation and parole records.
• Handwritten reports. An optical character recognition pipeline that turns handwritten assault and violence reports into text that extraction models can use.

What we're exploring

• Jailbreak robustness. Knowledge-aware, long-horizon attacks that test how well protected language models hold up against sustained adversarial pressure, so their safeguards can be strengthened.
• Retrieval-augmented generation. Query-aware diversity in retrieval, so models draw on a broader, less redundant set of sources when answering.