University of Twente Student Theses
Change Point Detection for Physiological Monitoring
Tjepkema, W.W.A. (2024) Change Point Detection for Physiological Monitoring.
Full text not available from this repository.
| Full Text Status: | Access to this publication is restricted |
| Embargo date: | 26 September 2026 |
| Abstract: | Continuous patient monitoring using wearable devices may improve patient outcomes by enabling earlier detection of patient deterioration in general wards. This research explores the application of an online unsupervised multivariate change point detection model on continuously measured vital sign data with the Checkpoint Cardio wearable at the UMC Utrecht. A comprehensive categorization of unsupervised change point detection methods was conducted, leading to the selection of an adaptive LSTM-Autoencoder Change-Point Detection (ALACPD) model for validation on continuously measured vital sign data. A mixed methods study was designed to validate the change point detection model. Ground truth change points were identified by medical service center staff through a custom-developed annotation interface, and inter-rater reliability was assessed using the Fleiss kappa statistic. The model’s performance was validated by comparing detected change points to the established ground truth set, and precision, recall, F1-score, and covering score were calculated for three ALACPD model variants: AE-skipLSTM-AR, AE-skipLSTM, and AR. A group of 52 patients was selected for model validation, including eight patients with known events. The ALACPD model demonstrated moderate performance, achieving an F1-score of 0.51, with a recall of 0.7, a precision of 0.44, and a covering score of 0.51. The model successfully detected all predefined events. While the model’s average F1-score (0.51) was lower than its performance on other datasets reported in the literature (0.67), it outperformed other change point detection models (0.49) from previous studies tested on various datasets. The AE-skipLSTM-AR variant showed the highest performance in general. In line with the expectation, the AR model, along with the AE-skipLSTM-AR model, was able to detect an abrupt change point in the quality control series with a mean shift. Conversely, the AE-skipLSTM model detected the gradual change point in the quality control series with a shift in the trend, although with a delay beyond the margin of error. The annotators demonstrated consistency in identifying change points related to shifts in mean or variance. Nonetheless, variability was observed when identifying gradual changes, potentially contributing to a lower Fleiss kappa score. The ALACPD model holds promise for continuous vital sign monitoring in ward patients by showing the potential to reduce alarm fatigue while still capturing important events. Additionally, the model can offer valuable insights into the interpretation of trends and assist in prioritising monitored patients in the future. Future research should focus on automated hyperparameter tuning, enhancing the annotation process to improve the reliability of the ground truth set, and addressing missing data challenges to optimize the model’s performance and generalizability in real-world clinical settings. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 44 medicine, 50 technical science in general, 54 computer science |
| Programme: | Biomedical Engineering MSc (66226) |
| Link to this item: | https://purl.utwente.nl/essays/103741 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page