Scalable Personalization of Long-Term Physiological Monitoring: Active Learning Methodologies for Epileptic Seizure Onset Detection

Guha Balakrishnan, Zeeshan Syed
Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics, PMLR 22:73-81, 2012.

Abstract

Patient-specific algorithms to detect adverse clinical events during long-term physiological monitoring substantially improve performance relative to patient-nonspecific ones. However, these algorithms often rely on the availability of expert hand-labeled data for training, which severely restricts the scalability of personalized monitoring within a real-world setting. While active learning offers a natural framework to address this issue, the relative merits of different active learning methodologies have not been extensively studied in the setting of developing clinically useful detectors for infrequent time-series events. In this paper, we identify a core set of principles that are relative to the specific goal of personalized long-term physiological monitoring. We describe and compare different approaches for initialization, batch selection and termination within the active learning process. We position this work in the context of epileptic seizure onset detection. When evaluated on a database of scalp EEG recordings from 23 epileptic patients, we show that a combined distance- and diversity-based measure to determine the data to be queried, max-min clustering for identification of the initialization set, and a comparison of consecutive support vector sets to guide termination results in an active learning-based detector that can achieve similar performance to a patient-specific detector while requiring two orders of magnitude fewer labeled examples for training.

Cite this Paper

BibTeX
@InProceedings{pmlr-v22-balakrishnan12b, title = {Scalable Personalization of Long-Term Physiological Monitoring: Active Learning Methodologies for Epileptic Seizure Onset Detection}, author = {Balakrishnan, Guha and Syed, Zeeshan}, booktitle = {Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics}, pages = {73--81}, year = {2012}, editor = {Lawrence, Neil D. and Girolami, Mark}, volume = {22}, series = {Proceedings of Machine Learning Research}, address = {La Palma, Canary Islands}, month = {21--23 Apr}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v22/balakrishnan12b/balakrishnan12b.pdf}, url = {https://proceedings.mlr.press/v22/balakrishnan12b.html}, abstract = {Patient-specific algorithms to detect adverse clinical events during long-term physiological monitoring substantially improve performance relative to patient-nonspecific ones. However, these algorithms often rely on the availability of expert hand-labeled data for training, which severely restricts the scalability of personalized monitoring within a real-world setting. While active learning offers a natural framework to address this issue, the relative merits of different active learning methodologies have not been extensively studied in the setting of developing clinically useful detectors for infrequent time-series events. In this paper, we identify a core set of principles that are relative to the specific goal of personalized long-term physiological monitoring. We describe and compare different approaches for initialization, batch selection and termination within the active learning process. We position this work in the context of epileptic seizure onset detection. When evaluated on a database of scalp EEG recordings from 23 epileptic patients, we show that a combined distance- and diversity-based measure to determine the data to be queried, max-min clustering for identification of the initialization set, and a comparison of consecutive support vector sets to guide termination results in an active learning-based detector that can achieve similar performance to a patient-specific detector while requiring two orders of magnitude fewer labeled examples for training.} }
Endnote
%0 Conference Paper %T Scalable Personalization of Long-Term Physiological Monitoring: Active Learning Methodologies for Epileptic Seizure Onset Detection %A Guha Balakrishnan %A Zeeshan Syed %B Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2012 %E Neil D. Lawrence %E Mark Girolami %F pmlr-v22-balakrishnan12b %I PMLR %P 73--81 %U https://proceedings.mlr.press/v22/balakrishnan12b.html %V 22 %X Patient-specific algorithms to detect adverse clinical events during long-term physiological monitoring substantially improve performance relative to patient-nonspecific ones. However, these algorithms often rely on the availability of expert hand-labeled data for training, which severely restricts the scalability of personalized monitoring within a real-world setting. While active learning offers a natural framework to address this issue, the relative merits of different active learning methodologies have not been extensively studied in the setting of developing clinically useful detectors for infrequent time-series events. In this paper, we identify a core set of principles that are relative to the specific goal of personalized long-term physiological monitoring. We describe and compare different approaches for initialization, batch selection and termination within the active learning process. We position this work in the context of epileptic seizure onset detection. When evaluated on a database of scalp EEG recordings from 23 epileptic patients, we show that a combined distance- and diversity-based measure to determine the data to be queried, max-min clustering for identification of the initialization set, and a comparison of consecutive support vector sets to guide termination results in an active learning-based detector that can achieve similar performance to a patient-specific detector while requiring two orders of magnitude fewer labeled examples for training.
RIS
TY - CPAPER TI - Scalable Personalization of Long-Term Physiological Monitoring: Active Learning Methodologies for Epileptic Seizure Onset Detection AU - Guha Balakrishnan AU - Zeeshan Syed BT - Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics DA - 2012/03/21 ED - Neil D. Lawrence ED - Mark Girolami ID - pmlr-v22-balakrishnan12b PB - PMLR DP - Proceedings of Machine Learning Research VL - 22 SP - 73 EP - 81 L1 - http://proceedings.mlr.press/v22/balakrishnan12b/balakrishnan12b.pdf UR - https://proceedings.mlr.press/v22/balakrishnan12b.html AB - Patient-specific algorithms to detect adverse clinical events during long-term physiological monitoring substantially improve performance relative to patient-nonspecific ones. However, these algorithms often rely on the availability of expert hand-labeled data for training, which severely restricts the scalability of personalized monitoring within a real-world setting. While active learning offers a natural framework to address this issue, the relative merits of different active learning methodologies have not been extensively studied in the setting of developing clinically useful detectors for infrequent time-series events. In this paper, we identify a core set of principles that are relative to the specific goal of personalized long-term physiological monitoring. We describe and compare different approaches for initialization, batch selection and termination within the active learning process. We position this work in the context of epileptic seizure onset detection. When evaluated on a database of scalp EEG recordings from 23 epileptic patients, we show that a combined distance- and diversity-based measure to determine the data to be queried, max-min clustering for identification of the initialization set, and a comparison of consecutive support vector sets to guide termination results in an active learning-based detector that can achieve similar performance to a patient-specific detector while requiring two orders of magnitude fewer labeled examples for training. ER -
APA
Balakrishnan, G. & Syed, Z.. (2012). Scalable Personalization of Long-Term Physiological Monitoring: Active Learning Methodologies for Epileptic Seizure Onset Detection. Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 22:73-81 Available from https://proceedings.mlr.press/v22/balakrishnan12b.html.

Related Material