Application of deep machine learning in dental education: a systematic review of effectiveness in dental students'...

Deep machine learning (DML) technologies, including convolutional neural networks (CNN) and transformer-based models, are increasingly used to support teaching and diagnostic reasoning in dental education. However, the evidence regarding their effectiveness in improving cognitive learning outcomes among dental students remains fragmented and highly heterogeneous.

To systematically evaluate the impact of DML-based educational and diagnostic interventions on cognitive learning outcomes among dental students, compared with conventional or non-artificial intelligence methods.

PRISMA 2020 guidelines were followed strictly. Comprehensive searches were conducted across PubMed/MEDLINE, Scopus, Web of Science, Cochrane CENTRAL, IEEE Xplore, Google Scholar, and preprint servers like medRxiv, bioRxiv and arXiv. Eligible studies included randomized trials, quasi-experimental designs, and controlled pre-post studies involving dental students, evaluating DML interventions with at least one measurable cognitive outcome like diagnostic accuracy, reasoning, and cognitive load. Two reviewers independently screened records and extracted data; inter-reviewer agreement exceeded κ > 0.80. Risk of bias was assessed using RoB 2.0, ROBINS-I, and AXIS tools. Certainty of evidence was appraised using GRADE. Due to marked heterogeneity in outcomes and model types, meta-analysis was not feasible.

Of 145 records, 16 studies met the inclusion criteria. Interventions consisted primarily of CNN-based radiographic systems, AI-assisted diagnostic tools, and transformer-based large language models. Across diagnostic accuracy outcomes, most studies reported improvements in sensitivity, F1 scores, and pattern recognition when using DML tools compared with traditional methods. Evidence for improvements in cognitive reasoning, assignment quality, or cognitive load was mixed and often limited by small samples, subjective evaluations, and nonrandomized designs. One study reported lower post-test scores among students using LLM-generated assistance compared with conventional teaching. No study assessed long-term learning retention. Overall risk of bias ranged from low to moderate, and the certainty of evidence was moderate for diagnostic accuracy and low for all other cognitive outcomes.

Heterogeneity in interventions, outcomes, and assessment methods precluded meta-analysis. Some outcomes were subjective, and long-term retention was rarely assessed.

DML-assisted interventions show promising but preliminary potential to enhance specific cognitive domains, particularly diagnostic accuracy in dental education. However, the overall evidence remains limited by study heterogeneity, small samples, and methodological weaknesses. Current findings support the adjunctive, not substitutive, use of DML tools in dental curricula. High-quality multicenter RCTs with standardized cognitive outcome measures and longitudinal follow-up are needed to determine the sustained educational value and practical feasibility of DML integration.