Learning data partitioning strategies for artificial intelligence studies in orthodontics
The Angle Orthodontist · 2026-08-31
DOI: 10.2319/010126-1.1 ↗Abstract
ABSTRACT Objectives To gain insight into which data partitioning design should be used for artificial intelligence (AI) studies in orthodontics. Materials and Methods Original research articles published from January 1, 2021, to December 31, 2025, in six orthodontic specialty journals, were screened on the journal websites. AI studies were defined as those that applied a deep learning algorithm, either when the study explicitly stated that it was an AI study or when it was conducted using a commercial or preexisting AI software product. Out of a total of 2194 original research articles, 128 were identified as AI studies. While scrutinizing 128 AI studies, the research topic, study formulation, data characteristics, and data partitioning methods were examined. Results The proportion of AI studies has increased over the past several years, from 4.3% in 2021 to 9.5% in 2025. Across various research topics and data characteristics, the data partitioning pattern also varied considerably, and no clear preference for a specific strategy could be identified. Among various data partitioning methods, 8:1:1 and 4:1 partitioning were the most frequent as the three- and two-part partitioning methods, respectively. Among cross-validation methods, the leave-one-out cross-validation was the most frequent. Conclusions Although no clear or specific guidelines were found for splitting the data, cross-validation was recommended when the sample size was relatively small. Based on patterns observed in the reviewed literature, when sufficient data were available (eg, over 1000), any data partitioning design seemed plausible without significant concern.
EUVIMED