29. Performance Comparison Between Deep Learning and Machine Learning Models for Gene Mutation-Based Text Classification of Cancer

Identifying genetic mutations that contribute to cancer tumors is the key to diagnosing cancer and finding specific gene mutation-based treatment. It is a very challenging problem and a time-consuming job. Currently, clinical pathologists classify cancer manually, and they need to analyze and organize every single genetic mutation in cancer tumors from clinical text. The text data analysis can be automated using Deep Learning and Machine Learning classification techniques to ease the manual work needed to extract information from clinical text. This paper aims to analyze the performance of Machine Learning and Deep Learning methods to classify cancer from gene mutation-based clinical text data. This paper uses Natural Language Processing techniques, namely, CountVectorizer and TfidfVectorizer, and Keras API’s One-Hot encoding and to-categorical utility, to vectorize the categorical and text data and transform them into numerical vectors. Machine Learning classification algorithms and Deep Learning methods are then applied to the extracted features, and the most accurate combination of feature extraction and a classifier is discovered. Keras API’s Embedding Layer (Word Embeddings) and Bidirectional Long-Term Short-Term Memory (Bidirectional LSTM) techniques using original and augmented text data from NLPAug library are applied as Deep Learning methods. The Keras Word Embeddings using augmented text data ha performed the highest with an accuracy of 80.67%, the weighted average precision of 0.81, recall of 0.81, F1 score of 0.81, and the log loss of 0.6391. As for the Machine Learning classification algorithms, Random Forest and Stacking classifiers are explored within this paper, and the highest accuracy of 67.02% is achieved from the Random Forest classifier with the weighted average precision of 0.70, recall of 0.67, F1 score of 0.65, and the log loss of 1.0523.