From January 2007 to December 2018, medical records of all patients who underwent debulking surgery in our center were comprehensively reviewed and evaluated by gynecologists. The inclusion criteria were as follows: 1. histopathologically confirmed EOC; 2. no history of other malignancies; 3. no other anti-cancer therapies before surgery or neoadjuvant chemotherapy. Patients were excluded if they met any of the following exclusion criteria: 1. suboptimal surgery; 2. incomplete clinical data; 3. non-standard surgery or non-standard first-line chemotherapy; 4. early-stage EOC patients who underwent fertility-saving procedures; 5. previous hysterectomy or unilateral/bilateral salpingo-oophorectomy before diagnosis (e.g., previous treatment of hysteromyoma or ectopic pregnancy); 6. operational records were not detailed enough to calculate PCI score, PSDSS, adjusted AGO score, or surgical complexity score (SCS); 7. death or loss to follow-up within 3 months after initial treatment. After evaluation, a total of 1173 patients were included in this study and were randomly allocated into a derivation cohort of 879 patients and an internal validation cohort of 294 patients, in a ratio of 3:1. Five hundred and seventy-seven patients from another three cancer centers served as the external validation cohort. This retrospective study was approved by the Medical Ethics Committee of all institutions involved and was conducted under relevant guidelines. The workflow is summarized in Figure 1 .

Factors that might impact OS and PFS were extracted and analyzed. Demographic characteristics including age, Eastern Cooperative Oncology Group performance, body mass index, hemoglobin and platelet counts, and neoadjuvant chemotherapy (NACT) status were acquired. Inflammatory data, including leukocytes counts or neutrophil and monocyte percentages, were collected. Tumor characteristics (pathological diagnosis and differentiation, maximum tumor diameter, presurgical CA125 and HE4s, ascites cytology, etc.) were obtained, as well as surgical information including the duration of surgery and amount of blood loss. Biochemical indexes including but not limited to ALT, AST, GGT, TP, ALB, CREA, PT, INR, and APTT, were measured.

We used PCI score and PSDSS to represent the level of tumor burden observed during surgery, as both of these scoring systems are objective, implementable, and frequently used in other malignancies (15–18). PCI score was calculated by assessing the extent of intra-peritoneal disease at diagnosis through quantitatively combining the cancer implant size with the tumor distribution throughout 13 abdominopelvic regions, producing a maximum score of 393. PSDSS was calculated by integrating clinical symptoms, the extent of carcinomatosis, and primary tumor pathology as reported by Antonio et al. (15). We also consulted the settings of the AGO score by utilizing three vital factors to assess the gross disease of patients (complete cytoreduction at initial surgery, no ascites, and good performance status at recurrence). As the AGO score was created for evaluation of secondary cytoreduction, this study used an adjusted AGO score, defined based on complete cytoreduction, no ascites, and good performance status at initial surgery. SCS was used to describe the complexity of the cytoreductive surgery (19, 20). The score sheet used for calculating PCI score, PSDSS, adjusted AGO score, and SCS is summarized in Supplementary Figure 1 .

For patients with comorbidity, poor performance status, and predicted suboptimal debulking surgery, neoadjuvant chemotherapy was performed before initial surgery. The possibility of suboptimal resection in the primary debulking setting was assessed by at least two experienced gynecological oncologists and one radiologist. All patients who underwent NACT were assessed with computed tomography (CT) and had Suidan score ≥3. Before NACT was initiated, fine needle biopsy (only for those whose clinical conditions couldn’t tolerate a diagnostic laparoscopy) or diagnostic laparoscopy was completed for histologic confirmation of diagnosis. For those who received interval debulking surgery, surgeries were performed after < 4 cycles of NACT. After surgery, adjuvant chemotherapy was performed when clinically indicated, according to the standard of care based on NCCN guidelines. All patients involved in this research received taxane/platinum-based systemic chemotherapy. Those who did not complete initial treatment or switched to a non-standard regimen were excluded from this study.

All surgeries were performed by experienced gynecological oncologists. For patients in the initial stages (FIGO stages I–II), comprehensive surgical staging was performed, with standard surgical treatment consisting of hysterectomy, bilateral salpingo-oophorectomy, omentectomy, and retroperitoneal (pelvic and para-aortic) lymphadenectomy. For patients with advanced disease (FIGO stages III–IV), cytoreductive surgeries were performed. Surgical approaches included but were not limited to the operations mentioned above, together with other tumorectomies of metastatic lesions if applicable, in an attempt to reach a state of no gross residual disease (complete surgical resection). Peritoneal washing was routinely conducted for all patients.

Patients were followed up every 3 months for the first 2 years after complete initial surgery and adjuvant chemotherapy. After 2 years, patients were followed up annually. Each follow-up consisted of a blood test to determine CA125 and HE4 levels, biochemical tests, and imaging such as abdominal ultrasound or contrast-enhanced CT/magnetic resonance imaging. Recurrence was defined as a persistent elevation in CA125 level or metastases detected by CT/ultrasound. Recurrent EOC was treated by further surgical resection, radiofrequency ablation, or chemotherapy according to the recurrence pattern. The end of follow-up was May 2021. The median follow-up period was 41 months (range: 6.0–160.7 months) for the cohort of patients from our center and 40 months (range: 6.0–149.5 months) for the external validation cohort.

