Conclusions and Future Work
8.2 Directions for further Research
The results obtained in this thesis open directions to further research and improvement. The following lists points to improvement and future work:
8.2.1 Mining the Results from Multiple patient seizure detection
The volume of results produced from the comprehensive multi-patient seizure- prediction is significant. Visualisation of these results is particularly difficult as characteristics such as combination of files, characteristics of patients, characteristics of patients in the feature-set and four performance measures of Accuracy, Sensitivity, Specificity and S1-score, demand multi-dimensional representation which, even if done, would be very difficult to interpret. In this thesis, we have mainly looked at the two special cases where data could easily be visualised. Using association rules and other data-mining tools, useful results could potentially be extracted, which could reveal more about how the characteristics of Patient-Files come in to play with respect to multi- patient analysis. We suspect that there are correlations between characteristics of patients, whether in their profiles, or seizure attributes, which makes them more predictive in generalising some patients rather than others.
8.2.2 Accounting for patient similarity in Multi-Patient Seizure Prediction
In the multiple-patient seizure-prediction analysis, we discovered that a patient is generalised differently by other patients; in most cases, the majority of patients produce a similar generalisation outcome on the specific training-set, whilst a small number of patients may produce exceptionally high performance outcomes, compared to other patients. This suggests the inherent similarity among groups of patients. It would be significantly useful to understand what makes these patients similar. By using, and automatically detecting these similarities, we are able to build more powerful multi- patient analysis predictors, where patients of similar characteristics are used in training/test combinations. A simple example of such methods, is a weighted average of classification results, where patients that are similar to the tested patient are given a higher weight compared to other patients.
8.2.3 Expand the Deep Belief Nets:
Amongst the multi-patient seizure detection methods, Deep Belief Nets yielded the highest test-set outcome on both skewed and balanced datasets. The results were however reported for 29 variations of hyper-parameter sets. Since we have used random jumps to find these parameters, it is likely that we are yet to find the global optima. With further fine-tuning these parameters by either introducing more random jumps or using Gaussian Processes (chapter 8), we are able to yield potentially better outcomes (Bergstra et al. 2011).
8.2.4 On-line Seizure Detection:
In this thesis, we have only looked at off-line seizure detection, where training and test data are at hand and can be pre-processed and analyzed in batches. Now that optimal, experimental settings have been identified, it is useful to evaluate them in variations of on-line learning which is closer to the real-time application (Anderson 2008). This is significantly useful for the case of advance prediction, as prediction windows have merely been simulated in the work presented in this thesis.
8.2.5 Other machine learning algorithms
In most experiments presented in this thesis we used a single, generally good, machine learning algorithm, namely Multi-Class SVM, which was suitable for the seizure classification task. We have successfully identified the optimal experimental settings for improved, individualised epileptic seizure detection and prediction. It would be of high advantage to apply other suitable machine learning algorithms such as Bayesian methods on the best experimental setups, in order to evaluate how each algorithm measures up against our benchmark, and whether further improvements are obtained using other machine learning tools.
8.2.6 Further Exploration of Features
This thesis presented a new set of features that yielded a higher performance than that of the benchmark results produced by a previous set of features. These results were averaged across all patients, with low variation amongst patients in various stages of the experiment. However, some feature-sets led to variation in the performance among
patients. In addition to this, the feature-rankings were observed as highly patient- specific. By having a closer look at each patient, and its feature rankings and respective classification-performance, and by looking at recurring patterns across patients, we may find patient-specific or patient-group specific characteristics, which led to the particular ranking order of features.
In addition to studying individual and cluster performance-factors of the patients on the derived features, we can also look towards further expanding the feature-set to incorporate additional features, particularly multivariate ones which involve all channels of the patient in order to verify the effect of such features in a multi-feature experimental setup.
8.2.7 Multi-Modal Training Set
In chapter 8 of this thesis, we introduced two machine learning algorithms, which yielded a high performance for multi-patient classification of seizures. These methods, namely Multi-Task Learning and Deep Belief Nets, are forms of Transfer Learning, where information from solving one problem is stored and used for solving a different, but related problem. In the multi-patient classification problem, we regarded each patient’s invasive EEG data as a specific task to the Multi-Task Learning algorithm, where the ictal and non-ictal states of the invasive EEG were a commonality between the related tasks. Transfer Learning methods are also particularly suitable for Multi- Modal training-sets, where the information about a task is obtained from different modes of data retrieval. Drawing from the success of the multi-patient analysis using Transfer Learning presented in this thesis, and other research conducted on multi-modal training using this method of learning (Yuan et al. 2012; Charuvaka & Rangwala 2012), we can formulate a new problem where a patient-specific learner is trained on multiple modes of epilepsy data, such as Invasive EEG, Scalp EEG, MRI scans etc. where all modes are related in the representation of the seizure and non-seizure characteristics of a specific patient.
8.2.8 Predicting Seizures Further in Advance
In chapter 7 of this thesis, we implemented prediction algorithms with alternating prediction window lengths, on all patients from the Freiburg EEG Database. We successfully found predictive markers up to 20 minutes in advance with high