Establishment and comparison of 3 fear of progression risk prediction models for gynecological malignancies patients based on machine learning

Ao Xiong¹Jiayi Wang²Zenan Wang²Dongyan Qi¹Yingqin Yu¹Lei Xia¹Junxiang Gao^1*

• ¹Second Hospital of Hebei Medical University, Shijiazhuang, Hebei Province, China
• ²Graduate School of Hebei Medical University, Shijiazhuang, Hebei Province, China

Objective: This study applied the Society Ecosystems Theory to investigate Fear of Progression (FoP) prevalence and predictors in gynecological malignancy patients. By constructing and comparing three machine learning models, we sought to identify the optimal scientifically validated predictive tool for FoP risk in clinical practice, thereby enabling early identification of high-risk populations and informing evidence-based targeted interventions. Methods: A convenience sample of 330 patients diagnosed with gynecological malignancies was recruited from a tertiary hospital in China between September 2023 and August 2024. Data were collected through validated instruments: the General Information Questionnaire, Fear of Progression Questionnaire-Short Form , Comprehensive Scores for Financial Toxicity , Chinese Dyadic Coping Inventory, Perceived Social Support Scale, and Chinese Memorial Symptom Assessment Scale. The dataset was partitioned into training (70%, n=231) and testing sets (30%, n=99) using stratified random sampling. Patients were classified into FoP and non-FoP groups based on diagnostic criteria. Three machine learning algorithms, logistic regression (LR), support vector machine (SVM), and random forest (RF) were implemented to develop FoP prediction models. Model performance was compared using accuracy, recall, precision, F1-score, and area under the ROC curve (AUC-ROC) to select the optimal model. Results: This study included 330 patients with gynecological malignancies, with a FoP incidence of 52.7% (n = 174). All three models identified social support, dyadic coping, mindset bias, and elevated tumor markers as significant predictors of FoP (P < 0.05). Additionally, symptom distress and financial toxicity demonstrated significant predictive value in the SVM and RF models. Comparative analysis revealed that the RF model outperformed the LR and SVM models in overall predictive performance. Conclusions: The Random Forest-based prediction model exhibited optimal performance, demonstrating high accuracy and reliability in identifying FoP risk among gynecological malignancy patients. It can provide a scientific foundation for early FoP detection and personalized intervention strategies. These findings underscore the clinical utility of combining machine learning approaches with social-ecological theory to advance precision nursing practices in psycho-oncology care.