Image: Neural networks trained to recognize hip joints and classify fractures (Photo courtesy of University of Bath)
A new machine learning process designed to identify and classify hip fractures has been shown to outperform human clinicians.
Two convolutional neural networks (CNNs) developed at the University of Bath (Somerset, UK) were able to identify and classify hip fractures from X-rays with a 19% greater degree of accuracy and confidence than hospital-based clinicians. The research team set about creating the new process to help clinicians make hip fracture care more efficient and to support better patient outcomes. They used a total of 3,659 hip X-rays, classified by at least two experts, to train and test the neural networks, which achieved an overall accuracy of 92%, and 19% greater accuracy than hospital-based clinicians.
Hip fractures are a major cause of morbidity and mortality in the elderly, incurring high costs to health and social care. Classifying a fracture prior to surgery is crucial to help surgeons select the right interventions to treat the fracture and restore mobility and improve patient outcomes. The ability to swiftly, accurately, and reliably classify a fracture is key: delays to surgery of more than 48 hours can increase the risk of adverse outcomes and mortality. Fractures are divided into three classes – intracapsular, trochanteric, or subtrochanteric – depending on the part of the joint they occur in. Some treatments, which are determined by the fracture classification, can cost up to 4.5 times as much as others.
As important are longer-term patient outcomes: people who sustain a hip fracture have in the following year twice the age-specific mortality of the general population. So, the team says, the development of strategies to improve hip fracture management and their impact of morbidity, mortality and healthcare provision costs is a high priority. One critical issue affecting the use of diagnostic imaging is the mismatch between demand and resource. Rising demand on radiology departments often means they cannot report results in a timely manner.
“Machine learning methods and neural networks offer a new and powerful approach to automate diagnostics and outcome prediction, so this new technique we’ve shared has great potential,” said Prof Richie Gill, lead author of the paper and Co-Director of the Center for Therapeutic Innovation, says. “Despite fracture classification so strongly determining surgical treatment and hence patient outcomes, there is currently no standardized process as to who determines this classification in the UK – whether this is done by orthopedic surgeons or radiologists specializing in musculoskeletal disorders.”
University of Bath