Source-linked AI summary
Defining and Categorising Human-AI Interactions in Clinical Trials: A Multidimensional Human-AI Classification Approach
Sandra Woolley, Tim Collins, Khalid Khattak, Illia Chernomorets, Ariane Arevalo, Chris Richardson
TL;DR
AI-related clinical-trial records vary in how they describe AI methods and human involvement, making consistent comparison difficult. This study develops a multidimensional framework and preliminarily examines LLM-assisted categorisation; the findings suggest LLMs may assist classification, while human judgement remains important for incomplete or ambiguous records.
Problem
Clinical-trial records vary in their descriptions of AI methods and human involvement, and reporting reviews identify persistent omissions about AI interventions.
Source: §1 · pp. 1–2Method
The study develops a multidimensional framework for trial-record categorisation and applies it to purposively sampled trials selected because categorisation was difficult.
Source: §7 · p. 12, §8 · p. 13Results
Across selected trials, AI was most often categorised as a tool, AI-first human-in-the-loop was the most frequent configuration, and health professionals were the most frequent interacting group.
Source: §8 · p. 13Takeaways & Limitations
LLMs may support completion of trial records by prompting additional detail and flagging ambiguous descriptions, while human judgement remains important for incomplete or ambiguous records.
Source: §7 · p. 13, §8 · p. 13The trial set was purposively selected to include difficult cases, and records sometimes lacked details about AI methods, interaction sequencing or human oversight.
Source: §7 · pp. 12–13
Abstract
from arXiv · showhide
This paper examines human-AI interactions (HAIIs) in clinical trials and presents a multidimensional categorisation framework that classifies interactions according to AI tasks, human-AI relationships, interaction configurations and interacting human groups. We define HAII, examine existing taxonomies and extend existing categorisation approaches through this novel multidimensional framework. We purposively sampled 15 clinical trials from a previously reported dataset. Each trial was independently categorised by two human reviewers and six large language model (LLM) classifiers. The proposed categorisation provides a structured method for the consistent identification, comparison and synthesis of human-AI interactions across clinical-trial records. The framework is intended to support more consistent comparison and synthesis of AI-related clinical trials and to make explicit the different forms of human involvement associated with AI interventions. The results demonstrate the potential for LLM-assisted categorisation while indicating the continuing importance of human judgement where trial records are incomplete or ambiguous. The principal contribution is a proposed multidimensional framework that brings together AI tasks, human-AI relationships, interaction configurations and interacting human groups within a single approach designed for clinical-trial records. Its significance lies in its potential to support more systematic identification, comparison and synthesis of how humans and AI interact in clinical trials.
1. Introduction
Clinical-trial records make AI-related trials difficult to identify, compare and synthesise because descriptions of AI and human involvement vary. Existing reporting guidance improves trial descriptions, but a framework is also needed to compare how humans and AI interact across trials.
1. Introduction
- Clinical-trial records vary in descriptions of AI methods, human involvement, inputs and outputs, and clinical context, limiting identification, comparison and synthesis.A structured categorisation is needed to compare interaction tasks, contribution order, exchange configurations, human groups and oversight.Source: §1 · pp. 1–2
- SPIRIT-AI and CONSORT-AI extend protocol and completed-RCT reporting guidance to cover AI interventions, intended use, settings, inputs and outputs, users, and human-AI interaction.SPIRIT-AI applies to trial protocols; CONSORT-AI applies to reports of completed randomised controlled trials.Source: §1 · pp. 1–2
- Reviews found persistent omissions: no reviewed ophthalmology trials reported every CONSORT-AI item, and oncology reports omitted details such as algorithm versions, input data and error assessment.The oncology review also identified omissions about poor-quality data handling.Source: §1 · p. 2
- AI clinical studies often fall outside RCT-focused guidance: Lam et al. found 39 of 11,839 articles included RCTs, while Wang et al. reported 2.4% of AI clinical studies were RCTs.Many trials also predate SPIRIT-AI and CONSORT-AI guidance.Source: §1 · p. 2
2. Related Work
Prior work addresses AI trial reporting, interaction patterns, task allocation and evaluation, but does not provide a multidimensional framework for categorising human-AI interactions in clinical-trial records. Incomplete reporting and limited information about human involvement motivate a framework for comparison across trials.
