Predicting Response to Treatment and Survival in Advanced Ovarian Cancer Using Machine Learning and Radiomics: A Systematic Review

Sabrina Piedimonte, Mariam Mohamed, Gabriela Rosa, Brigit Gerstl, Danielle Vicus

Background and Objective: Machine learning and radiomics (ML/RM) are gaining interest in ovarian cancer (OC) but only a few studies have used these methods to predict treatment response. The objective of this study was to review the literature on the applications of ML/RM in OC assessments, specifically focusing on studies describing algorithms to predict treatment response and survival. Methods: This is a systematic review of the published literature from January 1985 to December 2023 on the use of ML/RM in OC An extensive search of electronic library databases was conducted. Two independent reviewers screened the articles initially by title then by full text. Quality was assessed using the MINORS criteria. p-values were generated using the Pearson’s Chi-squared (x2) test to compare the performances of ML/RM models with traditional statistics. Results: Of the 5576 screened articles, 225 studies were included. Between 2021 and 2023, 49 studies were published, highlighting the rapidly growing interest in ML/RM. Median-quality scores using the MINORS scale were similar between studies published between 1985–2021 and 2021–2023 (both 8). Neural Networks (22.6%) and LASSO (15.3%) were the most common ML/RM algorithms in OC. Among these studies, 13 focused specifically on prediction of treatment response using radiomics. A total of 5113 patients were analyzed. The most common algorithms were Random Forest (4/13) followed by Neural Networks (3/13) and Support Vectors (3/13). Radiomic analysis was used to predict response to neoadjuvant chemotherapy in seven studies, with a median AUC of 0.77 (range 0.72–0.93), while the median AUC was 0.82 (range 0.77–0.89) in the six studies assessing the prediction of optimal or complete cytoreduction. Median model accuracy reported in 7/13 studies was 73% (range 66–98%). Additionally, four studies investigated the use of ML/RM for survival prediction for OC. The XGBoost model had 80.9% accuracy in predicting 5-year survival compared to linear regression, which achieved 79% accuracy. The Random Forest model has 93.7% accuracy in predicting 12-month progression-free survival, compared to 82% for linear regression. Conclusions: In conclusion, we found that the use of ML/RM algorithms is becoming a more frequent method to predict responses to treatment of OC. These models should be validated in a prospective multicenter trial prior to integration into clinical use.