Publications

E-SMOTE: Entropy Based Minority Oversampling for Heart Failure and AIDS Clinical Trails Analysis

Authors: Anbu Valluvan, Sainath Veerla

Published in: 2024 IEEE 48th Annual Computers, Software, and Applications Conference (COMPSAC)

Date of Conference: 02-04 July 2024

Date Added to IEEE Xplore: 26 August 2024

DOI: 10.1109/COMPSAC61105.2024.00291

Publisher: IEEE

Conference Location: Osaka, Japan

Abstract

Machine Learning (ML) algorithms often exhibit reduced performance in the presence of class imbalance, leading to biased results favoring the majority class in a dataset. This imbalance can be addressed through various sampling techniques, including oversampling of the minority class, undersampling of the majority class, or a combination of both. However, these techniques utilize the entire set of samples of datasets. In this paper, we introduce E-SMOTE, an Entropy-based Synthetic Minority Oversampling Technique (SMOTE), which extends the traditional SMOTE method. E-SMOTE is a novel oversampling technique designed to utilize a subset of the dataset from the minority class for the oversampling process. We employ entropy as a guiding metric to identify influential minority class instances located near decision boundaries. By generating additional instances near these boundaries within a binary classification system, E-SMOTE strengthens the decision boundary during the ML classifier training process. We conducted experiments on two datasets, Heart Failure Records and AIDS Clinical Trail Records, to demonstrate the effectiveness of E-SMOTE compared to traditional SMOTE. Our experimental results illustrate that E-SMOTE outperforms baseline classifier for both Heart Failure and AIDS clinical trial datasets. Additionally, it provides reasonable and comparable performance using a subset of the datasets compared to SMOTE oversampling technique using the entire dataset.

E-SMOTE Results

Deep Learning for Non-Invasive Blood Pressure Monitoring: Model Performance and Quantization Trade-Offs

Authors: Anbu Valluvan Devadasan, Soumyabrata Sengupta, Mohammad Masum

Published in: Electronics (MDPI)

Volume: 14, Issue 7

Article Number: 1300

Publication Date: 2025

DOI: 10.3390/electronics14071300

Publisher: MDPI

Journal: Electronics (ISSN 2079-9292)

Abstract

This research explores the application of deep learning techniques for non-invasive blood pressure monitoring, focusing on the critical balance between model performance and computational efficiency through quantization strategies. The study investigates various deep learning architectures and their effectiveness in accurately predicting blood pressure measurements from non-invasive signals, while addressing the practical constraints of deploying these models in resource-limited healthcare environments. Through comprehensive analysis of quantization trade-offs, this work provides insights into optimizing deep learning models for real-world medical applications where both accuracy and computational efficiency are paramount.