The medical expenses of cancer patients have been rising worldwide, and the current payment system based on diagnosis‐related groups is limited in reflecting the actual burden at the patient level. We aimed to develop a prediction model for reasonably high cost of hospitalization, taking gastric cancer (GC) patients undergoing surgery as an example.

In this real‐world patient cohort study, we collected clinical data from 3981 GC surgery inpatients in a tertiary hospital in Beijing, China. Reasonably high cost was defined as exceeding the 75th percentile of the hospitalization costs. The final model for reasonably high cost was derived based on univariate and multivariate unconditional logistic regression, with predictors selected according to the Akaike information criterion. Model performance was evaluated in terms of discrimination and calibration.

The final model consisted of 13 predictors, with the area under the receiver operating characteristics curve of 0.718 (95% confidence interval: 0.700–0.736). Leave‐one‐out cross‐validation generated an area under the receiver operating characteristics curve of 0.708 (95% confidence interval: 0.690–0.727). No significant differences in model discrimination were detected among the subgroups stratified by gender, age, N stage, surgical method, and admission year. The calibration plot showed good agreement between predicted probability and observed frequency of reasonably high cost.

Our prediction tool for reasonably high cost of GC surgery patients has demonstrated good performance. The combined use of diagnosis‐related groups payment and reasonably high cost assessment is expected to promote individualization and precision in medical cost control and ensure rational, efficient, and equitable provision of medical services.

To construct and validate a prognosis prediction model and a risk stratification tool for more precise and individualized evaluation of prognosis for patients following resection of esophageal squamous cell carcinoma (ESCC), and provide real-world evidence for informing optimal decision-making about adjuvant therapy.

The comprehensive clinical data and follow-up data were collected from consecutive patients with ESCC in the Anyang Cancer Hospital (Anyang center) from May 31, 2011 to July 31, 2018, and in the Cancer Hospital of Shantou University Medical College (Shantou center) from August 1, 2009 to December 31, 2018. Patients from the Anyang center formed the training cohort, and a two-phase selection based on backward stepwise multivariable Cox proportional hazard regression and minimization of AIC was used to construct prediction model for overall survival (OS). Bootstrap with 1 000 resamples was used for internal validation, and cohort from the Shantou center was used for external validation. Furthermore, a risk stratification tool was constructed according to the tertiles of the total points derived from nomogram in the training cohort.

A total of 4 171 eligible patients were included in the training cohort, and 1 895 patients were included in the validation cohort. The final model incorporated nine variables: age, sex, primary tumor location, T stage, N stage, number of lymph nodes harvested, tumor size, adjuvant treatment, and preoperative hemoglobin level. A significant interaction was observed between N stage and adjuvant treatment (P < 0.001), which means that N+ stage patients were likely to benefit from addition of adjuvant therapy as opposed to surgery alone, but adjuvant therapy did not improve OS for N0 stage patients. The C-index of the model was 0.728 (95% CI: 0.713-0.742) in the training cohort, 0.722 (95% CI: 0.711-0.739) after bootstrapping, and 0.679 (95% CI: 0.662-0.697) in the external validation cohort. Calibration plots demonstrated favorable agreement between model prediction and actual observation for 1-, 3- and 5-year OS. In both training and validation cohorts, this model outperformed the seventh edition of the AJCC TNM (tumor, lymph node, and metastasis) staging system in terms of the accuracy of prognostic prediction (P < 0.05). Moreover, within each TNM staging group, this model achieved ideal risk stratification.

The prediction model constructed in this study may provide individualized survival prediction for patients with resected ESCC in China. This study also demonstrated that the N stage may be a fundamental determinant in planning postoperative adjuvant therapy for ESCC patients.