Risk prediction model of survival in patients with low-grade serous ovarian cancer: a multicenter Cohort study.

A total of 155 eligible patients with confirmed LGSOC were finally enrolled in this multicenter study, and 4 (2.6%) patients have history of endometriosis. The median age at diagnosis was 47 years, ranging from 21 to 79 years. Among the 64 patients in the early FIGO stage, the majority received comprehensive staging surgery. Conversely, among the 91 patients in the advanced stage, cytoreductive surgery was predominantly performed, among whom 54 patients (59.3%) undergone optimal cytoreductive surgery. A total of 110 patients (71.0%) underwent pelvic and/or para-aortic lymphadenectomy, with 30 patients (27.3%) exhibiting pathological evidence of lymph node metastasis. With a median follow-up of 30 months, 54 patients experienced disease recurrence, and 27 patients died. Clinical features of the cohort are presented in Table 1 . Table 1 Basic clinical features of LGSOC patients Characteristic Total cohort(n=155) Age, year  ≤50 97 (62.6)  >50 58 (37.4) FIGO (2014)  I and II 64 (41.3)  III and IV 91 (58.7) CA-125, U/mL  ≤35 29 (18.7)  >35 126 (81.3) Rupture of tumor  None 36 (23.2)  Pre-operative 103 (66.5)  Intro-operative 16 (10.3) Cytoreductive surgery  Optimal (<1cm) 114 (73.5)  Suboptimal (≥1cm) 41 (26.5) Ascites cytology  Negative 47 (30.3)  Positive 27 (17.4)  Unknown 81 (52.3) Lymphadenectomy  No 47(30.3)  Yes 108(69.7) Adjuvant chemotherapy  No 21 (13.5)  Yes 134 (86.5) Values are presented as n (%) Abbreviation: LGSOC low-grade serous ovarian cancer, FIGO International Federation of Gynecology and Obstetrics, CA-125 carbohydrate antigen 125 Basic clinical features of LGSOC patients Values are presented as n (%) Abbreviation: LGSOC low-grade serous ovarian cancer, FIGO International Federation of Gynecology and Obstetrics, CA-125 carbohydrate antigen 125 All available clinicopathological variables were included in univariate Cox proportional hazard regression analysis. Univariate Cox analysis revealed significant associations of age at diagnosis, FIGO stage, cytoreductive surgery, ascites cytology, and lymphadenectomy with both DFS and OS (all P < 0.05; Table 2 ). Meanwhile, LASSO regression with ten-fold cross-validation was used to select prognostic features from the clinical-pathological variables. Notably, the variables identified by the LASSO regression were consist with the univariable Cox analysis, which underscored the robustness of these candidate variables (Fig. S2). Subsequently, the multivariate Cox analysis revealed that FIGO stage, rupture of tumor, cytoreductive surgery, and lymphadenectomy, exhibited independent prognostic significance for DFS. Moreover, age at diagnosis, FIGO stage, cytoreductive surgery, and lymphadenectomy were identified as independent prognostic factors for OS. Table 2 Univariable Cox proportional hazard regression analysis of DFS and OS Characteristic DFS OS HR (95%CI) P value HR (95%CI) P value Age, year  ≤50 Reference Reference  >50 1.83 (1.08-3.09) 0.025 3.01 (1.40-6.49) 0.005 FIGO (2014)  I and II Reference Reference  III and IV 6.38 (3.11-13.08) ＜0.001 6.48 (2.25-18.67) 0.001 CA-125, U/mL  ≤35 Reference Reference  >35 3.67 (1.33-10.16) 0.012 3.56 (0.85-14.99) 0.083 Rupture of tumor  None Reference - -  Pre-operative 7.00 (2.18-22.52) 0.001 -  Intro-operative 4.83 (1.21-19.32) 0.026 - Cytoreductive surgery  Suboptimal (≥1cm) Reference Reference  Optimal (<1cm) 0.26 (0.15-0.45) ＜0.001 0.22 (0.10-0.45) ＜0.001 Ascites cytology  Negative Reference Reference  Positive 2.39 (1.03-5.55) 0.043 4.10 (1.26-13.37) 0.019  Unknown 2.38 (1.17-4.81) 0.016 2.83 (0.94-8.46) 0.063 Lymphadenectomy  No Reference Reference  Yes 0.33 (0.20-0.57) <0.001 0.22 (0.10-0.46) <0.001 Adjuvant chemotherapy  No Reference Reference  Yes 2.63 (0.82-8.45) 0.104 1.27 (0.38-4.20) 0.699 Abbreviation: DFS disease-free survival, OS overall survival, HR