Shomique Hayat completes UNIQplus research project with AI4DH

Shomique Hayat completed his 2026 UNIQplus research internship with the AI for Digital Health group at the Institute of Biomedical Engineering, University of Oxford. His project, “Reasoning enhancement in large language models for healthcare applications”, was supervised by Associate Professor Tingting Zhu, with guidance from Zhikang Chen.

 The project asked what moves a language model’s antibiotic recommendation: new clinical information, or simply being disagreed with. Using de-identified MIMIC-IV hospital records, Shomique assembled a fixed cohort of adult patients with Enterobacterales bloodstream infections. He built a Python and DuckDB pipeline and an experimental framework assigning doctor and pharmacist roles to the same language model. Initial recommendations used information available before the laboratory susceptibility results; those later results provided a retrospective reference for assessing whether the recommended antibiotic covered all relevant organisms in each case.

Under scripted pressure, the model frequently changed its recommendation towards meropenem, although losses of previously adequate coverage were uncommon. In a separate extended debate condition, a substantial share of initially adequate recommendations no longer met the coverage criterion afterwards. Self-review produced far fewer changes, although differences in instructions, context length and turn count prevent attributing the contrast to disagreement alone.

When genuine susceptibility results were supplied, more recommendations met the coverage criterion, including most of those that had not met it before. Some individual revisions nevertheless reduced coverage. Fixed single-drug policies also achieved high coverage, showing why a high score alone could not establish that the system was selecting treatment for the individual patient. These findings concern the selected retrospective cohort and tested model configurations. Susceptibility coverage does not establish the uniquely appropriate treatment, and no recommendation was used in patient care.

For Shomique, an important lesson was learning to value a result that challenged his original expectations. Simple baselines and feedback from Tingting prompted him to revise his interpretation of what the experiment could demonstrate. Zhikang’s advice to complete one small, concrete step at a time helped him turn ideas into experiments. Learning unfamiliar microbiology and presenting the work strengthened his confidence in explaining decisions, taking questions and acknowledging uncertainty.

He also found the placement immensely enjoyable. Lab conversations, research talks and college events made the summer much more than time spent running experiments. He particularly valued learning from people with different expertise and becoming more comfortable asking for help and questioning his own assumptions. The experience deepened his interest in the evaluation, reliability and interpretability of healthcare AI and strengthened his ambition to pursue doctoral research.

Shomique presented the project on 20 August 2026 at the UNIQplus end-of-programme conference at Cohen Quad, Exeter College.

We thank Shomique for his contribution over the summer and wish him every success in his future studies and research.