JMIR Publications

ABSTRACT

Background:

Despite the global decline in the incidence of gastric cancer, the number of young individuals diagnosed with it continues to rise. Several studies have been conducted to predict the mortality of patients with gastric cancer. Therefore, we propose short-, medium-, and long-term mortality prediction models for young patients with gastric cancer based on a survival machine learning model.

Objective:

We propose short-, medium-, and long-term mortality prediction models for young patients with gastric cancer based on a survival machine learning model.

Methods:

Data of 1,200 young (<50 years) patients diagnosed with gastric cancer between 2013–2015 were obtained from the Gastric Cancer Public Staging Database. Data of 840 and 360 patients were used for training and testing, respectively. We employed the random survival forest (RSF), gradient boosting survival analysis (GBSA), extra survival tree (EST) prediction models and cox proportional hazards (CoxPH) for 1-, 3-, and 5-year survival prediction, and the concordance index (C-index) with 95% Confidence intervals (CI) metric to objectively assess the models. This study also examined the key determinants of mortality based on the prediction time points.

Results:

For each survival model, the C-index was reported with 95% CI. For the RSF model, it was 95.95%(95.93-95.97) for 1-year mortality, 95.32%(95.31-95.33) for 3-year mortality, and 93.45%(93.44-93.46) for 5-year mortality. In comparison, for the GBSA model, it was 96.89%(96.89-96.90) for 1-year mortality, 94.90%(94.90-94.90) for 3-year mortality, and 93.56%(93.56-93.56) for 5-year mortality. For the EST model, it was 97.08(97.07-97.09) for 1-year mortality, 95.45(95.43-95.46) for 3-year mortality, and 93.90(93.88-93.91) for 5-year mortality. In addition, the CoxPH model showed a C-index of 89.97(89.97-89.97) for 1-year mortality, 95.56(95.56-95.56) for 3-year mortality, and 92.19(92.19-92.19) for 5-year mortality. For the Tumour stage and size were the primary variables employed for training the models to predict mortality at different time points. The other variables exhibited varying degrees of consistency for each time point.

Conclusions:

The findings are expected to facilitate the identification of high-risk young patients with gastric cancer who may benefit from aggressive treatment by predicting their risk of death at various time points.