Intelligent alarms using principal component analysis on hemodynamic monitoring (Preprint)

BACKGROUND Patient monitors in modern hospitals give heartbeat waveform data that is reduced to aggregated variables and simple thresholds for alarms. Often, the monitors give a steady stream of non-specific alarms, leading to alarm fatigue in clinicians. An alarm can be seen as a classification problem, and by applying Principal Component Analysis (PCA) to the heart rate waveform of readily available monitoring devices, the accuracy of the classification of abnormality could be highly increased. OBJECTIVE To investigate whether physiological changes could be detected by looking at the heart rate waveform. METHODS A dataset of a healthy volunteer monitored with electrocardiography (ECG) and invasive blood pressure (BP) experiencing several tilts on a tilting table was investigated. A novel way of splitting continuous data based on the heartbeat was introduced. PCA was applied to classify the heartbeats. RESULTS A classification using only the aggregated variables heart rate (HR) and BP was able to correctly identify 20.7% of the heartbeats in the vertical tilt as abnormal. A classification using the full waveforms and combining the ECG and BP signals was able to correctly identify 83.5% of the heartbeats in the vertical tilt as abnormal. A humanistic machine learning (ML) method is then proposed based on the PCA classification. CONCLUSIONS A ML method for classification of physiological variability was described. The main novelty lies in the splitting of an ECG and BP signal by the heart rate and performing a PCA on the data-table.