Deep Learning Outperforms Standard Statistics in Brain Tumour Prognosis

Predicting the clinical course of adult-type diffuse glioma (ADG), a serious type of brain tumour, has long relied on traditional statistical tools. However, a new study suggests that the future of prognosis lies in artificial intelligence. Researchers systematically compared standard methods against advanced machine learning approaches to determine which could best predict patient survival using real-world data.

The team evaluated four distinct models, pitting the conventional Cox Proportional Hazards model against a Random Survival Forest and two neural network-based approaches. Using data from two public sources and one private cohort, the results were clear: the deep learning model known as DeepSurv delivered the most robust performance. It significantly outperformed both the standard statistical methods and the Random Survival Forest algorithm.

DeepSurv proved particularly superior in handling the ‘messy’ nature of retrospective medical data, which often contains high heterogeneity and missing values. By analysing key factors—including age, molecular pathology, chemotherapy, radiotherapy, and the extent of surgical resection—the model provided stable and accurate predictions. The researchers have made this powerful modelling tool publicly available, marking a significant step forward in using computational intelligence to navigate the complexities of cancer care.