Seizure prediction could greatly improve the quality of life of people suffering from epilepsy. Modern prediction systems leverage Artificial Intelligence (AI) techniques to automatically analyze neurophysiological data, most commonly the electroencephalogram (EEG), in order to anticipate upcoming epileptic events. However, the performance of these systems is normally assessed using randomized splitting methods, which can suffer from data leakage and thus result in an optimistic evaluation. In this review, we systematically surveyed the available scientific literature looking for research approaches that adopted more stringent assessment methods based on patient-independent testing. We queried three scientific databases (PubMed, Scopus, and Web of Science), focusing on AI techniques based on non-invasive EEG recorded from human subjects. We first summarize a standardized signal processing pipeline that could be deployed for the development and testing of cross-patient seizure prediction systems. We then analyze the research work that meets our selection criteria: 21 articles adopted patient-independent validation methods, constituting only 4% of the published work in the entire field of epileptic seizure prediction. Among eligible articles, the most common approach to deal with cross-patient scenarios was based on source domain adaptation techniques, which allow to fine-tune the predictive model on a limited set of data recorded from a set of independent target patients. Overall, our review indicates that epileptic seizure prediction remains an extremely challenging problem and significant research efforts are still needed to develop automated systems that can be deployed in realistic clinical settings. Our review protocol is based on the PRISMA 2020 guidelines for conducting systematic reviews, considering NHLBI and ROBIS tools to mitigate the risk of bias, and it was pre-registered in PROSPERO (registration number: CRD4202452317).
