PCG classification through spectrogram using transfer learning. (January 2023)

• Record Type: Journal Article
• Authors: Ismail, Shahid; Ismail, Basit; Siddiqi, Imran; Akram, Usman
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 1
• Publication Date: 2023-01
• Subjects: PCG -- Spectrograms -- Convolutional neural networks -- Transfer learning -- Remote-interpretation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104075
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Heart rate classification is a challenging problem primarily due to spectral overlap of normal heart sound with internal sources like extra heart sounds, extra systole, murmurs, respiration sounds and external sources like body motion. In order to address this challenging problem, we have proposed a technique that relies on signal filtering, time segmentation, spectrogram generation, hybrid classification and finally a voting based mechanism. The proposed method carries out analysis at cycle as well as at signal level. Evaluation of the proposed technique on a challenging public dataset (PASCAL 2011) results in precision, recall and accuracy values of greater than 95% using 5-fold cross validation. Furthermore, the reported results also validate our claim that 2–3 s of data suffices for classification. Graphical abstract: Highlights: This study introduces and validates the postulate that multi-class PCG signal classification can be carried out from 2–3 s of data. An overlap of 0.1*fs retains all major peaks in the signal and can be utilized to classify the signals with rare events like extra systole; (fs : sampling frequency). The study shows that spectrum limited to 800 Hz is required for PCG signal classification. A hybrid classifier (CNN and SVM) complemented with a voting based system is used for cycle classification. It is shown that the training time can be significantly reduced by using pre-trained off-the-shelf models.