hazard ratio, CI confidence interval, FIGO International Federation of Gynecology and Obstetrics, CA-125  carbohydrate antigen 125 Univariable Cox proportional hazard regression analysis of DFS and OS Abbreviation: DFS disease-free survival, OS overall survival, HR hazard ratio, CI confidence interval, FIGO International Federation of Gynecology and Obstetrics, CA-125  carbohydrate antigen 125 The nomogram of DFS and OS at 3- and 5-year based on the prediction risk is shown in Fig. 1 . The findings revealed that the FIGO stage exerted the most significant influence on the DFS prediction model, followed by rupture of tumor, lymphadenectomy and cytoreductive surgery. Each of these variable types assigns a score that enables the prediction of the likelihood of an event occurring at a specific time based on the cumulative score of all factors. Fig. 1 Nomograms for predicting prognosis in LGSOC patients. A Nomogram for predicting DFS in LGSOC patients. B Nomogram for predicting OS in LGSOC patients. Abbreviation: LGSOC, low-grade serous ovarian cancer; DFS, disease-free survival; OS, overall survival Nomograms for predicting prognosis in LGSOC patients. A Nomogram for predicting DFS in LGSOC patients. B Nomogram for predicting OS in LGSOC patients. Abbreviation: LGSOC, low-grade serous ovarian cancer; DFS, disease-free survival; OS, overall survival The discriminatory ability of the model was assessed using the C-index and AUC. The C-index for DFS prediction was determined to be 0.781, while the AUC values for 3-year and 5-year DFS were found to be 0.795 and 0.817, respectively (Fig. 2 A). The C-index for OS prediction was determined to be 0.802, while the AUC values for 3- and 5-year OS were found to be 0.810 and 0.798, respectively (Fig. 2 B). The calibration curves for DFS and OS at 3- and 5-year showed close agreement between predicted and observed probabilities (Fig. 2 C-D). Subsequently, decision curve analysis indicated the positive net clinical benefit of using the model for prediction. Furthermore, comparison of the 3- and 5-year data demonstrated enhanced prediction performance over longer time horizons, as illustrated in Fig. 2 E-F. Fig. 2 ROC curves, calibration curves and DCA of the nomogram.  A , B ROC curves for predicting 3-year and 5-year DFS and OS in LGSOC patients in the total cohort. C , D Calibration curves for 3-year and 5-year DFS and OS in LGSOC patients. E , F 3-year and 5-year DFS and OS benefit in the total cohort.Abbreviation: ROC, receiver operating characteristic; AUC, area under the receiver operating characteristic curve; DCA, decision curve analysis; LGSOC, low-grade serous ovarian cancer; DFS, disease-free survival; OS, overall survival ROC curves, calibration curves and DCA of the nomogram.  A , B ROC curves for predicting 3-year and 5-year DFS and OS in LGSOC patients in the total cohort. C , D Calibration curves for 3-year and 5-year DFS and OS in LGSOC patients. E , F 3-year and 5-year DFS and OS benefit in the total cohort.Abbreviation: ROC, receiver operating characteristic; AUC, area under the receiver operating characteristic curve; DCA, decision curve analysis; LGSOC, low-grade serous ovarian cancer; DFS, disease-free survival; OS, overall survival Simultaneously, a time-dependent AUC curve is constructed to demonstrate its discriminant ability in response to temporal variations (Fig. S3). Ultimately, the evaluation of the prediction model in relation to FIGO staging reveals its superior clinical applicability (Fig. S4). Furthermore, we compared the predictive ability for 3- and 5-year between the prediction model and the FIGO stage, utilizing the NRI and IDI metrics. In this study, the NRI values of the prediction model for the 3- and 5-year DFS were 0.329 (95%CI: 0.067-0.530, P = 0.028) and 0.317 (95%CI: 0.003-0.521, P = 0.046), respectively. Furthermore, the IDI values for the 3- and 5-year DFS were 0.109 (95%CI: 0.033-0.223, P = 0.002) and 0.070 (95%CI: 0.014-0.174, P = 0.010), respectively. As for the OS, the IDI values for 3-year and 5-year were 0.108 (95%CI: 0.022-0.284, P = 0.006) and 0.112 (95%CI: 0.025-0.282, P = 0.004), respectively. The NRI values were determined as 0.333 (95% CI: -0.015-0.606, P = 0.064) and 0.283 (95% CI: 0.028-0.562, P = 0.030), respectively. The detailed outcomes are depicted in Table S1. Compared to the Cox proportional hazards regression model, the predictive performance of the deep learning-based model based on the neural networks significantly superior, with higher C-indexes for predicting DFS (0.907 vs. 0.781) and OS (0.922 vs. 0.802), respectively. Additionally, the time-dependent AUC curve demonstrates that the deep learning-based model exhibits excellent discrimination ability in response to temporal changes (Fig. 3 A-B). The AUC values for predicting 3-year and 5-year DFS were 0.935 and 0.965, respectively, while the AUC values for predicting 3-year and 5-year OS were 0.954 and 0.946, respectively (Fig. 3 C-D). Based on the risk scores output by the neural network model, patients were divided into high-risk and low-risk groups. Kaplan-Meier survival analysis revealed that patients in the low-risk group had significantly better DFS and OS compared to those in the high-risk group ( P < 0.001), underscoring the strong risk stratification ability of this model (Fig. 3 E-F). Fig. 3 Performance of the deep learning-based model for DFS and OS in LGSOC patients.  A , B Time-dependent AUC curves for DFS and OS in LGSOC patients over a 5-year period. C , D ROC curves for predicting 3-year and 5-year DFS and OS in LGSOC patients using the deep learning-based model. E , F Kaplan-Meier analysis of DFS and OS in LGSOC patients, stratified into high-risk and low-risk groups based on risk scores from the deep learning-based model. Abbreviation: DFS, disease-free survival; OS, overall survival; LGSOC, low-grade serous ovarian cancer; AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic Performance of the deep learning-based model for DFS and OS in LGSOC patients.  A , B Time-dependent AUC curves for DFS and OS in LGSOC patients over a 5-year period. C , D ROC curves for predicting 3-year and 5-year DFS and OS in LGSOC patients using the deep learning-based model. E , F Kaplan-Meier analysis of DFS and OS in LGSOC patients, stratified into high-risk and low-risk groups based on risk scores from the deep learning-based model. Abbreviation: DFS, disease-free survival; OS, overall survival; LGSOC, low-grade serous ovarian cancer; AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic

This multi-center retrospective study was conducted using data from the population-based LGSOC cohort from 2011 to 2020, comprising four university-teaching hospitals across China, namely Qilu Hospital of Shandong University, Tongji Hospital at Tongji Medical College of Huazhong University of Science and Technology, Women’s Hospital School of Medicine at Zhejiang University, and Affiliated Hospital of Qingdao University. The inclusion criteria were as follows: (a). patients diagnosed with LGSOC according to FIGO 2014 between 2011 and 2020; (b). patients underwent surgical treatment at renowned medical institutions; (c). patients with availability of complete clinicopathological and follow-up records in the database. Patient stage was performed according to the FIGO 2014, and the anatomical staging definitions of FIGO 2014 are consistent with the FIGO 2021 system. Patients who lost to follow-up after treatment were excluded. The flowchart of study flowchart is presented in Fig. S1. The study was approved by the Ethical Committee of Qilu Hospital of Shandong University (KYLL-202011-158-1) and obtained a waiver for informed consent. Before the analysis, the privacy of each patient was maintained. A clinical cohort was established to gather clinical data, encompassing variables such as