Conference Paper

Purpose: The unpredictability of seizures significantly impacts the quality of life for individuals with epilepsy. While seizure forecast has emerged as a promising approach, lack of standardization in protocols for developing and testing algorithms inhibits direct comparison between state-of-the-art approaches, both hindering progress in the field and limiting clinical applicability. In this work we present a Python-based framework designed to standardize the development, evaluation, and reporting of individualized algorithms for seizure forecast. Method: The framework was tested using the My Seizure Gauge Seizure Forecasting Challenge dataset to develop three distinct models: (1) a periodicity model based on seizure times, (2) a logistic regression model using time series features from wearable devices, and (3) an ensemble model combining the two. The framework was used throughout the whole model development process. This included automation of cross-validation splits, labeling of data, postprocessing of forecasts, computation of deterministic and probabilistic evaluation metrics, and visualization of results. Observations from the development process were used to assess the framework's usability and effectiveness. The framework is publicly available at https://github.com/anascacais/SeFEF and on PyPI to facilitate broader adoption by the community. Results: Preliminary observations suggest that the framework simplifies data preparation, facilitates consistent evaluation across different approaches, and supports direct performance comparisons, even when using diverse data types and modeling approaches. While formal metrics were not collected, anecdotal evidence during its application indicated reduced development time and minimized the risk of unintentional methodological variations that could compromise comparability between approaches. Conclusion: This framework addresses critical challenges in seizure forecasting research by promoting methodological consistency and enabling reliable comparisons between diverse algorithmic approaches. Its adoption could drive future efforts into a convergent approach regarding study design and evaluation methodologies, supporting the clinical translation of seizure forecasting tools. Further validation is necessary to confirm its broader applicability.