Enrico Coiera

A. Baskar, Ashwin Srinivasan, Michael Bain et al.

Machine Learning (ML) has emerged as a powerful form of data modelling with widespread applicability beyond its roots in the design of autonomous agents. However, relatively little attention has been paid to the interaction between people and ML systems. In this paper we view interaction between humans and ML systems within the broader context of communication between agents capable of prediction and explanation. We formalise the interaction model by taking agents to be automata with some special characteristics and define a protocol for communication between such agents. We define One- and Two-Way Intelligibility as properties that emerge at run-time by execution of the protocol. The formalisation allows us to identify conditions under which run-time sequences are bounded, and identify conditions under which the protocol can correctly implement an axiomatic specification of intelligible interaction between a human and an ML system. We also demonstrate using the formal model to: (a) identify instances of One- and Two-Way Intelligibility in literature reports on humans interacting with ML systems providing logic-based explanations, as is done in Inductive Logic Programming (ILP); and (b) map interactions between humans and machines in an elaborate natural-language based dialogue-model to One- or Two-Way Intelligible interactions in the formal model.

Ashwin Srinivasan, Michael Bain, Enrico Coiera

Recent engineering developments in specialised computational hardware, data-acquisition and storage technology have seen the emergence of Machine Learning (ML) as a powerful form of data analysis with widespread applicability beyond its historical roots in the design of autonomous agents. However -- possibly because of its origins in the development of agents capable of self-discovery -- relatively little attention has been paid to the interaction between people and ML. In this paper we are concerned with the use of ML in automated or semi-automated tools that assist one or more human decision makers. We argue that requirements on both human and machine in this context are significantly different to the use of ML either as part of autonomous agents for self-discovery or as part statistical data analysis. Our principal position is that the design of such human-machine systems should be driven by repeated, two-way intelligibility of information rather than one-way explainability of the ML-system's recommendations. Iterated rounds of intelligible information exchange, we think, will characterise the kinds of collaboration that will be needed to understand complex phenomena for which neither man or machine have complete answers. We propose operational principles -- we call them Intelligibility Axioms -- to guide the design of a collaborative decision-support system. The principles are concerned with: (a) what it means for information provided by the human to be intelligible to the ML system; and (b) what it means for an explanation provided by an ML system to be intelligible to a human. Using examples from the literature on the use of ML for drug-design and in medicine, we demonstrate cases where the conditions of the axioms are met. We describe some additional requirements needed for the design of a truly collaborative decision-support system.

David Fraile Navarro, Farah Magrabi, Enrico Coiera

Ramaswamy et al. reported in \textit{Nature Medicine} that ChatGPT Health under-triages 51.6\% of emergencies, concluding that consumer-facing AI triage poses safety risks. However, their evaluation used an exam-style protocol -- forced A/B/C/D output, knowledge suppression, and suppression of clarifying questions -- that differs fundamentally from how consumers use health chatbots. We tested five frontier LLMs (GPT-5.2, Claude Sonnet 4.6, Claude Opus 4.6, Gemini 3 Flash, Gemini 3.1 Pro) on a 17-scenario partial replication bank under constrained (exam-style, 1,275 trials) and naturalistic (patient-style messages, 850 trials) conditions, with targeted ablations and prompt-faithful checks using the authors' released prompts. Naturalistic interaction improved triage accuracy by 6.4 percentage points ($p = 0.015$). Diabetic ketoacidosis was correctly triaged in 100\% of trials across all models and conditions. Asthma triage improved from 48\% to 80\%. The forced A/B/C/D format was the dominant failure mechanism: three models scored 0--24\% with forced choice but 100\% with free text (all $p < 10^{-8}$), consistently recommending emergency care in their own words while the forced-choice format registered under-triage. Prompt-faithful checks on the authors' exact released prompts confirmed the scaffold produces model-dependent, case-dependent results. The headline under-triage rate is highly contingent on evaluation format and should not be interpreted as a stable estimate of deployed triage behavior. Valid evaluation of consumer health AI requires testing under conditions that reflect actual use.

Dana Rezazadegan, Shlomo Berkovsky, Juan C. Quiroz et al.

Speech summarisation techniques take human speech as input and then output an abridged version as text or speech. Speech summarisation has applications in many domains from information technology to health care, for example improving speech archives or reducing clinical documentation burden. This scoping review maps the speech summarisation literature, with no restrictions on time frame, language summarised, research method, or paper type. We reviewed a total of 110 papers out of a set of 153 found through a literature search and extracted speech features used, methods, scope, and training corpora. Most studies employ one of four speech summarisation architectures: (1) Sentence extraction and compaction; (2) Feature extraction and classification or rank-based sentence selection; (3) Sentence compression and compression summarisation; and (4) Language modelling. We also discuss the strengths and weaknesses of these different methods and speech features. Overall, supervised methods (e.g. Hidden Markov support vector machines, Ranking support vector machines, Conditional random fields) performed better than unsupervised methods. As supervised methods require manually annotated training data which can be costly, there was more interest in unsupervised methods. Recent research into unsupervised methods focusses on extending language modelling, for example by combining Uni-gram modelling with deep neural networks. Protocol registration: The protocol for this scoping review is registered at https://osf.io.

