Frequency Domain Feature Statistics for Cardiovascular Disease Diagnosis on Low-Computational Devices

Towards providing a low cost and readily available healthcare intervention system for rural India, we propose novel methods to analyze ECG on Android mobile devices for diagnostic purposes. In the proposed system, ECG is compressed to reduce the volume of input so as to enable processing on a low computational mobile device. Morphological, wavelet and statistical features are extracted directly from the compressed ECG. The subset of most discriminative features is extracted using information theoretic Dynamic Weighting based Feature Selection. Hybrid classification using majority voting-based classifier fusion is applied to enhance the classification performance. For the 14 MIT Arrhythmia classes, the proposed system achieved classification accuracy of 99.3% and a sensitivity of 100 % and a specificity of 98.9 %. Comparison with other existing methods shows better performance of the proposed method.