Comparison of state-of-the-art machine and deep learning algorithms to classify proximal humeral fractures using radiology text

Introduction Proximal humeral fractures account for a significant proportion of all fractures. Detailed accurate classification of the type and severity of the fracture is a key component of clinical decision making, treatment and plays an important role in orthopaedic trauma research. This research aimed to assess the performance of Machine Learning (ML) multiclass classification algorithms to classify proximal humeral fractures using radiology text data. Materials and Methods Data from adult (16 + years) patients admitted to a major trauma centre for management of their proximal humerus fracture from January 2010 to January 2019 were used (1,324). Six input text datasets were used for classification: X-ray and/or CT scan reports (primary) and concatenation of patient age and/or patient sex. One of seven Neer class labels were classified. Models were evaluated using accuracy, recall, precision, F1, and One-versus-rest scores. Results A number of statistical ML algorithms performed acceptably and one of the BERT models, exhibiting good accuracy of 61% and an excellent one-versus-rest score above 0.8. The highest precision, recall and F1 scores were 50%, 39% and 39% respectively, being considered reasonable scores with the sparse text data used and in the context of machine learning. Conclusion ML and BERT algorithms based on routine unstructured X-ray and CT text reports, combined with the demographics of the patient, show promise in Neer classification of proximal humeral fractures to aid research. Use of these algorithms shows potential to speed up the classification task and assist radiologist, surgeons and researchers.