Parameter Tuned Unsupervised Fuzzy Deep Learning for Clinical Data Classification

• Published in: 2022 International Conference on Electronic Systems and Intelligent Computing (ICESIC) pp. 358 - 363
• Main Authors: Saranya, S. S., Sabiyath Fatima, N.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.04.2022
• Subjects: Computational modeling; Convolutional neural networks; Data models; DBN-SVM; dynamical local-best harmony search; healthcare systems; hyperparameters; improved PCA; Medical services; Support vector machines; Training; unsupervised fuzzy-CNN
• DOI: 10.1109/ICESIC53714.2022.9783488

In healthcare systems, medicinal information is essential to classify and diagnose a variety of disorders at an early stage. Because of the use of Cloud-enabled Internet-of-Things (IoT) technologies, the classification of a vast amount of medicinal data has become more complex. To combat this issue, hybridized Deep Belief Network and Support Vector Machine (DBN-SVM) classifier was designed. However, its computation cost was high while initialization because the hyperparameters were not optimized. Hence this article proposes a Parameter Tuned Unsupervised Fuzzy Convolutional Neural Network (PTU-FCNN) model to classify the clinical data. First, the medical data is collected and pre-processed using a filtering scheme to obtain a clean database. Then, an Improved Principal Component Analysis (IPCA) is applied to extract the relevant attributes, which are fed to the PTU-FCNN classifier. In this PTU-FCNN classification model, a Dynamical Local-Best Harmony Search (DLB-HS) scheme is adopted to fine-tune the hyperparameters used in the U-FCNN model. This DLB-HS considers the different hyperparameters to be fine-tuned as the harmony and creates the Harmony Memory (HM) once the harmony is produced. Then, the HM is changed according to the training loss. Thus, the attributes are classified by this PTU-FCNN model to diagnose the diseases properly. At last, the testing outcomes reveal that the PTU-FCNN on the IVF dataset achieves 98.7 % of accuracy than the other conventional classifier models.