age at diagnosis (≤50 years, >50 years), FIGO stage (2014) (I and II, III and IV), preoperative serum carbohydrate antigen 125 (CA-125) level (≤35, >35 U/mL), rupture of tumor (none, pre-operative, intro-operative), cytoreductive surgery (optimal, suboptimal), ascites cytology (negative, positive, unknown), lymphadenectomy (no, yes), adjuvant therapy (none, chemotherapy). The criterion for optimal cytoreductive surgery was established as a postoperative residual tumor diameter not exceeding 1 cm, while any diameter exceeding 1 cm was deemed indicative of suboptimal cytoreductive surgery. DFS and OS were the primary outcomes. The time origin for all survival analyses was uniformly defined as the date of primary cytoreductive surgery treatment for all patients, ensuring a consistent baseline. DFS was defined as the interval between the surgical treatment and the initial disease recurrence, or mortality attributed to the disease. Recurrence was determined by the presence of elevated CA-125 levels and/or the identification of imaging abnormalities through transvaginal ultrasound, computed tomography (CT), and/or positron emission tomography (PET). Patients without disease recurrence on the date of their last follow up were censored for DFS analysis. OS was defined as the interval between the surgical treatment and the occurrence of death. Patients who were alive at the last follow-up were censored for OS analysis. All survival times were calculated in months to minimize variability in follow-up scheduling inherent to retrospective studies. Univariate Cox proportional hazard regression analysis and least absolute shrinkage and selection operator (LASSO) were employed to examine the risk factors linked to DFS and OS. Candidate variables were initially screened by univariate Cox regression ( P < 0.05) and the LASSO regression with cross-validation, respectively. Subsequently, the variables identified by two methods are incorporated into a multivariate Cox regression, and stepwise backward regression is employed to determine the final variables for the establishment of prediction model respectively, using the minimum AIC value. The two models were evaluated against each other via receiver operating characteristic (ROC) curve analysis, and the model yielding the highest area under the receiver operating characteristic curve (AUC) was selected as the final model. Base on the selected optimal model, a nomogram was developed to forecast 3/5-year DFS and OS. The discrimination of the model was assessed using the concordance index (C-index) and AUC. The calibration curve was applied to evaluate accuracy by assessing the congruity between the model’s predicted recurrence/survival probability and the observed outcomes. To evaluate the potential benefits for patients, decision curve analysis (DCA) was employed. Additionally, a time-dependent AUC curve was constructed to assess the model’s discriminatory capability over time. Furthermore, the clinical utility of the prediction model and FIGO stage was compared by the C-index, net reclassification index (NRI) and integrated discrimination improvement (IDI). The deep learning model based on neural networks is capable of learning complex nonlinear relationships between features. By combining the neural network with the traditional Cox regression model, a deep learning-based model is constructed to predict the prognosis. The output of this model is a risk score, which can be either positive or negative. Based on these scores, patients are categorized into high-risk and low-risk groups, then Kaplan-Meier survival analysis was performed to compare the prognosis between these two groups. The efficacy of the deep learning-based model and the Cox regression model is further evaluated using the C-index. All statistical analyses were conducted using SPSS (version 25.0) and R software (version 3.4.1). Statistical significance was set at P < 0.05.

