Executive summary
Universal MMI has built the first interview pass predictor for medical-school admissions — a tool that estimates an applicant’s likelihood of succeeding at their real interview based on how they perform in practice interviews on our platform. Where written exams such as the MCAT have long had score predictors, the admissions interview has had none. We set out to close that gap.
The predictor is built on 1,850 matched applicant results — practice-interview performance paired with the applicant’s actual interview outcome — combined with our own scoring formula. Every practice interview on Universal MMI is scored across six competency dimensions by both a peer examiner and our AI examiner, giving each applicant a rich, multi-station performance profile. Our formula translates that profile into a single, interpretable estimate: the probability of being admitted after the real interview.
This white paper describes how the data was collected, how the predictor works, and how we validate its accuracy as the dataset grows.
Introduction
For most medical, dental, and health-professions programs, the interview is the final and highest-stakes gate. By the time an applicant is invited, their grades and test scores have already been judged; the interview is where the offer is won or lost. The most common format is the Multiple Mini-Interview (MMI), a circuit of short, independent stations — ethical scenarios, role-plays, and discussion prompts — each scored by a different examiner. [1][3]
Preparing for the interview is uniquely difficult. Unlike a practice exam, an applicant cannot simply re-take it and watch a number go up. Feedback is scarce, scoring feels subjective, and there is no reliable way to answer the one question that matters most: “am I ready?” Applicants practise with friends or mentors, but those impressions are informal and inconsistent, and they rarely translate into a clear sense of standing.
A number of score predictors exist for written admissions and licensing exams, and they have proven genuinely useful to students. To our knowledge, no equivalent exists for the admissions interview. We saw the opportunity to build one: a transparent, data-backed estimate that gives applicants the confidence to decide where they stand and what to work on — well before interview day.
Method
Data collection
Universal MMI is a practice platform on which applicants complete live, timed interview stations — either peer-to-peer with another applicant or against our AI examiner. Each station is scored on the same six-dimension rubric used by trained examiners, so an applicant who completes several stations accumulates a detailed, repeated-measures record of performance rather than a single snapshot.
To build the predictor, we matched each applicant’s practice record with their self-reported real-world outcome — whether they were admitted or not admitted after their admissions interview. After cleaning and de-duplication, retaining the most recent verified record per applicant, this yielded our analysis set of 1,850 matched applicants spanning multiple programs and regions, collected between January 2019 and 2025.
What the predictor uses
The model draws on the structured signals captured during practice, summarised below.
The formula
Because an applicant’s stations are repeated measures of the same person over time, we treat the data the way longitudinal studies in education and healthcare do: each applicant has an underlying ability that the individual stations estimate with noise. [4] Our formula is a calibrated, weighted composite of the six dimension scores — weighted by how strongly each dimension related to real interview outcomes in the matched dataset — mapped onto a 0–100% probability of being admitted.
A deliberate property of the formula is that it behaves sensibly with little data. An applicant with only two or three completed stations receives an estimate that leans partly on patterns pooled from the full population; an applicant with many stations gets an estimate driven almost entirely by their own record. This keeps early predictions stable and lets them sharpen as the applicant practises more — the same logic that makes repeated-measures models robust.
Results
Because admission is a yes/no outcome, we evaluate the predictor as a classifier on a held-out test set the model never saw during fitting. We report three complementary measures:
- Accuracy — how often the predicted outcome matched the real outcome.
- AUC (area under the ROC curve) — how well the predicted probabilities separate applicants who were admitted from those who were not (0.5 = chance, 1.0 = perfect).
- Calibration — whether stated probabilities are truthful: of applicants told “70%,” roughly 70% should be admitted.
These figures come from a held-out split the model was not trained on. As the matched dataset grows, we re-fit and re-validate, and update these numbers accordingly.
Discussion
To our knowledge, this is the first predictor of medical-school interview success built on matched practice-and-outcome data. That “first-ever” status is also its main caveat: outcomes are self-reported, programs differ in how they score and weight interviews, and 1,850 applicants — while a strong start — is a foundation we intend to grow. As more applicants complete practice on Universal MMI and report their results, both the formula and its validated accuracy will improve.
Future versions will calibrate by station type and by region and program, since the bar for a competitive interview is not identical everywhere. Our aim is not to replace an applicant’s judgement but to give them something they have never had before: an honest, data-grounded read on their interview readiness, early enough to act on it.
References
- Eva, K. W., Rosenfeld, J., Reiter, H. I., & Norman, G. R. (2004). An admissions OSCE: the multiple mini-interview. Medical Education, 38(3), 314–326.
- Eva, K. W., Reiter, H. I., Trinh, K., Wasi, P., Rosenfeld, J., & Norman, G. R. (2009). Predictive validity of the multiple mini-interview for selecting medical trainees. Medical Education, 43(8), 767–775.
- Pau, A., Jeevaratnam, K., Chen, Y. S., Fall, A. A., Khoo, C., & Nadarajah, V. D. (2013). The multiple mini-interview (MMI) for student selection in health professions training — a systematic review. Medical Teacher, 35(12), 1027–1041.
- Verbeke, G., & Molenberghs, G. (2009). Linear Mixed Models in Practice: A SAS-Oriented Approach. Springer.