Better Accuracies, Worse Reasoning - A Step-Level Audit of Medical Chain-of-Thought Distillation
Accepted to the EACL 2027 main conference. Distilled models answer more accurately, but the steps behind their answers are less factual.
KnowLab builds machine learning, multimodal language models and knowledge graphs for health. We are based at the University of Glasgow and University College London.
Accepted to the EACL 2027 main conference. Distilled models answer more accurately, but the steps behind their answers are less factual.
One question runs through it all: how knowledge is carried in health data, made explicit, learned by models, and tested before anyone relies on it.
Predicting risk and finding gaps in care for cardiovascular disease, stroke and type 2 diabetes.
15 stories →Depression, anxiety, psychosis and the drugs used to treat them, studied from health records.
9 stories →Who is most at risk, how the disease unfolds, and what the pandemic cost, across millions of people.
11 stories →Finding patients, speeding up diagnosis and understanding who is affected by rare conditions.
10 stories →AI that reads, writes and checks radiology reports and scans, and keeps radiologists in the loop.
19 stories →Learning from tumour and pathology reports and slides, and judging AI-written findings.
4 stories →Spotting falls, frailty and delirium in older adults' records.
3 stories →Reading the heart-rate trace in labour and catching at-risk babies early.
2 stories →Understanding what clinical AI knows, where it fails, and how to tell the difference.
Read the stories →Reasoning, summarisation and coding over health records, evaluated step by step.
Read the stories →Linking images, text and records at population scale to support research and care.
Read the stories →How medical knowledge is represented, as explicit graphs and ontologies and as structure learned inside models, so it can be inspected, traced and reused.
Read the stories →A three-stage method that detects, locates and corrects mistakes in radiology reports, using medical knowledge to help language models do it.
Giving a multimodal model the places radiologists look and the regions they mark improves the reports it writes, without retraining it.
Invited Comment in The Lancet Digital Health on moving from diagnosing fetal damage to identifying at-risk fetuses early in labour.
Ontology rankers remain useful for rare-disease diagnosis because each candidate can be traced to matched patient phenotypes. Large language models (LLMs) can generate differential diagnoses from...
Given a patient's clinical findings, a diagnostic system ranks possible diseases and must decide when to endorse its first prediction or defer it for review. This...
AI systems increasingly retrieve from records that revise themselves - issue threads, encyclopedic histories, policy logs, and long conversations. The challenge is not only finding relevant...
Geriatric syndromes (GS), such as falls, dementia, delirium and malnutrition, are complex clinical conditions affecting older adults which involve multiple organ systems and have major impact...
We present the design and implementation of a data curation framework to generate a large-scale clinical brain imaging dataset suitable for artificial intelligence (AI) enabled image...
We welcome collaborators, research visitors and prospective students. Get in touch to talk about working together.