Ultrasonographic scoring system for the diagnosis of adnexal torsion

Can Natural Language Processing Improve Adnexal Torsion Predictions?

STUDY OBJECTIVE

To create a decision support tool based on machine learning algorithms and natural language processing (NLP) technology, to augment clinicians' ability to predict cases of suspected adnexal torsion.

DESIGN

Retrospective cohort study SETTING: Gynecology department, university-affiliated teaching medical center, 2014-2022.

PATIENTS

This study assessed risk-factors for adnexal torsion among women managed surgically for suspected adnexal torsion based on clinical and sonographic data.

INTERVENTIONS

None.

MEASUREMENTS AND MAIN RESULTS

The dataset included demographic, clinical, sonographic, and surgical information obtained from electronic medical records. NLP was used to extract insights from unstructured free text and unlock them for automated reasoning. The machine learning model was a CatBoost classifier that utilizes gradient boosting on decision trees. The study cohort included 433 women who met inclusion criteria and underwent laparoscopy. Among them, 320 (74%) had adnexal torsion diagnosed during laparoscopy, and 113 (26%) did not. The model developed improved prediction of adnexal torsion to 84%, with a recall of 95%. The model ranked several parameters as important for prediction. Age, difference in size between ovaries, and the size of each ovary were the most significant. The precision for the "no torsion" class was 77%, with a recall of 45%.

CONCLUSIONS

Using machine learning algorithms and NLP technology as a decision-support tool for the diagnosis of adnexal torsion is feasible. It improved true prediction of adnexal torsion to 84% and decreased cases of unnecessary laparoscopy.

A Scoring System Developed by a Machine Learning Algorithm to Better Predict Adnexal Torsion

STUDY OBJECTIVE

To establish a clinically relevant prediction score for the diagnosis of adnexal torsion (AT) in women who were operated on for suspected AT.

DESIGN

A retrospective cohort study conducted between 2014 and 2021.

SETTING

A large tertiary teaching medical center.

PATIENTS

Women who underwent urgent laparoscopy for suspected AT.

INTERVENTIONS

Analyses included univariate and multivariate models combined with the machine learning (ML) Random Forest model, which included all information available about the women and reported the accuracy of the model and the importance of each variable. Based on this model, we created a predictive score and evaluated its accuracy by receiver operating characteristic (ROC) curve.

MEASUREMENTS AND MAIN RESULTS

A total of 503 women were included in our study, 244 (49%) of whom were diagnosed with AT during the surgery, and 44 (8.8%) cases of necrotic ovary were found. Based on the Random Forrest and multivariate models, the most important preoperative clinical predictive variables for AT were vomiting, left-side complaints, and concurrent pregnancy; cervical tenderness and urinary symptoms decreased the likelihood of surgically confirmed AT. The most important sonographic findings that predicted increased risk of surgically confirmed AT were ovarian edema and decreased vascular flow; in contrast, hemorrhagic corpus luteum decreased the likelihood of surgically confirmed AT. The accuracy of the Random Forest model was 71% for the training set and 68% for the testing set, and the area under the curve for the multivariate model was 0.75 (95% confidence interval [CI] 0.69-0.80). Based on these models, we created a predictive score with a total score that ranges from 4 to 12. The area under the curve for this score was 0.72 (95% CI 0.67-0.76), and the best cutoff for the final score was >5, with a sensitivity, specificity, positive predictive value, and negative predictive value of 64%, 73%, 70%, and 67%, respectively.

CONCLUSION

Clinical characteristics and ultrasound findings may be incorporated into the emergency room workup of women with suspected AT. ML in this setting has no diagnostic/predictive advantage over the performance of logistic regression methods. Additional prospective studies are needed to confirm the accuracy of this model.