APDF: An active preference-based deep forest expert system for overall survival prediction in gastric cancer

Abstract

Accurate survival prediction is crucial for doctors to choose appropriate treatment options and make prognoses for gastric cancer patients. While machine learning algorithms have shown promise for improving prediction accuracy, the lack of doctor involvement and guidance can reduce the credibility of prediction results in practical clinical applications. To address this issue, we propose a survival prediction expert system integrating deep forest and active preference learning to incorporate doctors’ empirical knowledge into the prediction process. Our expert system first reduces feature dimensionality through feature selection, with doctor input to ensure clinical relevance. We then train estimators to simulate clinical prediction, and doctors compare these estimators in pairs and give preferences based on their experience. A recursive function based on the radial basis function is used to learn the preferences and assign weights to all the estimators. We apply our approach to two publicly available datasets and one private dataset. Our system achieves an AUC of 0.87 and a C-index of 0.75 for predictive classification and regression analysis tasks. Importantly, our system outperforms the clinical gold standard method (Nomogram) and other state-of-the-art machine learning methods with a maximum improvement of 12% and 10%, respectively, demonstrating superior performance. In conclusion, our expert system effectively incorporates doctors’ empirical knowledge into the survival prediction process, enhancing the credibility of prediction results in practical clinical applications. Our approach has significant potential for improving treatment decisions and prognoses for gastric cancer patients.