Clinical research on artificial intelligence medical diagnostic devices: A scoping review

Abstract

Artificial intelligence medical diagnostic devices (AIMDDs) show strong potential but face barriers to clinical use, emphasizing the need for rigorous clinical research. We assessed current AIMDD research, key challenges, and future directions. A scoping review followed Arksey and O'Malley's methodological framework and the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews guidelines. PubMed, Web of Science Core Collection, and the Cochrane Database of Systematic Reviews (January 2020–December 2024) were searched on AIMDD design, implementation, and evaluation. Two independent researchers screened and extracted data from the literature using predefined criteria. Ninety-seven articles met the inclusion criteria. Machine learning and deep learning approaches dominated across diverse disease fields, with oncology being the most frequent (41 %). The key challenges identified include insufficient quantity, quality, representativeness, and diversity of data; research designs that do not adequately address clinical needs; poor patient selection; poorly defined gold standards; lack of external and prospective validation; and a disconnect between validation strategies and clinical practice. Additionally, issues such as the “black box” phenomenon, overfitting, and data privacy concerns hinder clinical translation. Completeness and standardization of reporting were also found to be lacking. Significant challenges remain in the development and clinical application of AIMDD. To facilitate their clinical translation, improvements are needed in dataset optimization, clinically driven research design, development of evaluation frameworks, enhanced interpretability, and standardized reporting and validation of algorithms. • This study used a scoping review to synthesize evidence in the complex, high-volume AI medical diagnosis field. • Oncology was the most studied discipline, comprising 41 % of all included publications. • Key barriers include poor data quality, limited external validation, and incomplete methodological reporting. • Future work should focus on dataset optimization, clinician-led design, explainable AI, and standardized validation.

Ähnliche Arbeiten