Applications of Data Analytic Modeling for Seizure Prediction

Abstract

Many recent studies on online seizure prediction from iEEG signal describe various machine learning methods for improved prediction performance. However, meaningful performance comparisons are difficult to achieve, in the absence of well-defined methodological guidelines (used for comparison). Many studies emphasize the impact of machine learning algorithms, but overlook the importance of experimental procedure (for performance evaluation) and proper selection of system design parameters. This thesis initially investigates diverse factors necessary for understanding online seizure prediction. In particular, we show that it is possible to demonstrate improved prediction performance, simply by adjusting some factors, such as specification of training and test data, definition of lead seizures, specification of prediction horizon etc. These methodological aspects arise in all seizure prediction studies because estimated predictive models are patient-specific, iEEG signal is non-stationary, and the number of recorded seizures is small. In this thesis, we investigate and quantify the effect of system design parameters (besides machine learning method itself) on overall prediction performance, using empirical analysis of real-life data, along with conceptual comparisons using synthetic data.Our second contribution is better understanding of the notion of ‘leading’ seizures (aka lead seizures) that is critical for evaluating performance of online systems. We show that specification of (naturally occurring) lead seizures can be done by analyzing seizure clusters present in long-term seizure recordings. Our analysis indicates that the pattern of seizure clusters is patient-specific (as expected). However, based on our analysis, it is possible to specify the same value of seizure-free period (T=3 days), that can detect all seizure clusters for all patients. This enables reliable detection of all lead seizures corresponding to natural seizure clusters observed in long-term iEEG recordings. However, this data-driven specification of lead seizures results in very few seizures in the training data, making the task of machine learning more challenging. Our last contribution is the development of a new machine learning method, called Group Learning, for estimating predictive models from very sparse high-dimensional data (commonly used for online seizure prediction). This approach allows training using small number of high-dimensional training samples. That is, a large number (d) of input features is represented as several (t) groups of features of lower dimensionality (d/t). This (modified) training data is then used to estimate a classifier for making predictions in (d/t)-dimensional space. During testing stage, prediction for a test input (with d input features) is made by combining all t predictions (made by a trained classifier) via intelligent post-processing rules. Such post-processing rules, proposed in our research, effectively reflect global properties of application data. We demonstrate the effectiveness of Group Learning for two real-life application domains, i.e., handwritten digit recognition and seizure prediction from iEEG signal. These empirical results show superior performance of the Group Learning approach for sparse data, under both balanced and unbalanced classification settings.