A predictive model for recurrence in patients with borderline ovarian tumor based on neural multi-task logistic regression.

BACKGROUND

Effective management of patients with borderline ovarian tumor (BOT) requires the timely identification of those at a higher risk of recurrence. Artificial neural networks have been successfully used in many areas of clinical event prediction, significantly affecting clinical decisions and practice.

OBJECTIVE

We developed and validated a novel clinical model based on neural multi-task logistic regression (N-MTLR) for predicting recurrence in patients with BOT who underwent initial surgeries, and compared its prediction performance with that of the Cox regression model.

METHODS

This retrospective study included 736 patients diagnosed with BOT from May 2011 to August 2022, with 84 recurrences. The synthetic minority oversampling technique (SMOTE) was used to balance the minority group such that the two patient types were 1:1. Using random sampling, the SMOTE-balanced dataset was divided into 80% of the sample (1043 patients) as the training set and 20% (261 patients) as the validation set. Both N-MTLR and Cox regression models were trained on the training set using SMOTE and evaluated on the validation set using the time-dependent area under the receiver operating characteristic curve (tdAUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy.

RESULTS

Among the 736 enrolled patients, only 84 (11.41%) were diagnosed with BOT recurrence. Using SMOTE, the balanced dataset (1304 patients) contained equal numbers of patients (652 patients) in both recurrence and non-recurrence groups. Multivariate Cox regression analysis of the training set revealed that independent risk factors for BOT recurrence were premenopause, laparoscopic surgery, tumor rupture, advanced clinical stage, undissected lymph nodes, bilateral tumors, and fertility-sparing surgery (FSS). The N-MTLR model was constructed by correlation screening of 34 features in the training set, and 10 variables were screened including FSS, completeness of surgery, comorbidities, International Federation of Gynecology and Obstetrics (FIGO) staging, age, omentectomy, lymphadenectomy, parity, menopausal status, and peritoneal implantation. The N-MTLR model outperformed the Cox regression model in terms of AUC, accuracy, specificity, PPV, and NPV at the quartiles of follow-ups (2, 4, and 7 years).

CONCLUSIONS

The N-MTLR model effectively predicts BOT recurrence. Identifying high-risk recurrence groups in patients with BOT can facilitate close monitoring, suitable treatment, and an opportune time for intervention.