Accurate prognostic prediction remains an unmet clinical need in the management of LGSOC, due to its rarity, the study of prognostic factors has been limited, impeding the development of prediction models. To address this gap, we identified four independent clinicopathological factors associated with DFS and OS, and developed a clinically practical prediction model for the prognostic risk assessment among LGSOC. The prediction model was visualized by a nomogram for predicting DFS and OS, which demonstrated robust discriminatory performance and prediction ability. Meanwhile, we further developed a deep learning-based model, which integrated the neural network with the Cox regression analysis, demonstrating the superior predictive performance, with C-index values of 0.907 for DFS and 0.922 for OS. Numerous prognostic prediction models have been developed for ovarian cancer, yet the majority are specifically designed for HGSOC [ 14 – 16 ]. As a rare subtype of ovarian cancer, LGSOC has distinct biological behavior, clinical course, and treatment response profile compared to HGSOC. Therefore, models derived from HGSOC data have limited generalizability to the LGSOC patient. In contrast, prognostic models for LGSOC remain exceedingly scarce, which forces clinicians to rely on limited traditional factors like FIGO stage for prognosis and fails to capture the heterogeneity of patients and impedes personalized risk prediction [ 8 ]. Therefore, this multicenter study aimed to develop a comprehensive prognostic prediction model specifically tailored for LGSOC patients in clinical guidance, incorporating key clinical prognostic factors - FIGO stage, age at diagnosis, cytoreductive surgery, and lymphadenectomy. Concurrently, the deep learning-based model provided a novel, data-driven approach to capture potential non-linear relationships beyond conventional methods, enabling a high-performance predictive benchmark for future research. The deep learning-based model is capable of identifying complex nonlinear relationships, demonstrating robust fitting performance and a pronounced ability to extract valuable insights from high-dimensional data [ 17 ]. Although deep learning approaches are increasingly integrated into oncological diagnostic such as radiological image analysis and cytopathological interpretation [ 18 , 19 ], their application in prognostic modeling for cancer survival prediction remains relatively underexplored. Prior studies have applied deep learning-based model to HGSOC prognosis using transcriptomic or histomorphometric data from large cohorts [ 20 , 21 ] .Our study is the first to establish deep learning-based model on a multi-center LGSOC clinical dataset, demonstrating its potential even with a limited sample size specific to this rare subtype. Our findings reaffirm the established prognostic triad in LGSOC–FIGO stage, complete cytoreduction, and patient age [ 3 , 22 ]. While the FIGO stage provides an anatomy-based framework for initial prognostic stratification by clinicians, it fails to incorporate critical determinants, such as surgical residual disease status and patient-specific factors. Consistent with prior evidence, FIGO stage was confirmed as the most influential predictor, with early-stage (I-II) disease being consistently associated with superior oncologic outcomes [ 23 , 24 ] and this pattern is corroborated by observations in HGSOC [ 25 ]. Our model represented a significant advance over the conventional FIGO staging system by effectively mitigating the prognostic heterogeneity inherent within each FIGO stage. This superior predictive accuracy of our prediction model, quantitatively demonstrated by NRI and IDI metrics in LGSOC, is consistent with prior research conveying the enhanced performance of prediction model compared to the FIGO system alone in other cancers [ 7 , 26 ]. Given the characteristically indolent growth and intrinsic chemoresistance of LGSOC, achieving complete macroscopic cytoreduction during primary surgery is paramount and reaffirmed as a cornerstone of management [ 27 ]. The critical impact of surgical effort on long-term survival is strongly supported by high-level evidence, including the GOG-182 study [ 28 ]. This multicenter study further confirms that optimal cytoreductive surgery is an independent prognostic factor, which suggests gynecologic oncologists to prioritize and execute maximal cytoreductive effort during the initial intervention of LGSOC. The therapeutic role of systematic lymphadenectomy in LGSOC patients remains controversial, with conflicting evidence derived from heterogeneous analyses [ 28 – 31 ]. Although an evaluation of the National Cancer Database suggested that LGSOC patients could have a survival benefit from lymphadenectomy [ 30 ], a separate retrospective study enrolled 126 LGSOC patients, and found no statistically significant prognostic difference between lymphadenectomy and non-lymphadenectomy groups [ 28 ]. Another retrospective investigation demonstrated a significant improvement in DFS among patients with early-stage LGSOC undergoing systematic lymphadenectomy ( P = 0.007) [ 29 ]. This multicenter study indicates that