Evaluation of optimum classification measures used to define textbook outcome among patients undergoing curative-intent resection of gastric cancer

Prospectively maintained electronic databases of patients with gastric adenocarcinoma undergoing gastrectomy with curative intent in two academic surgical centres were reviewed. The Kraków cohort included patients treated in the First Department of Surgery, Jagiellonian University Medical College between 1996 and 2020. The Lublin cohort recruited patients treated in the Department of Surgical Oncology, Medical University of Lublin between 2010 and 2020. The first cohort was used to evaluate the candidate definitions of TO and identify potential predicting factors. Subsequently, both cohorts were used to develop and validate a nomogram predicting the odds of achieving TO (Figure S1, supporting information). This study has been approved by the Bioethics Committee of the Jagiellonian University.

Data related to surgical treatment and staging were reclassified according to the most recent guidelines [24, 25]. The postoperative course was evaluated using the Clavien-Dindo classification of surgical complications [26]. TO was defined using six definitions previously reported for gastric cancer (Table S1, supporting information) [18‐23]. Individual components used to formulate TO included macroscopically complete resection according to the surgeon, microscopically radical resection by the pathologist (R0), minimum number of lymph nodes retrieved and examined, no intraoperative complication, no grade II or higher postoperative complication according to the Clavien–Dindo classification, no reintervention (surgical, endoscopic or radiological), no readmission to the ICU, no postoperative mortality, duration of hospital stay, no readmission within 30 days after discharge from hospital, and compliance with chemotherapy as recommended by the National Comprehensive Cancer Network (NCCN) clinical practice guidelines. The minimum number of lymph nodes was evaluated adopting two cut-off values of 15 and 16 nodes [19, 22]. Prolonged hospitalization was defined as hospital stay exceeding 19 days, 21 days, and 75^th percentile of the cohort [19, 22, 23]. Postoperative mortality was defined as in-hospital death or death within 30 days or 90 days after surgery [19, 20]. The primary evaluation criterion for TO was overall survival defined as time from the date of surgical resection to the date of death from any cause or date of the last follow-up. All death dates were verified by data from the national census office.

The differences in proportions between groups were evaluated using the chi-square test, and the Mann–Whitney U test was used to detect differences in quantitative variables. Overall and conditional survival (under the condition of surviving the first postoperative 30 days and hospital stay) was analysed according to the Kaplan–Meier method and the log-rank test was used to detect differences between groups. Multivariate analysis was performed using a Cox proportional hazards model with a backward stepwise selection procedure. The goodness of fit of individual Cox models were compared using the Akaike (AIC) and Bayesian information criterion (BIC), while their predictive value with Harrell's concordance index (c-index) [27, 28].

After checking for multicollinearity, multivariable logistic regression was used to identify variables associated with achieving a textbook outcome. Three penalized regression methods (ridge, lasso, and elastic net regression) were used for selection of predictive variables and formulation of a multivariable logistic regression model based on minimising the value of AIC [29]. Internal validation of the predictive model was performed by bootstrapping with 1000 resamples [30]. Calibration plots were generated to evaluate the accuracy of prediction and model discrimination was quantified using the c-index corresponding to the area under the Receiver Operating Characteristic (ROC) curve. A nomogram was formulated based on the results of the logistic regression model. The significance level (P) < 0.05 in a two-tailed test was considered statistically significant. All statistical analyses were performed using the IBM® SPSS® Statistics 28 software package (IBM Corporation, NY) and RStudio (Integrated Development Environment for R) version 2021.9.2.382 with packages survival (3.3–1), rms (6.2–0), pROC (1.16.2), and caret (6.0–91).

The study population includes 1,743 patients with non-metastatic gastric adenocarcinoma treated in two academic surgical centres, corresponding to Kraków (n = 1,479) and Lublin (n = 264) cohorts. Baseline patient characteristics is shown in Table 1.

At the time of final follow-up, 1059 of 1479 (72%) patients in the Kraków cohort had died. The median follow-up for all surviving subjects was 101 months. Six Cox proportional hazards models were developed corresponding to previously published definitions of TO (Table S2, supporting information). Of all models tested, the model formulated by Busweiler et al. showed the best goodness of fit by achieving the lowest values of AIC and BIC and the best predictive performance based on the highest value of c-index. Therefore, the 10-item definition was selected as the optimum definition of TO for the current study (no intraoperative complication, macroscopically and microscopically radical resection, ≥ 15 lymph nodes examined, no Clavien–Dindo grade II or higher complication, no reintervention within 30 days after surgery, no readmission to ICU within 30 days after surgery, no postoperative mortality within 30 days after surgery, hospital stay ≤ 21 days, no hospital readmission within 30 days after discharge). Supplementary Figure S2 and S3 shows Kaplan–Meier overall and conditional survival curves for patients achieving individual TO metrics except the item referring to postoperative 30-day mortality. Apart from hospital readmission within 30 days and the number of evaluated lymph nodes, all individual parameters were associated with superior survival by the Cox proportional hazards model (Table S3, supporting information). Moreover, there was a gradual reduction in relative hazard of death when increasing the number of achieved TO measures from 5 or less items up to 10 items (Table S4, supporting information).

The absence of Clavien-Dindo grade 2 or higher complications (47%) was the most common limiting factor for TO, while the prevalence of other factors did not exceed 24% (Fig. 1). LOESS curve fitting showed a gradual increase in the annual proportion of patients achieving TO from 18.4% in 1996–2000 to 34.3% in 2016–2020 (P = 0.001). This trend was largely due to increased rates of resections with curative intent, lack of intraoperative complications, resections with at least 15 lymph nodes, and shortened hospital stay (Figure S4, supporting information). No major variability was found for 30-day mortality, 30-day readmissions, and microscopically radical resections. Figure S5 demonstrates changes in Clavien-Dindo severity scores according to the study period and figure S6 shows severity scores related to the type of surgery.

The overall median survival was significantly longer for patients with (69.0 months, 95% CI 53.3–96.3) than without textbook outcome (20.1 months, 95% CI 17.0–22.4, P < 0.001) (Fig. 2). Since postoperative mortality is a component of TO, a conditional survival analysis was also carried out excluding deaths in-hospital and within 30 days after surgery, producing similar Kaplan–Meier survival curves. Additionally, TO maintained its prognostic value in a Cox proportional hazards model controlling for age, sex, comorbidities, treatment, and tumour related variables (Table S5, supporting information) and was associated with a 39% lower risk of death (HR 0.61, 95% CI 0.51 – 0.73, P < 0.001).

Figure 3 shows odds ratios for all the variables potentially associated with TO by univariate analysis and the final model coefficients identified by multivariable regression. Patients’ age, ASA class 3 or 4, tumour size greater than 70 mm, tumour infiltrating gastric serosa or surrounding organs (pT4a or pT4b), and multivisceral resections were associated with a lower likelihood of achieving TO. Patients with lymph node metastases, tumours located in the distal third of the stomach, as well as treated by total gastrectomy or D2 lymphadenectomy had an increased odds of having TO.

Nine variables identified as predictors of TO were used to develop a nomogram combining datasets from two patient cohorts. Figure 4 shows the graphical representation of the nomogram. Model calibration by bootstrapping and calibration belt showed very good correlation between the predicted and actual probability of achieving TO. The AUC of ROC obtained from the logistic regression model was 0.752 (95% CI 0.727 – 0.781).