Ovarian torsion: developing a machine-learned algorithm for diagnosis

Ovarian torsion is a common concern in girls presenting to emergency care with pelvic or abdominal pain. The diagnosis is challenging to make accurately and quickly, relying on a combination of physical exam, history and radiologic evaluation. Failure to establish the diagnosis in a timely fashion can result in irreversible ovarian ischemia with implications for future fertility. Ultrasound is the mainstay of evaluation for ovarian torsion in the pediatric population. However, even with a high index of suspicion, imaging features are not pathognomonic.

We sought to develop an algorithm to aid radiologists in diagnosing ovarian torsion using machine learning from sonographic features and to evaluate the frequency of each sonographic element.

All pediatric patients treated for ovarian torsion at a quaternary pediatric hospital over an 11-year period were identified by both an internal radiology database and hospital-based International Statistical Classification of Diseases and Related Health Problems (ICD) code review. Inclusion criteria were surgical confirmation of ovarian torsion and available imaging. Patients were excluded if the diagnosis could not be confirmed, no imaging was available for review, the ovary was not identified by imaging, or torsion involved other adnexal structures but spared the ovary. Data collection included: patient age; laterality of torsion; bilateral ovarian volumes; torsed ovarian position, i.e. whether medialized with respect to the mid-uterine line; presence or absence of Doppler signal within the torsed ovary; visualization of peripheral follicles; and presence of a mass or cyst, and free peritoneal fluid. Subsequently, we evaluated a non-torsed control cohort from April 2015 to May 2016. This cohort consisted of sequential girls and young adults presenting to the emergency department with abdominopelvic symptoms concerning for ovarian torsion but who were ultimately diagnosed otherwise. These features were then fed into supervised machine learning systems to identify and develop viable decision algorithms. We divided data into training and validation sets and assessed algorithm performance using sub-sets of the validation set.

We identified 119 torsion-confirmed cases and 331 torsion-absent cases. Of the torsion-confirmed cases, significant imaging differences were evident for girls younger than 1 year; these girls were then excluded from analysis, and 99 pediatric patients older than 1 year were included in our study. Among these 99, all variables demonstrated statistically significant differences between the torsion-confirmed and torsion-absent groups with P-values <0.005. Using any single variable to identify torsion provided only modest detection performance, with areas under the curve (AUC) for medialization, peripheral follicles, and absence of Doppler flow of 0.76±0.16, 0.66±0.14 and 0.82±0.14, respectively. The best decision tree using a combination of variables yielded an AUC of 0.96±0.07 and required knowledge of the presence of intra-ovarian flow, peripheral follicles, the volume of both ovaries, and the presence of cysts or masses.

Based on the largest series of pediatric ovarian torsion in the literature to date, we quantified sonographic features and used machine learning to create an algorithm to identify the presence of ovarian torsion - an algorithm that performs better than simple approaches relying on single features. Although complex combinations using multiple-interaction models provide slightly better performance, a clinically pragmatic decision tree can be employed to detect torsion, providing sensitivity levels of 95±14% and specificity of 92±2%.