A Hybrid Model for Classification of Biomedical Data Using Feature Filtering and a Convolutional Neural Network

Deep learning is known for its capabilities in analysing large and complex sets of data, without the need of applying noise reduction methods, which is a necessary step for improving the performance of conventional machine learning models. Indeed, the superiority of deep learning over conventional machine learning models resides in their capabilities of analysing large sets of data to learn features directly from the data without the need for manual feature extraction. However, this paper aims to evaluate the hypothesis that by using feature filtering as a preprocessing step prior to feeding the data into the deep learning model, the quality of the data is improved which also leads to a better performing deep learning model. Two complex biomedical datasets which contain a large number of features and sufficient number of patient cases for deep learning were selected for the evaluations. A selection of feature filtering methods were applied to identify the most important features (i.e. top 20% ranked features) at the input level, prior to the data being fed into a deep learning classifier. Once the most important features are selected, these are fed into a deep learning algorithm, and in particular the Convolutional Neural Network, which has been tuned for the particular task. Experiment results demonstrate that applying feature filtering at the input level improves the performance of the deep Convolutional Neural Network, even for the most complex biomedical data such as those utilised in this paper. In particular, for the first dataset, PANCAN, an improvement of 20% was reported in Accuracy, whereas for the second dataset GAMETES Epistasis, an improvement of 10.63% was reported in Accuracy. The results are promising and demonstrate the benefits of filter filtering when deep learning methods are adopted for biomedical classification tasks.