A Hybrid CNN-Multi-Class SVM Framework for Biomedical Document Gene-Disease Datasets Classification
Jose Mary Golamari, Dasari Sai Naga Haritha · International Journal of Electronics and Communication Engineering · 2023
Healthcare investigators and clinicians need biomedical document classification to organize and handle the large volume of biomedical literature. Conventional classification methods use manually designed features, which may be timeconsuming and may not represent biomedical text complexity. Biomedical data’s high dimensionality and sparsity may also challenge current approaches. For big datasets, CNNs are computationally costly. Increasing feature extraction efficiency reduces training and inference durations. The proposed method intends to improve the accuracy of document classification in the biomedical sector considerably. It functions in two stages: feature extraction and classification. The proposed method employs a hybrid approach to biomedical document classification, focusing on the intricate interactions between genes, diseases, and chemical treatments via the use of a CNN Multi-class Support Vector Machine (M-SVM) model. CNN is utilized to extract features, while M-SVM is employed as a classifier. This work discusses Improved CNNs, which may extract more discriminative and informative features from input data, resulting in a more accurate representation of underlying patterns and connections. Error-Correcting Output Coding (ECOC) based on M-SVM is used to manage noisy data by merging the outputs of many binary classifiers, enabling it to recover from faults in individual classifiers and thereby lowering the risk of overfitting. The study’s results show that the proposed model is successful, with an accuracy of 99.28% and an F1-score of 99.84% across biomedical document datasets.