Jennkaie Lain

PPG is one of the most widely utilized physiological signals, demonstrating proven efficacy in assessing cardiovascular health. By leveraging low-cost finger PPG measurements, our CAD recognition model based on CNN AE deep feature representations achieved a mean recall of 96.65% in detecting CAD confirmed via cardiac catheterization. This novel classification technique significantly outperforms traditional time-domain feature-based methods and surpasses more sophisticated approaches in the CAD diagnosis literature, including those utilizing pre-trained networks for time-series image feature extraction. The promising results suggest that this CAD identification model has strong potential to become a viable tool for large-scale screening and early detection of cardiovascular conditions.

Chen-Wei Lee, Jenn-Kaie Lain, and Tin-Kwang Lin, “An automated coronary artery disease identification using photoplethysmography signals with deep feature representations,” under review.

About Me

Jenn-Kaie Lain is currently a Professor and Associate Chair of Dept. of Electronic Engineering at National Yunlin University of Science and Technology in Douliu City, Yunlin, Taiwan. He has more than twenty years of experience working in academic positions. His current interests are applications of machine (deep) learning including, biomedical signal processing, and optical wireless communications.

This study successfully developed a real-time portable system for glaucoma screening based on a deep convolutional neural network (CNN). Using an object detection algorithm, we first identified the optic nerve head (ONH), a critical region for glaucoma diagnosis. The YOLOv8 algorithm demonstrated exceptional performance in object detection, achieving a precision of 100%, recall of 99.4%, and mean average precision (mAP) of 99.5%. These results enabled precise localization of the ONH for subsequent classification tasks.

For classification, we trained multiple models, including EfficientNet B0, MobileNet V3, a lightweight VGG model enhanced with a squeeze-and-excitation (SE) module, and a Siamese network for few-shot learning (2-way k-shot). Among these, the Siamese network with 2-way 5-shot learning achieved the best performance, with a precision of 98.15%, recall of 96.85%, and an F1 score of 97.50%. Subsequently, the trained ONH detection model and the optimal 2-way 5-shot Siamese classification model were converted to the TensorFlow Lite (TFLite) format. We developed a mobile application in Android Studio, enabling the models to run on smartphones. This portable system facilitates cost-effective glaucoma screening in underdeveloped regions or remote areas, leveraging the widespread availability of smartphones to increase screening accessibility and reduce costs.

Guo-Kai Tu and Jenn-Kaie Lain, "Research on Real-Time Portable Glaucoma Screening with Deep Convolutional Neural Networks," draft manuscript.

In contrast to conventional architectures that sequentially combine feature extraction networks with regression networks, this study adopts a novel two-stage U-Net architecture. Each stage simultaneously performs feature learning and arterial blood pressure (ABP) waveform estimation from input signals, characterized by a progressive enhancement of ABP estimation performance. As highlighted in the literature, the first-stage network functions as an approximation network, while the second-stage network serves as a refinement network, iteratively improving and completing the blood pressure estimation process.

To achieve this, three key strategies were employed: (1) spatial feature alignment using additive attention, (2) channel attention via the squeeze-and-excitation (SE) module, and (3) preprocessing of PPG signals with VPG-assisted signal enhancement. These techniques collectively improved the performance of the blood pressure estimation model, successfully meeting the requirements of both the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS), as demonstrated in the results (refer to the screenshot).

Jenn-Kaie Lain, Jun-Hao Sheu, and Chon-Gan Kuo, "Cuffless blood pressure estimation from photoplethysmogram based Stacked U-Net," draft manuscript.

PUBLICATIONS

Google bots are awesome and they have compiled the complete list, accessible at Google Scholar page.