Amy Rouillard

Amy Rouillard, Sitwala Mundia, Linda Camara et al.

Evaluating medical AI systems using expert clinician panels is costly and slow, motivating the use of large language models (LLMs) as alternative adjudicators. Here, we evaluate an LLM jury composed of three frontier AI models scoring 3333 diagnoses on 300 real-world middle-income country (MIC) hospital cases. Model performance was benchmarked against expert clinician panel and independent human re-scoring panel evaluations. Both LLM and clinician-generated diagnoses are scored across four dimensions: diagnosis, differential diagnosis, clinical reasoning and negative treatment risk. For each of these, we assess scoring difference, inter-rater agreement, scoring stability, severe safety errors and the effect of post-hoc calibration. We find that: (i) the uncalibrated LLM jury scores are systematically lower than clinician panels scores; (ii) the LLM Jury preserves ordinal agreement and exhibits better concordance with the primary expert panels than the human expert re-score panels do; (iii) the probability of severe errors is lower in \lj models compared to the human expert re-score panels; (iv) the LLM Jury shows excellent agreement with primary expert panels' rankings. We find that the LLM jury combined with AI model diagnoses can be used to identify ward diagnoses at high risk of error, enabling targeted expert review and improved panel efficiency; (v) LLM jury models show no self-preference bias. They did not score diagnoses generated by their own underlying model or models from the same vendor more (or less) favourably than those generated by other models. Finally, we demonstrate that LLM jury calibration using isotonic regression improves alignment with human expert panel evaluations. Together, these results provide compelling evidence that a calibrated, multi-model LLM jury can serve as a trustworthy and reliable proxy for expert clinician evaluation in medical AI benchmarking.

Bruce A. Bassett, Amy Rouillard, Sitwala Mundia et al.

Background: Large language models (LLMs) are increasingly proposed for diagnostic support, but few evaluations use real-world multimodal inpatient data, particularly in low and middle-income country (LMIC) public hospitals. Methods: We conducted VALID, a retrospective evaluation of 539 multimodal inpatient cases from a tertiary public hospital in South Africa. Inputs included radiology imaging (CT, MRI, CXR) and reports, laboratory results, clinical notes, and vital signs. Expert panels adjudicated 300 cases (balanced and discordant subsets) to establish ground truth diagnoses, differentials, and reasoning. Ten multimodal LLMs generated zero-shot outputs. A calibrated three-model LLM Jury scored all outputs and routine ward diagnoses across diagnostic accuracy, differential quality, reasoning, and patient safety (>10,000 evaluations). Primary outcomes were composite scores ($S_3$, $S_4$) and win rates. Results: (i) LLM performance was tightly clustered (<15% variation) despite large cost differences; low-cost models performed comparably to top models. (ii) All LLMs significantly outperformed routine ward diagnoses on average diagnostic and safety scores. (iii) Top performance was achieved by GPT-5.1, followed by Gemini models. (vi) Adding radiology reports improved performance by 6%. (v) Diagnostic and reasoning scores were highly correlated ($ρ= 0.85$). (vi) Output rates varied (65-100%) due to input constraints. Results were robust across subsets and evaluation design. Conclusions: Across a real-world LMIC dataset, multimodal LLMs showed similar diagnostic performance despite large cost differences and outperformed routine care on average safety metrics. Affordability, robustness, and deployment constraints may outweigh marginal performance differences in LMIC settings.

Amandine M. Caut, Amy Rouillard, Beimnet Zenebe et al.

An important part of data science is the use of visualisations to display data in a way that is easy to digest. Visualisations often rely on underlying statistical or machine learning models -- ranging from basic calculations like category means to advanced methods such as principal component analysis of multidimensional datasets -- to convey insights. We introduce an analogous concept for word descriptions of data, which we call wordalisations. Wordalisations describe data in easy to digest words, without necessarily reporting numerical values from the data. We show how to create wordalisations using large language models, through prompt templates engineered according to a task-agnostic structure which can be used to automatically generate prompts from data. We show how to produce reliable and engaging texts on three application areas: scouting football players, personality tests, and international survey data. Using the model cards framework, we emphasise the importance of clearly stating the model we are imposing on the data when creating the wordalisation, detailing how numerical values are translated into words, incorporating background information into prompts for the large language model, and documenting the limitations of the wordalisations. We argue that our model cards approach is a more appropriate framework for setting best practices in wordalisation of data than performance tests on benchmark datasets.

Amandine M. Caut, Beimnet Zenebe, Amy Rouillard et al.

The rapid advancement and impressive capabilities of large language models (LLMs) have given rise to the field of prompt engineering, the practice of crafting inputs to guide LLMs toward high-quality, task-relevant outputs. A critical challenge facing the field is the lack of standardised prompt documentation and evaluation practices. Prompts can be long, complex and difficult to evaluate on subjective tasks. To address this challenge, we propose the use of prompt cards, structured summaries of prompt engineering practices inspired by the concept of model cards. Through prompt cards, the specific goals, considerations and steps taken during prompt engineering can be systematically documented and assessed. We present the prompt card approach and illustrate it on a specific task called wordalisation, in which structured numerical data is transformed into text. We argue that a well-structured prompt card can enable better reproducibility, transparency, improve prompt methodology and give an effective alternative to benchmarking for judging the quality of generated texts. By systemically capturing underlying model details, prompt intent, contextualisation strategies, evaluation practices and ethical considerations, prompt cards make explicit the often implicit design decisions that shape system behaviour. Documenting these choices is important as prompting increasingly involves complex pipelines with multiple moving parts.