Development of a dynamic prediction model with the inclusion of time-dependent inflammatory biomarker enhances recurrence prediction after curative surgery for stage II or III gastric cancer

Abstract

Background: Postoperative recurrence prediction models for gastric cancer often rely on preoperative or immediate postoperative data, overlooking time-dependent biomarkers from follow-up visits. By incorporating longitudinal biomarker data through a landmarking approach, this study aims to enhance recurrence risk prediction.

Methods: This multicenter study included patients who underwent curative surgery for stage II-III gastric cancer from January 2010 to December 2016 in three hospitals in Tokyo, Japan. Their demographic, clinical, and biomarker data were collected from medical records. Biomarkers were collected at surgery and 3, 6, 9, and 12 months postoperatively. Three prediction models-baseline model, landmarking 1.0, and landmarking 1.5-were developed and compared in terms of their prediction accuracy using four measures: concordance probability, calibration plot, Kaplan-Meier curves stratified with predicted risk, and Net Reclassification Improvement. The models aimed to predict recurrence within three years after surgery, with predictions made one year postsurgery.

Results: The study included 274 patients with gastric cancer, with 62 (22.6%) events occurring within three years. As a result of the variable selection process, lymphatic venous Invasion (LVI), pathological T (pT) stage, pathological N (pN) stage, and baseline prognostic nutritional index (PNI) were chosen. Additionally, in landmarking 1.0 and 1.5, S1 treatment status and PNI-change were also selected as time-dependent predictors. Landmarking 1.5, which incorporates time-dependent biomarkers until one year postsurgery, showed superior performance to the other models in all four measures.

Conclusions: Prediction models incorporating postoperative information could serve as a decision-making tool in clinical practice to more precisely distinguish between patients with high and low risk of recurrence.

Keywords: dynamic prediction; gastric cancer; landmarking; recurrence; surveillance.

Figure 1

Flow chart.

Figure 2

Kaplan–Meier curve of RFS stratified by event risk obtained from the baseline model at 0 year. Patients were classified into three risk groups based on their predicted risk: high risk (<67.0%), middle risk (<86.3%), and low risk (≥86.3%). The P-value was obtained from the log-rank test.

Figure 3

Kaplan–Meier curve of RFS stratified by event risk obtained from (a) baseline model and (b) landmarking 1.5 at 1 year. P-values were obtained from the log-rank test. For (a), patients were classified into three risk groups based on their predicted risk: high risk (<67.0%), middle risk (<87.3%), and low risk (≥87.3%). For (b), patients were classified into three risk groups based on their predicted risk: high risk (<86.9%), middle risk (<92.6%), and low risk (≥92.6%). Landmarking 1.5 demonstrates improved predictive accuracy at the 1-year postoperative mark. Landmarking 1.5 better stratifies patients by risk, particularly in identifying high-risk patients.

Figure 4

Calibration plot of the prediction models the calibration of the landmarking 1.5 shows superior alignment between predicted and observed outcomes compared to the baseline model.