Deep learning predicts hip fracture using confounding patient and healthcare variables

• DOI: 10.1038/s41746-019-0105-1
• Publication Date: 2019-04-30T10:04:20Z
• Authors: Marcus A. Badgeley; John R. Zech; Luke Oakden-Rayner; Benjamin S. Glick...; Manway Liu; William Gale; Michael V. McConnell; Bethany Percha; Thomas M. Snyder; Joel T. Dudley
• Keywords: Hip Fracture

Hip fractures are a leading cause of death and disability among older adults. also the most commonly missed diagnosis on pelvic radiographs, delayed leads to higher cost worse outcomes. Computer-aided (CAD) algorithms have shown promise for helping radiologists detect fractures, but image features underpinning their predictions notoriously difficult understand. In this study, we trained deep-learning models 17,587 radiographs classify fracture, 5 patient traits, 14 hospital process variables. All 20 variables could be individually predicted from radiograph, with best performances scanner model (AUC = 1.00), brand 0.98), whether order was marked "priority" 0.79). Fracture moderately well 0.78) better when combining data 0.86, DeLong paired AUC comparison, p 2e-9) or plus 0.91, 1e-21). prediction test set that balanced fracture risk across significantly lower than random 0.67, unpaired 0.003); variables, performed randomly 0.52, 95% CI 0.46-0.58), indicating these were main source model's predictions. A single directly combines features, patient, outperforms Naive Bayes ensemble an image-only prediction, data. If CAD inexplicably leveraging in predictions, it is unclear how should interpret context other known Further research needed illuminate decision processes so computers clinicians can effectively cooperate.