JMIR Publications

ABSTRACT

Background:

Competition for UK Internal Medicine Training (IMT) has risen sharply, from 1.70 applications per post in 2015 to 5.27 in 2026. Selection is by multi-station interview, which tests structured clinical and professional reasoning delivered under time pressure rather than knowledge alone. One common way candidates prepare is peer role-play, but it is not equally available: some lack a suitable practice partner or train remotely or overseas, and others opt out because they are reluctant to rehearse in front of colleagues.

Objective:

To describe MedPrep, an iOS application that provides on-demand simulated interview practice with automated feedback, and to report early user engagement during the IMT 2026 selection window as an initial feasibility and acceptability evaluation.

Methods:

In MedPrep, candidates give a timed spoken response to an interview station prompt and receive automated structured feedback and a score from a multimodal large language model (LLM). We retrospectively extracted backend usage data on 22 May 2026 for all users who registered during the observation window (1 October 2025 to 13 February 2026). Developer and test accounts, and users who completed no stations, were excluded. Metrics are reported descriptively. AI-generated scores are reported as platform outputs and were not validated against human examiners.

Results:

Of 149 registered users, 37 (24.8%) completed at least one station (real users) and 26 (17.4%) completed more than one (active users). All selected the IMT track and were recruited through social media and word of mouth. Active users completed a median of 6 stations (IQR 2-25; maximum 71) over a median of 42 app sessions (IQR 13-100; maximum 272). Real users completed 472 stations across all five IMT activities. Per available station, practice was most intensive on the formulaic portfolio and presentation tasks (12.3 completions per station) and lowest on the much larger clinical bank (5.7). The most-completed station was the 2-minute presentation (25/37 users). Of 472 completions, 445 carried a stored AI score (mean 6.35/10, SD 1.35), concentrated in the developing (4-6) and competent (7-8) bands.

Conclusions:

MedPrep is a feasible and scalable way to deliver self-directed interview practice, with preliminary acceptability indicated by sustained voluntary use and organic recruitment. The automated feedback has not yet been validated and no learning or selection outcomes were measured. The next step is a controlled evaluation with validated outcomes and examiner-referenced scoring.