Adnexal torsion refers to partial or complete rotation of the ovary and/or tube on the infundibulo-pelvic and utero-ovarian ligaments that support the adnexa. This causes partial or complete obstruction of the blood supply, which in turn can cause ischemia, necrosis, and loss of ovarian or tube function. Therefore, this condition represents an emergent gynecologic condition [1].

The incidence of adnexal torsion is based on estimates because the definitive diagnosis is made during surgery and in some patients, may be misdiagnosed preoperatively. The prevalence is approximately 2% to 6% [2],[3]. It usually occurs in women of reproductive age; although it can also be found in adolescents and pregnant patients [4].

The period in which a twisted ovary becomes necrotic or irreversibly damaged is unknown and challenging to study. Recognizing the onset of symptoms that might lead to ovarian necrosis or permanent damage, as well as early diagnosis and prompt treatment are crucial to preserving ovarian and/or tubal function and to prevent other associated morbidities [5]. The diagnosis of adnexal torsion is based on clinical and sonographic evaluation. The classic clinical presentation is acute onset of localized, lower abdominal pain that can vary in intensity and severity but usually occurs in 90–100% of patients [6]. Additional symptoms can be nausea in 70% and vomiting 50%; vomiting is an independent predictor [7]. Physical examination can strengthen the suspected diagnosis, even though it is nonspecific and varies significantly among patients [4]. Almost all patients have pain on physical examination [8]. Even though the final decision regarding surgical intervention is based on the clinical presentation, ultrasound imaging significantly impacts the decision.

Based on the literature, the common sonographic signs of torsion include unilaterally enlarged adnexa, whirlpool sign (79-90%), stromal edema with or without peripherally displaced antral follicles (70-80%), and fluid in the pelvis (70%) [9]. Absence of Doppler flow can help establish the diagnosis. However, normal Doppler flow does not necessarily exclude ovarian torsion because preserved flow can be attributed to incomplete occlusion, intermittent torsion, or collateral blood supply [10]. An ovarian cyst or mass, enlarged ovary (following hyperstimulation or polycystic ovary) or pregnancy are among the risk factors for adnexal torsion [1].

The treatment for adnexal torsion is an emergent surgery to confirm the diagnosis and assess ovarian viability, and detorsion of the adnexa with or without cystectomy. The decision regarding salpingectomy or oophorectomy is based on the patient's age, family planning and surgical findings [11,12]. Laparoscopy is generally safe but usually requires general anesthesia. Complications of any surgery may include infection, bleeding, and injury to adjacent organs [13,14].

Despite many years of research, the accuracy of the preoperative diagnosis is suboptimal. Data show that the preoperative diagnosis of adnexal torsion is confirmed in only 46% to 74% of cases [15,16].

The accuracy of ultrasonography in diagnosing adnexal torsion remains uncertain, with studies reporting correct diagnoses in only up to 66% of cases. One study reported a positive predictive value of 82% of the ultrasound diagnosis of adnexal torsion, based on a relatively small sample size of 129 women [17].

As the current tools for diagnosing adnexal torsion are limited and due to the importance of a more accurate diagnosis to avoid unnecessary surgery and possible complications, scoring systems and questionnaires have been created [18,19]. None of the existing tools have been shown to perform satisfactorily; therefore, they have not been incorporated into clinical practice.

To address this challenge, natural language processing (NLP) methods and tools suitable for unlocking insights from unstructured clinical notes have been leveraged. Specifically, the current study employed Azure Text Analytics for health [20]. This is a cloud-based application programming interface (API) service, that applies machine-learning intelligence for surfacing and labelling relevant medical information from a variety of unstructured, multilingual texts. Structured patient record fields enriched with insights extracted from free-text admission notes and imaging narratives, served as a basis for automated reasoning and developing a predictive model for ovarian torsion.

This study used machine learning algorithms and NLP technology to create a decision support tool that would augment clinicians’ abilities to predict cases of suspected ovarian torsion. Leveraging NLP technology allowed clinical parameters that would otherwise be hidden within unstructured text to be unlocked.