No code machine learning: validating the approach on use-case for classifying clavicle fractures.

Area of Science:

• Medical Imaging
• Machine Learning
• Artificial Intelligence in Medicine

Background:

• Physician access to machine learning (ML) tools is limited by programming requirements.
• Developing AI models for medical image analysis typically requires specialized expertise.
• There is a need for user-friendly platforms enabling clinicians to leverage ML for diagnostic tasks.

Purpose of the Study:

• To create and evaluate a no-code machine learning (NML) platform for physicians.
• To enable non-programming physicians to develop and test NML models.
• To classify clavicle radiographs for fracture presence using the developed NML platform.

Main Methods:

• A retrospective study utilized 4135 clavicle radiographs from 2039 patients across 13 hospitals.
• The NML platform automatically retrieved DICOM images via web access to hospital archives.
• The platform trained an ML model, providing performance metrics like sensitivity, specificity, and AUC.

Main Results:

• The NML platform successfully retrieved 94.7% of eligible radiographs.
• The trained ML model achieved 90% sensitivity, 87% specificity, and 88% accuracy in identifying clavicle fractures.
• The model demonstrated a high diagnostic performance with an AUC of 0.95.

Conclusions:

• A no-code machine learning platform is feasible for physicians to create diagnostic models.
• The NML platform facilitates the development and testing of ML models using multicenter imaging datasets.
• This technology can empower clinicians to utilize AI for radiograph classification, such as detecting clavicle fractures.