Effect of Decision Boundary for Logistic Regression Classifiers on Sleep Apnea Screening Accuracy with Wearable SpO2 Data
- Resource Type
- Conference
- Authors
- Liang, Zilu
- Source
- 2023 Fourteenth International Conference on Mobile Computing and Ubiquitous Network (ICMU) Mobile Computing and Ubiquitous Network (ICMU), 2023 Fourteenth International Conference on. :1-4 Nov, 2023
- Subject
- Communication, Networking and Broadcast Technologies
Signal Processing and Analysis
Measurement
Logistic regression
Databases
Sensitivity and specificity
Ubiquitous computing
Sleep apnea
Mobile computing
ubiquitous computing
sleep apnea
SpO2
logistic regression
imbalanced classification
- Language
Sleep apnea is a common sleep disorder characterized by short and frequent stop of breath during sleep. Many sleep apnea patients remain undiagnosed due to a lack of easily accessible screening method. While a considerable number of studies have intended to develop easy sleep apnea screening methods using public sleep datasets, the issue of the class imbalance presented in those datasets was rarely investigated. In this study, we developed logistic regression models for sleep apnea screening using one of the largest public sleep datasets-the SHHS database and examined how the model performance was affected by the logistic regression threshold for the original imbalanced dataset and for the resampled balanced dataset. We found a reverse $V$ -shape relationship between the overall model performance and the logistic regression threshold for both cases. The AUC peaked between 0.40-0.55 for the class-balanced model and between 0.75-0.85 for the non-balanced model. Meanwhile, MCC peaked between 0.30-0.45 for the balanced model and between 0.65-0.80 for the non-balanced model. The results are alerting given that many logistic regression classifiers use the default threshold 0.5. Our results imply that the decision boundary for logistic regression needs to be carefully adjusted in order to achieve the best model performance.