Session PB3.2

Automated Identification of Abnormal Fetuses Using Fetal ECG and Doppler Ultrasound Signals

AH Khandoker*, Y Kimura, M Palaniswami

The University of Melbourne
Melbourne, Australia

In this study, we propose an automated algorithm (support vector machines, SVM) to recognize the abnormal fetus using the timings of fetal cardiac events on the basis of analysis of simultaneously recorded fetal ECG (FECG) and Doppler ultrasound (DUS) signal. FECG and DUS signals from 23 fetuses [18 normal (N) and 5 abnormal (AB)] were analyzed. Multiresolution wavelet analysis was used to link the frequency contents of the Doppler signals with the opening(o) and closing(c) of the heart's valves [Aortic (A) and mitral(M)]. Five types of feature, namely 1) R-R intervals (msec) [N: 421±33, AB: 413±22], 2) time intervals from R-wave of QRS complex of FECG to opening and closing of aortic valve, i.e. R-Ao (msec) [N (mean±SD) :64.3±16.4 AB:67±21.4] and 3) R-Ac (msec) [N:209.7±12.3, AB: 208.5±12.9] respectively, 4) for the mitral valve R-Mc (msec) [N:16.7±8.4, AB: 18.5±7.4] and 5) R-Mo (msec) [N:277.6±20.3, AB: 278±25.8] were extracted from 60 beats and used as inputs to the SVM. Three kernel functions were tested. Using leave-one-fetus out cross validation technique, an SVM with polynomial kernel (d=2, C=10) correctly recognized 5 (five) abnormal (heart anomalies) fetuses out of 23 fetuses with only one false positive. The best feature subset was found to be of 3 features (R-R, R-Mc, R-Ao). We speculate that due to differences in control mechanism for heart contraction in abnormal fetuses from that in healthy ones, correlations among three features are different. These results could be useful in diagnosing fetal heart failure.

(Abstract Control Number: 56)