Biomedical signal processing and machine learning for wearable data
Goal: The goal of this project was to predict health states using photoplethysmography (PPG) signals from users of a wearable device.
Method: I implemented an innovative method to compute non-conventional heart rate variability (HRV) metrics from PPG signals collected with a wearable device, and built a machine learning model to classify health conditions, achieving an accuracy of 80-90%.
Result: This work resulted in the discovery of a set of digital biomarkers with promising value for real-time health monitoring.
Author: Eric Casey, Ph.D.
ivneuro: a Python package for analyzing neurophysiological signals
ivneuro provides tools for analyzing neural signals recorded in-vivo during behavior. It focus on time series analyses of continuous variables such as Local Field Potentials and is optimized to process either single signals in a single condition as well as multiple signals in multiple conditions simultaneously. It also provides a sub-package for extracting data from Nex files.
Useful links: Package site, GitHub, Documentation
Author: Eric Casey, Ph.D.
Microscopy image processing and unsupervised machine learning for biomarker quantification
Te goal of this study was to evaluate the existence of sub-regions with different functional properties within a brain area. To achieve this goal, I designed a method consisting in making a profile of pharmacologically induced cell activity across the region of interest.
Expected results and interpretation: an homogeneuos profile would indicate that the whole region doens't show detectable topological variations under the tested conditions, while an heterogeneous profile would reveal sub-region with specific patterns of cell activity induced by different pharmachological agents.
Approach: 4 different drugs and vehicle (5 treatments in total) were administered to animal models and the pharmacological response was assessed by the number of active cells, detected through the visualization of a molecular biomarker using histological techniques. A Python algorithm for images processing that quantifies cells from fluorescence microscopy photos, and an unsupervised machine learning algorithm (Principal Component Analysis) followed by visualization of data clusters using 2D and 3D scatterplots, was developed. This approach resulted in clustering anatomical samples based on their cell activation after administration of specific drugs.
Results: 5 previously unknown anatomical areas were identified, each with unique anatomical and pharmacological properties. The findings of this work are part of a scientific research published in a journal at the top 25% of impact.
Author: Eric Casey, Ph.D.
