Artificial Intelligence promising for diagnosing hip fracture from hip radiographs: JAMA

A recent systematic review and meta-analysis by Johnathan Lex and team suggests that there is potential for Artificial intelligence (AI) in the diagnositic process of hip radiographs. The findings of this study were published in the Journal of American Medical Association.

Strong models made possible by artificial intelligence can be used to construct clinical diagnostic and prognostic tools for hip fractures; however, the effectiveness and potential consequences of these recently established algorithms are yet unclear. In order to compare the efficacy of AI algorithms built to detect hip fractures on radiography and determine postoperative clinical outcomes after hip fracture surgery to existing procedures, this study was carried out.

Using the databases for Embase, MEDLINE, and the Cochrane Library, a thorough evaluation of the literature was conducted for all publications released up through January 23, 2023. To find any more pertinent papers, a manual reference search was also done on the listed articles. Included were studies using machine learning (ML) models to diagnose hip fractures from radiographs of the hip or pelvis or to forecast any postoperative patient outcome following hip fracture surgery.

The key findings of this study were:

A total of 39 papers that satisfied all requirements and were included in the study employed AI models in 18 (46.2%) of them to identify hip fractures on plain radiographs and in 21 (53.8%) of them to forecast patient outcomes after hip fracture surgery.

For training, validating, and testing ML models specific to diagnostic and postoperative outcome prediction, respectively, a total of 39 598 plain radiographs and 714 939 hip fractures were employed.

The most likely results were mortality and length of hospital stay.

In comparison to doctors, the OR for diagnostic mistake of ML models for hip fracture radiographs on pooled data analysis was 0.79.

The average (SD) sensitivity, specificity, and F1 score for the ML models were 89.3% (8.5%), 87.5% (9.9%), and 0.90, respectively (0.06).

In conclusion, when it came to identifying hip fractures, artificial intelligence performed on par with experienced radiologists and surgeons. The existing applications of AI for outcome prediction, however, do not appear to offer a significant advantage over multivariable predictive statistics.

Reference:

Lex, J. R., Di Michele, J., Koucheki, R., Pincus, D., Whyne, C., & Ravi, B. (2023). Artificial Intelligence for Hip Fracture Detection and Outcome Prediction. In JAMA Network Open (Vol. 6, Issue 3, p. e233391). American Medical Association (AMA). https://doi.org/10.1001/jamanetworkopen.2023.3391