A 13.34 μW Event-Driven Patient-Specific ANN Cardiac Arrhythmia Classifier for Wearable ECG Sensors

Event-Driven ECG Classification Using an Open-Source, LC-ADC Based Non-Uniformly Sampled Dataset

“…1) MDL1: The LC-ADC model 1, hereby named MDL1, has been taken from the LC-ADC design in [11]. In this model, the MIT-BIH Arrhythmia database is converted to event-driven ECG using the LSB definition in (1) multiplied by a factor of 2.…”

“…2 for patient-specific (PS) The output layer classifies each beat into one of the following 4 categories: normal beat (N), supraventricular ectopic beats (S), ventricular ectopic beats (V), and fusion beats (F). The prominent morphological features of an ECG beat is contained in an 800ms window [11] centered around the QRS peak as shown in Fig. 1(b).…”

“…The pre-RR and post-RR intervals for each beat are used as the last two features. The conditional grouping scheme defined in [11] is applied in this design, which excludes 14 records from evaluation, in addition to the 4 paced records. To handle the inherent class imbalance in the dataset, the output layer uses a weighted cross-entropy…”

“…The MIT-BIH database is sampled at 360Hz using an 11-bit ADC. The accuracy (ACC), sensitivity (SE), positive predictivity (+P), and false positive rate (FPR) are used to assess the classifier performance as in [11]. The average results overall beat types (N, S, V, F) using each LC-ADC model and MDL0 are shown in Table II.…”

“…LC-ADCs have also been used in low-power QRS-detection of ECG signals [9] [10]. More recently, clinically acceptable arrhythmia classification has been achieved using event-driven data from LC-ADCs [11]. To the best of our knowledge, performance analysis of various LC-ADC designs for eventdriven ECG classification has not yet been done.…”