Date Available

12-9-2021

Year of Publication

2021

Degree Name

Master of Science in Electrical Engineering (MSEE)

Document Type

Master's Thesis

College

Engineering

Department/School/Program

Electrical and Computer Engineering

First Advisor

Dr. Kevin D. Donohue

Abstract

Sleep has a significant impact on cognitive abilities such as memory, reaction time, productivity, and creative thinking; however, there are many aspects of this important activity that are not clearly understood. Over the last century, researchers have developed technology and animal models to assist in the study of sleep. Manual sleep scoring is time consuming, reduces productivity, and is impacted by human scorer subjectivity. On the other hand, automatic sleep stage categorization can enhance consistency and reliability, aiding professionals in identifying sleep related health problems.

In recent times various studies reported significant achievements for automatic vigilance detection and overcome the drawback of REM stage detection. Two models that reported very good performance are SCOPRISM and UTSN-L that replicate the manual scoring criteria. In this study, the performance of these models is documented on an independent dataset. The same dataset is also employed in feature-based machine learning approaches, where features from EEG and EMG signals are incorporated to the scoring process and NB, LDA, DT, KNN, SVM and RF models are assessed to do a comparative study on the same feature set.

Results show that, the random forest model achieves the highest overall accuracy of 84.7%, while the SCOPRISM and UTSN-L models achieve 76.1% and 77.1% respectively. When evaluated on an animal-by-animal basis, this RF model exhibits a reduced standard deviation with higher accuracy. However, despite the fact that the random forest model performs better than SCOPRISM and UTSN-L, it lacks REM sensitivity and still exhibits lower classification performance for genetically engineered mice of higher age groups. Animal-wise feature normalization is carried out, which resulted in findings that outperform all prior outcomes and reported the best result for vigilance stage detection with an overall accuracy of 90.8% and a REM sensitivity of 90%. The animal-wise evaluation also shows, this approach exhibits a more robust performance over the set of test animals than prior models.

Furthermore, the algorithm trained on 28 animal datasets is applied to the recordings utilized in the UTSN-L model, and overall accuracy was found 40%, with a REM recall of 16.6%. This reinforces the issue that while the machine learning algorithm excels at detecting key patterns in the dataset, performance varies depending on the equipment employed in different environments.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2021.442

Codes.zip (11127 kB)

Share

COinS