2. Related Work
- Plana et al. identified 41 published machine-learning RCTs, none fully compliant with CONSORT-AI, and noted limited reporting of race and ethnicity.Only 27% of trials reported race and ethnicity; common omissions included performance-error analysis and algorithm availability.Source: §2 · p. 2
- Gomez et al. reviewed 105 empirical studies and found simple AI assistance predominated, while reciprocal and adaptive collaboration was less common.Their taxonomy emphasises sequencing, initiative and interaction structure.Source: §2 · p. 2
- You and Li reviewed 85 publications and found collaboration patterns associated with clinical-task risk and automation, alongside a division of labour between doctors and AI.Their review also considered technical performance, clinical efficiency and user experience.Source: §2 · p. 3
- Existing reviews address interaction patterns, task allocation and evaluation, but do not provide a multidimensional framework designed for categorising interactions in clinical-trial records.The reviewed taxonomy principally describes interaction processes rather than human participants, AI roles or clinical context.Source: §2 · p. 3
- Earlier work categorised AI use and interacting human groups, but trial records often left AI functionality, interaction sequencing and oversight ambiguous.These observations motivated the present framework; AI is also used in trial processes such as recruitment, analysis and outcome prediction.Source: §2 · pp. 2–3
3. Study Aims and Research Questions
The study aims to develop and test a multidimensional framework for categorising human-AI interactions in clinical trials and to examine LLM-assisted categorisation. Its research questions cover the definition of interaction, relevant classification dimensions and categories, their application to trial records, and preliminary comparison of human and LLM classifications.
3. Study Aims and Research Questions
- The study aims to analyse human-AI interaction in clinical trials and observe trends in its use.The stated motivation is to understand these interactions in clinical trials.Source: §3 · p. 3
3.1 Study Aims
- The study builds on categorisation of interacting human groups by developing and testing a framework spanning AI tasks, human-AI relationships and interaction configurations.The additional dimensions were informed by existing taxonomies and refined through observation, pilot categorisation and application to trials.Source: §3.1 · pp. 3–4
3.2 Research Questions
- The research questions ask how HAII is defined, which dimensions and categories classify it, how they apply to trial records, and what human-LLM classification comparisons show.The questions include both a definition and the minimal unit of interaction.Source: §3.2 · p. 4
4. What is Human-AI Interaction?
The paper defines human-AI interaction broadly enough to include one-way influence: AI outputs can affect people, and human inputs can affect AI systems. It treats a human’s perception of an AI output or an AI’s detection of human input as a minimal interaction, even without an observable response.
4.1. Definitions of Human-AI Interaction
- The paper defines HAII as occurring when AI or its output influences human perception, interpretation, judgement or action, or human input influences AI operation or output.This includes a clinician whose judgement is influenced by an automatically presented AI assessment, even without clinician input to the system.Source: §4.1 · p. 4
- The definition includes one-way human-to-AI or AI-to-human influence, unlike reciprocal definitions that require two-way interaction.AI characteristics such as prediction, adaptation and probabilistic reasoning may also make conventional HCI concepts insufficiently precise.Source: §4.1 · p. 4
4.2. What is a minimal unit of Human-AI Interaction?
- On the human side, an AI output must enter the person’s experience; presentation without a realised human encounter is not enough.Perception may affect attention, interpretation or judgement without producing observable behaviour.Source: §4.2 · p. 5
- AI detection can count as an interaction when the system excludes detected input or changes internal processing without a visible output.An input that is presented but not detected or registered by the AI does not meet the human-to-AI boundary described in the table.Source: §4.2 · p. 5
- A minimal HAII occurs when a person perceives an AI output or an AI detects human input, even if no externally observable response follows.The lower boundaries are considered from both AI-to-human and human-to-AI perspectives.Source: §4.2 · p. 5
5. Methodology
The study developed and refined a human-AI interaction framework, then applied it to 15 purposively selected clinical trials using two human reviewers and six LLM classifiers.
5. Methodology
- The framework was developed and refined using existing taxonomies, prior research, and pilot categorisation of clinical-trial records.Its categories are not mutually exclusive, so a trial record may receive multiple categories within a dimension.Source: §5 · p. 5, §5.1 · p. 6
5.2 Selected Trials and Selected LLMs
- The 15 purposively selected trials varied in AI type, clinical application, human groups, and interaction configuration, and included records expected to pose classification challenges.For AI type, reviewers relied only on explicit trial-record descriptions and left the classification unspecified when methodology was not stated.Source: §5.2 · p. 6
- Two expert human reviewers and six LLMs independently classified each trial against the framework, assigning True, False, or Uncertain and a confidence rating.The six LLMs comprised five small locally hosted models and one cloud-hosted frontier model for comparison.Source: §5.2 · pp. 6–7
6. Results
Across 15 trials, eight classifiers produced 3,480 category-level classifications. AI as a tool was the most frequently classified category as True, while agreement and human–LLM consensus differences varied across categories.
