We have presented a framework for reducing false alarms
in cardiac continuous monitoring using a wireless sensor.
Machine-learning-based classifiers are used for (i) classifying
arrhythmia heartbeat types and signal quality levels from
features extracted from ECG signals, and (ii) classifying
activity types from 3D acceleration signals. Taking signal
quality levels and activity types into consideration, a rule-based
expert system is used for determining whether abnormal heart
beats should trigger alarms or should be ignored. Signals from
two datasets, i.e., ECG signals from the MIT-BIH arrhythmia
database (DS1) and a collection of ECG signals and 3D
acceleration signals acquired through BSN nodes from 10
healthy human subjects when they performed activities of daily
living (DS2) were used for framework evaluation. The
experimental results showed that the overall accuracy for
arrhythmia classification on DS1 was 97.69%, with high
sensitivity, specificity, and selectivity. When evaluated on
DS2, our proposed framework yields the accuracy of 98.42%,
with the arrhythmia false alarm rate being reduced by wide