Real-Time Migraine Prediction Using Mathematical and Time–Series Models
Abstract
Migraine prediction is a difficult rare time-series problem with episodic occurrence, non-linear dynamics and extreme imbalance between classes. Conventional statistical models like the autoregressive models are able to provide a structured time analysis but do not reflect nonlinear dependence and dynamic risk behaviour. Recent progress in machine learning and deep learning, especially recurrent and hybrid models, have enhanced predictive accuracy, but the current research on the topic does not pay enough attention to such critical problems as imbalance of classes, long-horizon stability, uncertainty quantification and standardized evaluation. As wearable and multimodal physiological data have become even more widespread, the problem of migraine prediction has become a real-time prediction problem, bringing with it the added complexities of sparsity and personalization of the data. In this paper, I am going to provide a systematic review of the methodologies of migraine prediction, starting with classical time-series models to deep learning based on probabilities. The paper establishes the major gaps in research and suggests a single point of view on the creation of adaptive, imbalance-sensitive and uncertainty-adjusted forecasts. Comparative analysis reveals that the hybrid models, i.e., ARIMA-LSTM and probabilistic deep learning models, are more robust and their precision-recall performance is better than in the rare-event forecasting situations.