Recent Study Suggests AI Could Improve Surgeon Diagnostic Accuracy

Sarasota, FL (WorkersCompensation.com) – With comments from the likes of Elon Musk, and recent reports of a suicide stemming from encouragement to do so by an AI chatbot, AI has been a hot topic with wildly varying opinions, especially when it comes to healthcare.

According to statistics from the AMA, in 2022 surveys showed that physicians were warming up to use of more technology. At the same time, there were other reports that those in the healthcare industry were feeling overwhelmed by the technological advances being used to address staffing shortages.
Researchers are also testing the AI waters when it comes to healthcare’s most important job of all, treating patients. In fact, the results of one 2021 study suggested that AI was accurate at detecting mild cognitive decline preceding the actual diagnosis.

One area that needs more research is Orthopedics, as around 10 percent of suspected hip fractures are not diagnosed in the initial pelvic x-rays. Additionally, surgical delays over 24 hours have been associated with a 20 percent increased risk of 30-day mortality, as well as twice the risk of complications, longer hospital stays, and increased costs.

Canadian researchers from the Division of Orthopaedic Surgery at the University of Toronto questioned how accurately current AI algorithms could diagnose a fracture and predict post-operative outcomes. The researchers reviewed 39 studies of which 46.2 percent used AI models to diagnose hip fractures on plain x-rays, and 53.8 percent used AI to predict outcomes following hip fracture surgery. Overall, the study utilized a total of 39,598 x-rays, and 714,939 hip fractures. Complications, length of hospital stay, and mortality were noted, as well as accuracy of diagnosis.

Compared to clinicians, the odds ratio for diagnostic error of the hip x-rays of the AI models was .79. The sensitivity rate was 89.3 percent, and a specificity rate of 87.5 percent. The mean area under the curve for mortality prediction was 0.84 with AI models, compared with 0.79 under the alternative controls. For 30-day mortality, the average accuracy of the AI models was 72.8 percent.

The performance of the AI models was comparable to the radiologists and orthopedic surgeons. However, the AI models consistently outperformed trainees and non-specialized clinicians. Additionally, the researchers found consistent improvement in the diagnosis accuracy of specialists with the help of AI.

The researchers found a higher overall accuracy in the AI models, however there was more inconsistency in the sensitivity and specificity across the AI models in comparison to clinician performance. The researchers speculate that the results could be s negatively skewed by one study that attempted to classify fractures into 3 different categories despite a relatively small sample.
The researchers noted one study that accomplished 86.0 percent sensitivity and 90.0 percent specificity rate with AI algorithms. Although the algorithm was able to detect pelvic imaging position, as well as the presence of hardware, and pelvic and acetabular fractures, it was most accurate at diagnosing proximal femoral fractures. The researchers believe that in order to reduce errors, the training of the algorithm should not only include a large sampling, but also a large number of outliers.
Overall, the researchers believe the potential applications of AI are promising, however they concede that more standardized studies need to be done to verify the usefulness of AI. They also believe that as a precaution surgeon review should be included in all steps of algorithm training, and that it ultimately comes down to the specific steps of training.

Share This Article: