A novel machine learning model may help reduce negative appendectomies in pediatric patients with a high clinical probability of acute appendicitis, according to a recent study published in Scientific Reports.
Diagnosing and referring patients for appendectomy based on clinical presentation, standard laboratory findings, and radiologic imaging, such as ultrasound, results in negative appendectomies in about 10% to 15% of cases. Although rare, complications may still occur during or after appendectomy.
The appendicitis inflammatory response score, which has shown high specificity and positive predictive values, was introduced to aid physicians in properly diagnosing acute appendicitis in this patient population. However, even with the use of this score, the rate of negative appendectomies remains high.
In the recent study, researchers used clinical, anthropometric, and laboratory data from 551 pediatric patients aged 0 to 17 years who underwent surgery for suspected acute appendicitis between January 2019 and July 2023 to develop, train, and validate a machine learning model. Among the patients involved in the study, they identified 252 patients with uncomplicated appendicitis, 252 patients with complicated appendicitis, and 47 patients with no appendicitis.
The researchers trained three machine learning algorithms—random forest, eXtreme gradient boosting, and logistic regression—to improve the detection of negative appendicitis without reducing the detection of positive cases by tuning the model’s hyperparameters. During threshold shifting, they incorporated a custom metric to locate the highest specificity on the receiver operator characteristic curve where sensitivity was still 1.
After testing the three algorithms, the researchers found that the random forest model achieved a mean specificity of 0.17 ± 0.01 and a mean sensitivity of 0.997 ± 0.001 when predicting acute appendicitis. Additionally, it achieved a joint mean specificity of 0.129 ± 0.009 and a mean sensitivity of 0.994 ± 0.002 when predicting discrimination between complicated appendicitis and either uncomplicated acute appendicitis or no appendicitis. Further, the random forest model demonstrated greater diagnostic accuracy compared with the appendicitis inflammatory response score.
The researchers suggested that although the machine learning model missed about 0.3% of cases where surgery was required, the use of the novel model could spare up to 17% of pediatric patients with a high clinical probability of acute appendicitis from undergoing unnecessary appendectomies. The researchers emphasized that their model for predicting complicated appendicitis may be used to determine which patients with complicated cases should undergo immediate appendectomy. They indicated that external validation studies may be needed to confirm their findings.
The study authors concluded: “These findings suggest the potential use of [machine learning] models in assisting clinicians in making accurate decisions [and] addressing the needs of centers that endorse performing emergency surgery only in patients with complicated appendicitis. In these centers, implementing such a model could be beneficial for appendicitis management.”
A full list of disclosures can be found in the original study.