Abstract
242237
Introduction: The FIGO/TNM staging system, with emphasis on lymph node involvement, is vital in cervical cancer prognosis. Identifying lymph node status pre-treatment aids personalized planning, reducing unnecessary surgeries or extensive radiotherapy. Our study explores disease-free survival in patients initially diagnosed with lymph node involvement but exhibiting negative post-treatment PET results.
Methods: Our study considered a cohort of 64 patients diagnosed with cervical cancer with the average age of 47.9±14.5. The patient selection spanned from 2015 to 2020, and included both pre-treatment PET scans and a post-treatment PET scan, conducted at a median of 5.54 months after radiotherapy. All patients had confirmed histological diagnosis of cervical cancer, with 61.9% showing lymph node involvement in their pre-treatment PET (Pre-PET) and 31.7% (n=20) experiencing recurrence following curative-intent radiotherapy. Notably, in this study we focused on the subset of cases, constituting 73.0% (n=46), characterized by negative post-treatment PET results. Disease-free survival after radiotherapy was determined based on follow-up imaging selected as per clinical practice. Our analysis focused on several key features extracted from Pre-PET scans, including lymph node involvement, the total metabolic tumor volume (TMTV) of the primary tumor and lymph nodes, and the maximum standardized uptake value (SUVmax) of the primary tumor, and the number of involved lymph nodes. We employed a robust approach to hyperparameter tuning using grid search, and assessed the predictive performance of machine learning algorithms (Gradient Boosting, Ridge, eXtreme Gradient Boosting (XG Boost), Linear Regression, Random Forest regressor, Lasso, K nearest Neighbors regressor, and Support Vector Regression (SVR)) for disease-free survival analysis through nested cross-validation. The hyperparameter tuning process was facilitated by GridSearchCV using negative mean squared error as the scoring metric. The outer loop iterates through folds, utilizing the best models obtained from grid search for each regression algorithm. The concordance index (c-index) is employed as an evaluation metric for disease-free survival prediction. Additionally, the Mann-Whitney U test is conducted to assess significant differences in performance between pairs of regressors.
Results: Within the subgroup of cases with negative post-treatment PET results, 21.7% (n=10) encountered recurrence. Kaplan-Meier plots and log-rank test results revealed a significant difference in clinical disease-free survival of patients with lymph node (LN) involvement, for the subgroup with negative post-treatment PET ('LN Involvement & PostPET Negative') compared to those with positive post-treatment PET ('LN Involvement & PostPET Positive'), with a p-value of 0.033. On the other hand, we did not detect statistically significant difference in disease-free survival of patients without LN involvement, for the subgroups with negative vs. positive post-treatment PET results ('No LN Involvement & PostPET Negative' and 'No LN Involvement & PostPET Positive'), with a p-value of 0.58. Gradient Boosting, Ridge and XGBoost outperformed other approaches with c-indices of 0.671±0.120, 0.669±0.246 and 0.659±0.204 respectively (p-value<0.05) for progression prediction in cases with LN involvement and negative post-PET.
Conclusions: Our study highlighted the crucial interplay of lymph node involvement in influencing clinical disease-free survival. Notably, Gradient Boosting, Ridge, and XGBoost machine learning demonstrated superior predictive performance, underscoring their potential as effective models for prognosis in this specific patient population. These findings offer valuable insights to enhance outcomes in the management of cervical cancer, particularly for patients with node involvement in pre-treatment PET scans whose post-treatment PET results are negative.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.