The endpoints of this study were PFS and OS. PFS was calculated as the interval between the last day of initial treatment and the date when a recurrence was confirmed. OS was defined as the period between the last day of initial treatment and the date of the patient’s death or the date of the last follow-up. Kruskal–Wallis analysis of variance (ANOVA) tests and ANOVA t-tests were performed to compare continuous data. For discrete data, χ²-tests were carried out. Survival was calculated by the classical Kaplan–Meier method and compared between groups by log-rank tests. Variables screened as significant by univariate analysis were used in multivariate Cox proportional hazards regression analysis to identify independent prognostic factors.

Based on the multivariate analysis of the training cohort, two nomograms were established. Final model selection was performed by a backward step-down process with the Akaike information criterion. The two nomograms were assessed by Harrell’s concordance index (C-index) and calibration curves comparing model-predicted vs real-world Kaplan–Meier estimates of OS and PFS.

Nomograms were then validated in the internal and external validation cohorts. The nomograms were also compared with the FIGO staging system by analysis of receiver operating characteristic (ROC) curves. SPSS (version 22.0, IBM, Armonk, NY) and R (version 3.6.3) software was used during those procedures. P < 0.05 was defined as statistically significant.

This multicenter study recruited 1750 patients with a median age of 52.0 (interquartile range: 45–60) years. Of these patients, 630 patients underwent interval debulking surgery (36.0%), whereas others received primary debulking surgery (74.0%). The clinicopathological data of the patients in the training and validation cohorts, including demographic factors, imaging results, surgical factors, and preoperative blood test results, are summarized in Supplementary Table 1 .

Univariable and multivariable analyses were performed in the training cohort. All significant factors (P < 0.05) in the univariable analysis were entered into the multivariable analysis via a Cox regression model. The results showed that FIGO stage, pathological differentiation, NACT, number of organ sites of metastases, blood loss, PSDSS, SCS, and HE4 level were independent risk factors associated with PFS, whereas FIGO stage, pathological differentiation, NACT status, number of organ sites of metastases, ascites cytology, PSDSS, adjusted AGO score, and SCS were independent risk factors associated with OS ( Figure 2 ).

To predict OS and PFS of patients with EOC, two nomograms were formulated using R software based on the findings of the multivariable analysis ( Figure 3 ). Each value of these variables was assigned as a subtype, therefore a score on the point scale. By adding up the scores for each variable and positioning it on the total point scale, a user can predict the PFS and OS of an individual patient. Red points in the graph represent the values of an observed case. The size of the green boxes represents the proportion of this type of patient in the training cohort, and the density represents the distribution of the variable in the data set.

The Harrell’s C-index values for the PFS and OS nomograms in the training cohort were 0.82 (95% CI: 0.76–0.88) and 0.84 (95% CI: 0.78–0.90), respectively. In the internal validation cohort, the C-index values for the PFS and OS nomograms were 0.80 (95% CI: 0.74–0.86) and 0.81 (95% CI: 0.75–0.87), respectively. In the external validation cohort, the C-index values for the PFS and OS nomograms were 0.84 (95% CI: 0.79–0.89) and 0.87 (95% CI: 0.82–0.92), respectively. The calibration plots showed remarkable agreement between the nomograms’ predictions and actual observations for 1-, 3-, and 5-year PFS and OS in the training cohort ( Figures 4A–D ), internal validation cohort ( Figures 4B–E ), and external validation cohort ( Figures 4C–F ).

We compared the predictive abilities of the PFS and OS nomograms with those of the FIGO staging system using ROC curve analysis. Our nomograms showed better discriminatory power in the training cohort than did the conventional FIGO staging system ( Figure 5 ). For the PFS nomogram, the C-index value was 0.82 (95% CI: 0.76–0.88), substantially higher than that of the FIGO staging system (0.74, 95% CI: 0.71–0.81; P < 0.001). For the OS nomogram, the C-index value was 0.84 (95% CI: 0.78–0.90), significantly higher than that of the FIGO staging system (0.76, 95% CI: 0.70–0.78; P < 0.001).

We determined appropriate cutoff values by grouping the patients in the training cohort evenly into three subgroups after sorting by total score. The OS nomogram could also stratify patients into low risk (score ≤ 172), medium risk (score 172 to 220), and high risk (score > 220) groups. The 5-year PFS rates were 87.4%, 53.2%, and 27.5%, respectively, in the training cohort (P < 0.001) ( Figure 6A ). Similar results were obtained in the internal validation cohort ( Figure 6B ) and in the external validation cohort ( Figure 6C ). Similarly, the PFS nomogram was able to accurately stratify patients as low risk (score ≤ 172), medium risk (score 173-208), or high risk (score > 208) of recurrence, where each group represented a distinct prognosis ( Figure 6D ). The 5-year PFS rates were 71.1%, 25.1%, and 13.6%, respectively, in the training cohort (P < 0.001). Similar results were obtained in the internal validation cohort ( Figure 6E ) and the external validation cohort ( Figure 6F ). While stratifying patients into different risk groups, we took the nomogram’s score of patients from training group into the Cox-model, to explore the relationship between the nomogram score and the probability of >5-year PFS/OS. Using nomogram score as the horizontal axis, and the probability of >5-year PFS/OS as the vertical axis, we found that those two variables are almost linear ( Supplementary Figure 3 ).

The online versions of our nomograms are available for public use; clinicians and researchers can visit https://eocnomogram.shinyapps.io/eocpfs/ and https://eocnomogram.shinyapps.io/eocos/ to predict the long-term survival of postoperative patients with EOC. Predicted survival probability across time can be easily determined by inputting clinical features and reading the output figures and tables generated by the webserver ( Supplementary Figure 2 ).