Juan C Quiroz, Liliana Laranjo, Catalin Tufanaru et al.

Bayesian modelling and statistical text analysis rely on informed probability priors to encourage good solutions. This paper empirically analyses whether text in medical discharge reports follow Zipf's law, a commonly assumed statistical property of language where word frequency follows a discrete power law distribution. We examined 20,000 medical discharge reports from the MIMIC-III dataset. Methods included splitting the discharge reports into tokens, counting token frequency, fitting power law distributions to the data, and testing whether alternative distributions--lognormal, exponential, stretched exponential, and truncated power law--provided superior fits to the data. Results show that discharge reports are best fit by the truncated power law and lognormal distributions. Our findings suggest that Bayesian modelling and statistical text analysis of discharge report text would benefit from using truncated power law and lognormal probability priors.

Zubair Shah, Paige Martin, Enrico Coiera et al.

Background: Studies examining how sentiment on social media varies depending on timing and location appear to produce inconsistent results, making it hard to design systems that use sentiment to detect localized events for public health applications. Objective: The aim of this study was to measure how common timing and location confounders explain variation in sentiment on Twitter. Methods: Using a dataset of 16.54 million English-language tweets from 100 cities posted between July 13 and November 30, 2017, we estimated the positive and negative sentiment for each of the cities using a dictionary-based sentiment analysis and constructed models to explain the differences in sentiment using time of day, day of week, weather, city, and interaction type (conversations or broadcasting) as factors and found that all factors were independently associated with sentiment. Results: In the full multivariable model of positive (Pearson r in test data 0.236; 95\% CI 0.231-0.241) and negative (Pearson r in test data 0.306; 95\% CI 0.301-0.310) sentiment, the city and time of day explained more of the variance than weather and day of week. Models that account for these confounders produce a different distribution and ranking of important events compared with models that do not account for these confounders. Conclusions: In public health applications that aim to detect localized events by aggregating sentiment across populations of Twitter users, it is worthwhile accounting for baseline differences before looking for unexpected changes.

Didi Surian, Adam G. Dunn, Liat Orenstein et al.

Clinical trial registries can be used to monitor the production of trial evidence and signal when systematic reviews become out of date. However, this use has been limited to date due to the extensive manual review required to search for and screen relevant trial registrations. Our aim was to evaluate a new method that could partially automate the identification of trial registrations that may be relevant for systematic review updates. We identified 179 systematic reviews of drug interventions for type 2 diabetes, which included 537 clinical trials that had registrations in ClinicalTrials.gov. We tested a matrix factorisation approach that uses a shared latent space to learn how to rank relevant trial registrations for each systematic review, comparing the performance to document similarity to rank relevant trial registrations. The two approaches were tested on a holdout set of the newest trials from the set of type 2 diabetes systematic reviews and an unseen set of 141 clinical trial registrations from 17 updated systematic reviews published in the Cochrane Database of Systematic Reviews. The matrix factorisation approach outperformed the document similarity approach with a median rank of 59 and recall@100 of 60.9%, compared to a median rank of 138 and recall@100 of 42.8% in the document similarity baseline. In the second set of systematic reviews and their updates, the highest performing approach used document similarity and gave a median rank of 67 (recall@100 of 62.9%). The proposed method was useful for ranking trial registrations to reduce the manual workload associated with finding relevant trials for systematic review updates. The results suggest that the approach could be used as part of a semi-automated pipeline for monitoring potentially new evidence for inclusion in a review update.

Adam G. Dunn, Enrico Coiera, Florence Bourgeois

Objectives: Trial registries can be used to measure reporting biases and support systematic reviews but 45% of registrations do not provide a link to the article reporting on the trial. We evaluated the use of document similarity methods to identify unreported links between ClinicalTrials.gov and PubMed. Study Design and Setting: We extracted terms and concepts from a dataset of 72,469 ClinicalTrials.gov registrations and 276,307 PubMed articles, and tested methods for ranking articles across 16,005 reported links and 90 manually-identified unreported links. Performance was measured by the median rank of matching articles, and the proportion of unreported links that could be found by screening ranked candidate articles in order. Results: The best performing concept-based representation produced a median rank of 3 (IQR 1-21) for reported links and 3 (IQR 1-19) for the manually-identified unreported links, and term-based representations produced a median rank of 2 (1-20) for reported links and 2 (IQR 1-12) in unreported links. The matching article was ranked first for 40% of registrations, and screening 50 candidate articles per registration identified 86% of the unreported links. Conclusions: Leveraging the growth in the corpus of reported links between ClinicalTrials.gov and PubMed, we found that document similarity methods can assist in the identification of unreported links between trial registrations and corresponding articles.