Using "big data" to guide implementation of a web and mobile adaptive learning platform for medical students.

Background: Adaptive learning platforms (ALPs) can revolutionize medical education by making learning more efficient, but their potential has not been realized because students do not use them persistently.

Methods: We applied educational data mining methods to study United States medical students who used an ALP called Osmosis ( www.osmosis.org ) from 1 August 2014 to 31 July 2015. Multivariate logistic regressions modeled persistence on Osmosis as the dependent variable and Osmosis-collected variables as predictors.

Results: The 6787 students included in our analysis responded to a total of 887,193 items, with 2138 (31.5%) using Osmosis persistently. Number of items per student, mobile device use, subscription payment, and group membership were independently associated with persisting (p < 0.001 in all models). Persistent users rated quality more favorably (p < 0.01) but were not more confident in answer selections (p = 0.80). While persisters were more accurate than non-persisters (55% (SD 18%) vs 52% (SD 22%), p < 0.001), after adjusting for number of items, lower accuracy was associated with persistent use (OR 0.93 [95% CI 0.90-0.97], p < 0.01).

Conclusions: Our study of a large sample of U.S. medical students illustrates big data medical education research and provides guidance for improving implementation of ALPs and further investigation.