Session P73.1

Distant Prediction of Paroxysmal Atrial Fibrillation Using HRV Data Analysis

YV Chesnokov*

Kuban State University
Krasnodar, Russia

Paroxysmal atrial fibrillation (PAF) is the most common abnormal heart rhythm encountered in clinical practice, and has serious associated morbidity and mortality as a sudden stroke. Distant prediction of PAF onset can help to avoid it using pacing techniques.
The data for analysis was taken from Physionet databases. We used atrial fibrillation prediction database (AFPDB), consisted of 30 minute ECG segments from patients immediately before PAF onset and distant ones, and MIT-BIH AF database (AFDB), consisted of 10 hour recordings with PAF and non-PAF heart rhythms. We annotated each ECG record using our developed algorithm [Y.V. Chesnokov et al, Individually Adaptable Automatic QT Detector. In Proceedings of Computers Cardiol: 17-20 September 2006, Valencia.] and extracted HRV data. We used entire 30 minute segment for analysis from AFPDB database. 10 hour AFDB database records were divided to consecutive overlapping 30 minute segments. Obtained HRV segments were interpolated to 2Hz and analyzed with power spectral density (PSD) estimation in the 0.0 – 0.5Hz frequency range. In order to obtain automatic classification we applied artificial neural networks (ANN). First we trained ANN on the HRV data from AFPDB database to distinguish between immediate and distant PAF segments. Then we applied that ANN model to AFDB database on the non-PAF HRV data before PAF onset (ranging from 50 minutes to 7 hours in duration) from 16 patients.
PSD analysis of the 30 minute HRV records from AFPDB database immediately before PAF onset provided statistically significant (p<0.001, KS-test) increase in the energy for entire spectrum range compared to records distant from PAF. This data was split to 50% for training and 50% for validation and testing using ANN classifier. Results of classification in terms of sensitivity (Se) and specificity (Sp) for train set (Se:90.4%, Sp:100%), validation (Se:90.4%, Sp:93.7%) and test set (Se:72.7%, Sp:88.2%). Trained on AFPDB database ANN classifier was further used for testing of distant prediction of PAF onset on AFDB database. The results we obtained on 14 patients using ANN classifier are promising, as the classifier marked 30 minute segments before PAF onset as positives leading to PAF 75 +- 43 minutes in advance, and the rest of distant segments were classified as negatives, not leading to PAF.

(Abstract Control Number: 43)