Themes
Signal Processing and Analysis of Physiological Data
I work on physiological data analysis by using waveform-based signal processing method and machine/deep learning approaches. I make efforts in processing the raw noisy signals and extracting the features that contribute the applications, including vital sign monitoring, disease diagnosis, and biometric identification. I have several papers published in IEEE EMBC, ICPR, CinC, CVPR.
Ultramarathon Data Exploration - Prediction of Acute Kidney Injury at Pre-race Stage
An acute kidney injury (AKI) prediction algorithm has been proposed for ultramarathon runners. In this study, blood, urine, and body composition data from the ultrarunners were collected. We process the data based on our medical expertise and leverage traditional support vector machine algorithm to accomplish AKI prediction. Our algorithm achieves an accuracy of 85% and sufficiently high sensitivity and specificity in the validation dataset. More importantly, we found out that high muscle mass, low fat mass, and healthy heart functioning prevent an ultrarunner from acquiring AKI after participating an ultramarathon run. Our work will be published at Journal of Translational Sports Medicine.
Continuous Motion Artifact Resilient Cuff-less Blood Pressure Monitoring
A cutting edge cuff-less blood pressure monitoring technology is proposed in this project. The monitoring requires only one 3-axis accelerometer. Signal processing and feature extraction techniques for blood pressure estimation have been developed to process noisy raw data. We are able to reconstruct BCG-like signals by exploiting three-dimensional accelerations and seek the informative features through the BCG-like waveforms. Our work was presented at 2019 EMBC and 2021 EMBC.
Early Prediction of Sepsis from Clinical Data (2019 PhysioNet Challenge)
A video classification neural network is leveraged to accomplish early sepsis prediction (see Figure below: the horizontal axis is the provided variables such as vital signs and lab measures; the vertical axis is the time-step). To tackle the issues of largely missing value and irregularly sampled data, we adopt a data-driven signal processing methodology to reconstruct the signals. Furthermore, we conducted the classification experiments on several baseline machine learning algorithms (SVM, XGBoost, Random Forest, Naive Bayes,...) and other neural network models. Our work was presented at 2019 Computing in Cardiology Conference (CinC) and will be published around March of 2020.
Dynamic Parameter Estimation of Brain Mechanisms
This is my research survey topic in my research examination. I reviewed various approaches of parameter estimation in the field of computational neuroscience. Given the variety of the tactics, I focused on the popular parameter estimation methods, including variational Bayesian, particle filtering, Markov Chain Monte Carlo, Gauss-Newton algorithms and constrained optimization. It is difficult to determine the best strategy because the data, or the subjects, alter from one experiment to another. I would suggest to have all the parameter estimation methods tested on an identical large dataset, which currently seems unavailable.
Here's the link to the review
Neuroimaging Data Analysis - Biomarkers Identification of Mild Traumatic Brain Injury
I collaborated with doctors from department of radiology and VA Healthcare on this project. I developed feature selection algorithm for diffusion MRI data and tweaked the conventional SVM algorithm to achieve biomarkers searching. The efficacy of the biomarkers found shows a 94% accuracy, 88% sensitivity and 90% specificity. Our work was presented at Society of Neuroscience Conference of the year 2018 and published in Diagnostics in 2022.
Signal Processing and Analysis of Electrocardiogram
The periodicity quantification problem is formulated in differential equations with boundary conditions. Experiments have been conducted on single and multiple ECG waveforms. My discovery shows that the rising of R wave is informative of the subject's heartbeat stability. I would like to show my gratefulness to Professor Chieh-Hsiung Kuan for his supervision. This project was funded by the Ministry of Science and Technology in Taiwan.