This paper focuses on human‑AI interactions (HAIIs) in clinical trials and introduces a multidimensional classification framework. The framework splits interactions into four dimensions: AI tasks, the nature of the human‑AI relationship, interaction configurations, and the groups of humans involved. It first defines HAII, reviews existing taxonomies, and highlights their shortcomings when applied to trial records. The proposed framework then combines the four dimensions to fully describe any AI‑mediated intervention in a trial.
A purposive sample of 15 trials from a previously reported dataset was used. Each trial was independently categorized by two human reviewers and by six large language model (LLM) classifiers. Results show that LLMs can rapidly assign consistent labels in most cases, yet human judgment remains essential when records are incomplete or ambiguous.
The main contribution is a unified language that enables systematic identification, comparison, and synthesis of AI‑related clinical trials. By making the forms of human involvement explicit, the framework aims to improve reproducibility and transparency of AI interventions in clinical research.
Review