Agreement between an artificial intelligence video-analysis platform and human evaluators scoring emergency medical technician psychomotor skills: a fully crossed reliability study
Year: 2026
Type: preprint
Abstract: Objectives Psychomotor competency verification for emergency medical technician (EMT) candidates is set by each state and, where states approve program-verified competency, rests on programs and their local evaluators. Commercial artificial intelligence (AI) video-analysis platforms are marketed to programs, but none has been evaluated against human evaluators on EMT skills. We measured the inter-rater reliability of certified evaluators scoring identical video-recorded performances, then evaluated one platform against that baseline. Methods Five human evaluators and one AI platform independently scored every video of three skills in a fully crossed design: cervical collar application (15 performances), bag-valve-mask ventilation (14), and medical patient assessment (15), using identical dichotomous checklists with critical-failure criteria. We computed item-level Fleiss κ, Krippendorff α, and Gwet AC1; absolute-agreement intraclass correlation coefficients (ICC) for percentage scores; and κ for critical-failure and pass/fail decisions, for the human panel alone and with the platform as a sixth rater. Results Human item-level agreement was moderate on all three skills (κ 0.41–0.47). Single-rater score reliability was moderate (ICC 0.54–0.69) against 0.85–0.92 for the five-rater mean. Agreement on an objective 75% threshold was moderate (κ 0.54–0.57), but agreement on the recorded pass/fail decision was fair (κ 0.30–0.39) and on the critical-failure determination near zero for cervical collar (0.03) and medical assessment (0.12). Platform alignment tracked rubric observability, not task complexity: correlation with the five-human mean was 0.857 for cervical collar, exceeding every human evaluator, −0.173 for bag-valve-mask, and 0.664 for medical assessment. Adding the platform improved panel reliability for cervical collar and reduced it for the others. Conclusions The platform rated a discrete, visually observable skill as well as the panel it joined and degraded where credit depended on rates, volumes, or verbalization. Formative practice feedback and adjunct rating are defensible now; autonomous summative use is not. The human baseline is least reliable precisely where it decides outcomes.
Source: medRxiv
Authors: Joel H. Otte, Autumn Cartagena
Institution: Des Moines Area Community College
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7204806104