HRV analysis of patients prone to atrial fibrillation using

1
HRV analysis of patients prone to atrial
fibrillation using a Neural Network approach.
Yuriy Chesnokov, Dmitry Nerukh, Robert Glen.
Simple and robust fully automated methods for the
screening and prediction of PAF events is of high
clinical importance for the detection of the most
frequent cardiac arrhythmias.
Method
Automatic ECG annotation
30 min segments
Training data normalization 5 layer Screening ANN
41 15 10 5 1
PSD spectrum 0.01 0.4 Hz
HRV extraction
Training data normalization 5 layer Prediction
ANN 41 15 10 5 1
Results
10 min distant PAF prediction on 20 min HRV
segment
PAF screening on single 30 min non-PAF HRV segment
PAF screening training data 45 prone to PAF
635 healthy
PAF screening results Test set 48 healthy
and PAF Se 74.0, Sp 61.9, Ac 68.7
PAF prediction training data 78 10 min
distant from PAF 1997 healthy and more than
45 min PAF distant
Normal HRV database testing Test set 6385
healthy 30 min segments screening ANN 83.3
accuracy prediction ANN 87.1 accuracy
Immediate PAF prediction on single 30 min HRV
segment
PAF prediction results Test set 175 healthy
and PAF preceding Se 67.3, Sp 64.2, Ac
65.1 with energy normalization
PAF prediction training data 21 immediately
before PAF 658 healthy and PAF distant
PAF prediction results Test set 175 healthy
and PAF preceding Se 69.3, Sp 68.2, Ac
68.5 with minmax normalization Se 81.6,
Sp 53.9, Ac 61.7 with sigmoidal
normalization
Conclusions 1) Differentiation of PAF patients
from healthy patients 2) Immediate
prediction of PAF 3)
Prediction of PAF 10 minutes in advance
This automated method produced good results with
high Se and Sp values providing the possibility
of implementing the method in portable devices
for people at risk of cardiac arrhythmia and also
the possible adaptation of the neural network for
individual patients.
Unilever Centre for Molecular Sciences
Informatics, University Chemical Laboratory,
Cambridge University Lensfield Road, Cambridge,
CB2 1EW, UK