Towards real-time non-invasive detection of hyperlipidemia through finger pulse image analysis using deep learning.

Hyperlipidemia detection involves invasive, time-consuming procedures requiring clinical laboratories and blood samples. Often asymptomatic in its early stages, hyperlipidemia significantly increases the risk of cardiovascular diseases. The objective of this study was to investigate whether hyperlipidemia produces detectable changes in pulse wave patterns and to develop a non-invasive, cost-effective diagnostic approach using deep learning techniques applied to finger pulse images. Pulse waves were recorded from 81 hyperlipidemia patients and 65 participants in the control group, with 700 single pulse wave cycles selected from each group. These waveforms were preprocessed and divided into training (70%), validation (15%), and testing (15%) subsets. Custom Convolutional Neural Network (CNN) architectures trained from scratch were developed and evaluated to identify the most effective classification model. After model selection, hyperparameter tuning was applied to enhance predictive performance. In parallel, pre-trained models such as Visual Geometry Group 16 (VGG16) were fine-tuned and optimized. The models were assessed using accuracy, precision, recall, and F1-score. The custom CNN models achieved the highest performance, with the top model reaching approximately 95%-96% for accuracy, precision, recall, and F1-score. The VGG16 models also performed well, with all metrics around 91%. Training and validation curves for both model types indicated strong learning capabilities with minimal overfitting or underfitting, showcasing their potential for generalization to unseen data. Deep learning models effectively differentiated pulse waves between individuals with hyperlipidemia and those in the control group, indicating that hyperlipidemia causes detectable changes in pulse wave patterns. This study could lead to the development of a reliable, efficient, and non-invasive device for hyperlipidemia screening.