lymphadenectomy may confer a prognostic benefit. Future studies should incorporate standardized intro-operative assessment of lymph node status and postoperative pathological metastasis confirmation, in order to definitively establish the relationship between lymphadenectomy and prognostic improvement. Adjuvant therapy was not an independent prognostic factor in this multicenter study and a large-scale database study reported the same finding [ 10 ]. This may be attributed to the inherent chemoresistance of LGSOC and heterogeneity in treatment regimens. The prognostic significance of adjuvant therapy in LGSOC therefore remains an open question, requiring further investigation in larger cohorts [ 9 , 10 , 32 ]. The independent prognostic value of preoperative serum CA-125 in LGSOC remains unclear [ 9 , 23 ]. In this study, it was not selected into the final prediction model, likely because its prognostic information is statistical collinearity with the FIGO stage. Consistent with this result, one study suggests that the prognostic utility of CA-125 in LGSOC is critically dependent on the timing of assessment, with post-treatment dynamics being more informative [ 23 ]. Several limitations should be discussed. First, the retrospective design introduces inherent selection and information biases. The unavailability of transcriptomic or other multi-omics data restricted our analysis to clinicopathological variables. The integration of such molecular data in future models could enhance biological insight and predictive power. Second, owing to the rarity of LGSOC, despite drawing from four major tertiary medical centers in China, the final multi-center retrospective cohort comprised only 155 patients, which may limit the robustness of the prediction model. Finally, external validation across expanded, multi-institutional cohorts remains essential to confirm the generalizability of our model. In conclusion, we have developed and internally validated a clinically practical prediction model specifically for LGSOC, which was visualized by the nomogram. The nomogram is applicable for clinical use to guide postoperative counseling and surveillance strategy, addressing a critical gap in the risk stratification of LGSOC. Prospective, large-scale studies are essential for external validation of our models.

Low-grade serous ovarian cancer (LGSOC), as a rare subtype, accounts for approximately 5–10% of ovarian cancers [ 1 – 3 ]. Histologically, LGSOC and high-grade serous ovarian cancer (HGSOC) are distinct subtypes within the serous ovarian cancer category, exhibiting different clinical presentations and prognostic outcomes [ 3 , 4 ]. Numerous investigations have demonstrated that LGSOC exhibited indolent progression, less sensitivity of chemotherapy, and a better prognosis relative to HGSOC [ 4 – 6 ]. However, owing to the rarity of LGSOC, its clinical diagnosis and treatment primarily rely on retrospective clinical studies or subgroup analysis of clinical trials pertaining to HGSOC. Currently, no validated prediction models specifically tailored for LGSOC in Chinese cohorts exist. Consequently, there remains a dearth of personalized and precise prognostic models for evaluating the prognosis of LGSOC patients. The International Federation of Gynecology and Obstetrics (FIGO) staging system, employed in clinical settings, serves to delineate tumor progression and offer an initial prognosis for patients. While it provides a preliminary understanding of tumor invasion, it fails to consider crucial factors such as patient age, surgical satisfaction, and other clinical indicators, thereby limiting its ability to accurately assess individual patient prognoses [ 7 ].While some studies have individually confirmed the prognostic value of factors like FIGO stage and optimal cytoreduction in LGSOC, there is an absence of integrated multivariable prediction models [ 8 – 10 ]. Multivariable prediction models have demonstrated superior predictive accuracy compared to models based on single variable. The nomogram, a graphical scoring tool derived from statistical models like Cox proportional hazards regression, effectively visualizes such integrated models, translating complex statistical outputs into an intuitive, points-based system for risk calculation [ 7 , 11 – 13 ]. Compared to the traditional FIGO staging system, prediction model combines multiple prognostic factors and demonstrate superior predictive ability for clinical use. The deep learning-based model, a prognostic tool based on deep learning, leverages advanced machine learning techniques such as neural networks to process data and can uncover complex nonlinear relationships, providing a methodological reference for future large-cohort retrospective studies in LGSOC. This study aimed to develop a prediction model for predicting the prognosis of LGSOC patients and present it as a nomogram for clinical application. The goal is to establish a robust risk stratification system that provides reliable individualized survival information, thereby facilitating personalized therapeutic decision-making and guiding follow-up strategies.

Supplementary Material 1. Supplementary Material 1.