Sivajohan, B., Elgendi, M., Menon, C., Allaire, C., Yong, P., & Bedaiwy, M. A. (2022). Clinical use of artificial intelligence in endometriosis: a scoping review. NPJ digital medicine. https://doi.org/10.1038/s41746-022-00638-1
Sivajohan B, Elgendi M, Menon C, Allaire C, Yong P, Bedaiwy MA. Clinical use of artificial intelligence in endometriosis: a scoping review. NPJ digital medicine. 2022. doi:10.1038/s41746-022-00638-1
Sivajohan, Brintha, et al. "Clinical use of artificial intelligence in endometriosis: a scoping review." NPJ digital medicine, 2022.
For EndNote, Zotero or Mendeley:
Open access
No open license is recorded for this paper. Reuse terms are set by the publisher.
AI models show good accuracy in simple endometriosis classification
A 2022 scoping review of 36 studies found that AI models separated endometriosis cases from controls with good accuracy in simple comparisons. The models used different data types, such as biomarkers, symptoms, genes and imaging. The authors report that none of the studies involved a human comparator.
Key Findings
The review identified 1309 titles across four databases, screened 115 full texts, and included 36 studies: 20 retrospective and 16 prospective, with no randomized studies.
Prediction models made up 44.4% (n = 16) of the studies, diagnostic models 47.2% (n = 17), and studies of disease understanding 8.33% (n = 3).
Sensitivity pooled within input categories came to between 81.7 and 96.7%, and specificity pooled within input categories to between 70.7 and 91.6%.
Pooled sensitivity and specificity were 81.7% and 91.6% for clinical variables and symptoms, and 96.7% and 70.7% for genetic variables.
Logistic regression was the method used most often (n = 15). The average sample size was 245 individuals for studies exploring diagnosis and prediction.
Interpretation
This scoping review maps the published research. The authors took a sample-size-weighted average of each study's best model. A meta-analysis (a formal statistical combining of results) was not possible because the studies varied too much. Twenty of the 36 studies were retrospective, and none was randomized. Many gave little detail on how endometriosis was confirmed. The studies differed in inputs, endometriosis type and stage, and AI method, which makes the models difficult to compare. The authors note that without a human comparator, AI's advantage over clinicians is difficult to judge. The authors recommend prospective controlled trials. The search covered English-language papers through March 2022.
RRM Context
The authors name laparoscopic visualization with histologic confirmation as the gold standard for diagnosis. The paper cites diagnostic delays of up to 6 to 12 years. Restorative reproductive medicine treats pelvic pain as a question for evaluation before treatment. Cycle charting records when pain and bleeding occur. That timing adds to the symptom history, which several of these models used as input.
Our editorial summary of this paper, not the article's abstract.
Abstract
Endometriosis is a chronic, debilitating, gynecologic condition with a non-specific clinical presentation. Globally, patients can experience diagnostic delays of ~6 to 12 years, which significantly hinders adequate management and places a significant financial burden on patients and the healthcare system. Through artificial intelligence (AI), it is possible to create models that can extract data patterns to act as inputs for developing interventions with predictive and diagnostic accuracies that are superior to conventional methods and current tools used in standards of care. This literature review explored the use of AI methods to address different clinical problems in endometriosis. Approximately 1309 unique records were found across four databases; among those, 36 studies met the inclusion criteria. Studies were eligible if they involved an AI approach or model to explore endometriosis pathology, diagnostics, prediction, or management and if they reported evaluation metrics (sensitivity and specificity) after validating their models. Only articles accessible in English were included in this review. Logistic regression was the most popular machine learning method, followed by decision tree algorithms, random forest, and support vector machines. Approximately 44.4% (n = 16) of the studies analyzed the predictive capabilities of AI approaches in patients with endometriosis, while 47.2% (n = 17) explored diagnostic capabilities, and 8.33% (n = 3) used AI to improve disease understanding. Models were built using different data types, including biomarkers, clinical variables, metabolite spectra, genetic variables, imaging data, mixed methods, and lesion characteristics. Regardless of the AI-based endometriosis application (either diagnostic or predictive), pooled sensitivities ranged from 81.7 to 96.7%, and pooled specificities ranged between 70.7 and 91.6%. Overall, AI models displayed good diagnostic and predictive capacity in detecting endometriosis using simple classification scenarios (i.e., differentiating between cases and controls), showing promising directions for AI in assessing endometriosis in the near future. This timely review highlighted an emerging area of interest in endometriosis and AI. It also provided recommendations for future research in this field to improve the reproducibility of results and comparability between models, and further test the capacity of these models to enhance diagnosis, prediction, and management in endometriosis patients.
