Session P74.1

ECG Signal Quantization Effects in the Analysis of Atrial Fibrillation

C Vayá*, JJ Rieta

Universidad de Castilla-La Mancha
Cuenca, Spain

During last years spectral analysis techniques applied to the ECG have contributed to the knowledge and management of atrial fibrillation (AF). In this matter, parameters from the atrial activity (AA) time-frequency domain have proved their utility for episode analysis and characterization. These techniques are mostly applied to high resolution 12-lead ECGs. Nevertheless, these high resolution recordings are not always available in the clinical practice and recordings from low resolution acquisition systems, e.g. Holter systems, must be used instead. In this work we analyze the effects of ECG quantization on the AA extraction quality carried out by one of the most recent techniques, independent component analysis (ICA), and the suitability of time-frequency parameters in low resolution recordings.
We analyzed a total of 18 ECG signals of sixteen seconds in length, sampling rate equal to 500 samples per second and digitized with 16 bits. Quantizations from 15 to 4 bits of the original signals were obtained and then the FastICA algorithm was used to extract the estimated AA from every signal. Chebyshev type II ninth order digital low pass filtering was applied to all the estimated AA signals with cut-off frequency 70Hz and 40dB ripple in the stop-band. Next, the spectrograms of the filtered AA signals were computed using 1024 points Fast Fourier Transform (FFT), 1024 sample size Hamming window and 75% overlapping. Besides, time-domain sequences of the AA spectrograms characteristic parameters, i.e. the mean peak frequency (fp) and the spectral concentration (SC), were constructed. Furthermore, the Sample Entropy (SampEn) of the estimated AA signals was also computed.
The cross-correlation (Rxy) between the extracted AA signals at different quantification levels and the AA signals obtained operating with 16 bit resolution was calculated as a measure of signal estimation quality. Results revealed that Rxy remained higher than 0.7 when the number of quantization bits was higher or equal to nine. For eight or seven bits, Rxy was 0.65, decreasing significantly when the number of bits was lower or equal to six. On the other hand, the Minimum Absolute Error (MAE) of fp, SC and SampEn was also computed. For fp, MAE did not surpassed 1.3 Hz when the number of bits was higher or equal to seven but it increased drastically below seven bits. In the same way, MAE of SC maintained fewer than 6% when at least seven quantification bits were used, but higher MAE values were obtained when a more reduced number of bits was considered. Finally, MAE denoted a good approximation of SampEn when seven or more bits were used but it worsens considerably in other case. As conclusion, this study revealed that eight bits can be considered as the minimum real resolution threshold than can provide acceptable results. However, taking into account the headroom margin that has to be considered to avoid recording overflow, the minimum recommended number of bits is twelve.

(Abstract Control Number: 61)