6. Results
- The eight classifiers produced 3,480 category-level classifications across the 15 trials, reported as individual responses, consensus classifications, and agreement.Figures 1(i)–(iii) show category-level classifications for trials 1–5, 6–10, and 11–15, respectively.Source: §6 · p. 8
6.1 Classification Frequencies
- AI as a tool was classified True in 114/120 classifications, the highest frequency; Treat was False in 108/120, and AI-follows human-in-the-loop was Uncertain in 47/120.The 120 classifications represent eight classifiers across 15 trials.Source: §6.1 · p. 8, Table 6 · p. 12
- By consensus, AI as a tool was True in all 15 trials, followed by AI-first human-in-the-loop interaction in 13/15 and classification and decision-support resource in 11/15 each.Health professional–AI interaction was True in 10/15 trials, while detection or assessment and human-on-the-loop interaction were each True in 9/15.Source: §6.1 · p. 8
- Three categories received no True classification by consensus in any trial: Treat as the AI task, AI as a collaborator, and Other human-AI interaction.Each category nevertheless received at least one True classification from an individual classifier.Source: §6.1 · p. 8
6.2 Classifier Agreement
- AI as a tool had the highest mean agreement alongside predict/prognose at 7.6/8; human-on-the-loop was lowest at 4.1/8.AI-follows human-in-the-loop followed at 4.9/8, and three of the five lowest-agreement categories concerned interaction configuration.Source: §6.2 · p. 8, Table 7 · p. 12
6.3 Human and LLM Classifications
- Human reviewers assigned Uncertain more often than LLMs, and opposing True and False consensus classifications were most frequent for AI-first human-in-the-loop interaction (6/15 trials).Opposing consensus classifications occurred in 4/15 trials for human-on-the-loop and 2/15 for human-out-of-the-loop interaction.Source: §6.3 · p. 8
7. Discussion
Discussion shows that ambiguity in trial records complicates categorisation, especially around human oversight, while applying the framework prompted clarifications to its categories and definitions.
7. Discussion
- Iterative piloting exposed subjective, ambiguous narratives; patient-AI interaction was classified as absent when patients only supplied data and did not engage with AI or its outputs.Source: §7 · p. 12
- In NCT00647413, the record left unclear whether the expert system contributed to counselling or whether a health professional reviewed its tailored letters.Source: §7 · p. 12
- For NCT04289025, nurses checked exercise feasibility and removed unsuitable exercises, but the record left unclear whether AI-generated reports were reviewed or who recommended the exercises.Source: §7 · pp. 12–13
- Trials evaluating AI development posed categorisation challenges, prompting a distinct “AI in development/training/optimisation” category.Source: §7 · p. 13
- AI performance measures such as sensitivity and recall were not treated as scoring/ranking tasks; categories instead reflect the system’s clinical or health-related function.Source: §7 · p. 13
- Findings suggest LLM tools may support trial-record completion by prompting additional detail and flagging ambiguous descriptions of AI methods and human-AI interactions.Source: §7 · p. 13
8. Conclusions and Further Work
The study proposes a multidimensional framework for categorising human-AI interactions in clinical trials and reports preliminary human and LLM classifications. The selected trials showed common interaction patterns alongside substantial ambiguity in trial-record reporting.
8. Conclusions and Further Work
- The framework combines AI tasks, human-AI relationships, interaction configurations, and interacting human groups to categorise clinical-trial interactions.Source: §8 · p. 13
- Across selected trials, AI was most often categorised as a tool, AI-first human-in-the-loop was the most frequent configuration, and health professionals were the most frequently identified interacting group.Source: §8 · p. 13
- Application revealed substantial ambiguity in records about AI functionality, interaction sequencing, and human oversight; LLMs may aid classification, but human judgement remains important when records are incomplete or ambiguous.Source: §8 · p. 13
- Further work will refine category definitions and decision rules, include an AI-development category, and evaluate the framework first across 100 trials and then a wider dataset.Source: §8 · p. 13
- The conclusion calls for practical support for consistent description and categorisation of human-AI interactions, in addition to reporting guidance.Source: §8 · p. 13
A. AI prompt (markdown format):
The prompt asks a classifier to inspect a ClinicalTrials.gov record and categorise AI's role using specified dimensions. It also sets rules for confidence, evidence-based classification, and a structured summary.
A. AI prompt (markdown format):
- The prompt frames the task as classifying AI's role in a ClinicalTrials.gov trial record using the description, intervention details, and other record fields.Source: A. AI prompt (markdown format) · pp. 15–16
- It asks classifiers to identify AI tasks and human-AI relationships, with category lists covering clinical functions and roles such as tool, educator, decision support, collaborator, or intervention.Source: A. AI prompt (markdown format) · p. 15
- Interaction categories include AI-first or AI-follow decisions, iterative or integrated exchanges, human oversight, and autonomous action without oversight.Source: A. AI prompt (markdown format) · pp. 15–16
- The prompt asks classifiers to identify interacting human groups and quote the AI type from the trial record, using “unknown” when the type is not clearly stated.Source: A. AI prompt (markdown format) · p. 16
- The output includes a short summary describing the AI task, AI type, human interaction, interacting people, and allocated categories.Source: A. AI prompt (markdown format) · p. 16
- Classifiers assign an overall low, medium, or high confidence and use true, false, or uncertain only according to whether the record makes a classification clear, without assumptions or inferences.Source: A. AI prompt (markdown format) · p. 16