By analyzing lab values and ultrasound data, an artificial intelligence (AI) algorithm can be highly accurate for diagnosing acute appendicitis and could potentially help avoid unnecessary surgery in two-thirds of patients without appendicitis, according to research published online September 25 in PLOS One.
A team of researchers led by Josephine Reissmann of Charité Universitätsmedizin Berlin trained an AI algorithm to provide an automated diagnosis of appendicitis based on the analysis of full blood counts, C-reactive protein (CRP), and appendiceal diameters on ultrasound examinations. In testing, the algorithm was 90% accurate for diagnosing appendicitis.
"The presented method has the potential to change today's therapeutic approach for appendicitis and demonstrates the capability of algorithms from AI and [machine learning] to significantly improve diagnostics even based on routine diagnostic parameters," the authors wrote.
Acute appendicitis represents one of the major causes for emergency surgery but remains a challenging diagnosis. As a result, the researchers set out to establish a decision-making model for suspected acute appendicitis in children based on reliable nonclinical parameters that are unbiased from interpretation or expert opinion. They also focused on differentiation between uncomplicated (phlegmonous) and complicated (gangrenous/perforated) appendicitis.
"Early diagnosis of complicated inflammation is particularly important, because this severe type of disease primarily requires surgical treatment," they wrote. "In contrast, for uncomplicated appendicitis conservative strategies are under investigation and will most probably be primarily applied in the near future, as shown by a current multicenter randomized controlled trial."
Reissmann and colleagues first gathered data from 590 pediatric patients who had received surgery for suspected acute appendicitis at their institution between December 2006 and September 2016. Of the 590 patients, 473 had histopathologically proven appendicitis and 117 had negative histopathological findings. The classification model was trained on 35% of the patients, with the remaining 65% used for validation. The AI model found two distinct biomarker signatures for diagnosing appendicitis and complicated appendicitis, respectively.
"For the diagnosis of appendicitis, a selective biomarker signature was developed containing basophils, leukocytes, monocytes, neutrophils, CRP, and the appendiceal diameter," they wrote. "For the differential diagnosis of complicated versus uncomplicated appendicitis, a selective biomarker signature was developed including basophils, eosinophils, monocytes, thrombocytes, [and] CRP, supplemented by the appendiceal diameter."
| Performance of AI model on pediatric patients with suspected appendicitis | |||
| Sensitivity | Specificity | Accuracy | |
| Diagnosing appendicitis | 93% | 67% | 90% |
| Identifying complicated inflammation | 95% | 33% | 51% |
If used clinically, the model would be capable of avoiding unnecessary surgery in two out of three patients without appendicitis and one out of three patients with uncomplicated appendicitis, according to the researchers.
"Due to the retrospective nature of our study we do not present a ready-to-use clinical algorithm, but our approach demonstrates significant improvements compared to today's diagnosis and enables secure translation into clinical practice," they wrote. "Our approach also demonstrates significant value in ruling out complicated appendicitis with high sensitivity. Investigations on the [omics] level such as genome-wide gene expression profiling of specific cell compartments could be a path to increase the specificity.
