Semi-supervised detection of intracranial pressure alarms using waveform dynamics

Patient monitoring systems in intensive care units (ICU) are usually set to trigger alarms when abnormal values are detected. Alarms are generated by threshold-crossing rules that lead to high false alarm rates. This is a recognized issue that causes alarm fatigue, waste of human resources, and increased patient risks. Recently developed smart alarm models require alarms to be validated by experts during the training phase. The manual annotation process involved is time-consuming and virtually impossible to achieve for the thousands of alarms recorded in the ICU every week. To tackle this problem, we investigate in this study if the use of semi-supervised learning methods, that can naturally integrate unlabeled data samples in the model, can be used to improve the accuracy of the alarm detection. As a proof of concept, the detection system is evaluated on intracranial pressure (ICP) signal alarms. Specific morphological and trending features are extracted from the ICP signal waveform to capture the dynamic of the signal prior to alarms. This study is based on a comprehensive dataset of 4791 manually labeled alarms recorded from 108 neurosurgical patients. A comparative analysis is provided between kernel spectral regression (SR-KDA) and support vector machine (SVM) both modified for the semi-supervised setting. Results obtained during the experimental evaluations indicate that the two models can significantly reduce false alarms using unlabeled samples; especially in the presence of a restrained number of labeled examples. At a true alarm recognition rate of 99%, the false alarm reduction rates improved from 9% (supervised) to 27% (semi-supervised) for SR-KDA, and from 3% (supervised) to 16% (semi-supervised) for SVM.