Objective : The heart sounds, recorded through a phonocardiogram (PCG) can assist in the early detection of valvular heart diseases (VHD), which is important for minimizing the chances of cardiac complications like heart failure and sudden cardiac death. Despite a large number of previous studies with highly accurate classifications, PCG-based deep learning (DL) methods for VHD detection are not suitable for clinical practice as they lack transparency and interpretability. Methods: The proposed work utilizes analytic continuous wavelet transform (CWT) scalograms as the time-frequency representations (TFR) of the PCG signals. A 2D convolutional neural network (CNN) is designed for the multiclass classification (Aortic stenosis, Mitral regurgitation, Mitral stenosis, Mitral valve prolapse, Normal) of PCG signal’s TFR. Besides introducing a VHD classification method in this paper, we also carry out the interpretation of the proposed CNN architecture by using occlusion maps and deep dream images for local and global explanations respectively. Results: The highest accuracy achieved during fivefold cross validation is 99.6% and the overall accuracy is 98.32% for a publicly available PCG database. The accuracy of the proposed method for binary classification (Abnormal, Normal), tested on the PhysioNet database is 93.07%. DL visualization methods assisted in determining what features or regions of TFR of PCGs the proposed network was “looking at” for making a class-specific correct prediction. Class-specific morphological and timefrequency features were observed. Conclusion: DL visualization makes the network decision more reliable. Significance: The novel classification framework along with its interpretations would enable successful clinical translation of PCG-based VHD detection.

An exoskeleton needs reference joint angle profiles at various speeds and slopes for its application in real-world scenarios. Recording these profiles and their implementation in the control system of an exoskeleton are time-consuming and complex processes. Therefore, there is a need for an artificial system that can predict subject-specific joint angle profiles for various real-world scenarios from a minimum amount of input data. This study aims to propose a predictive neural network (NN) called MoveNet, for joints angle profile prediction across variable walking speeds and slopes. MoveNet consists of three NN modules, namely, encoder, mapper, and decoder. The encoder module is trained to convert the knee joint angle profile into its 6-D latent representation. The mapper module is trained to map the latent representation of angle profile at zero degrees angle of inclination (AOI) to latent representation of angle profile at the target AOI. The decoder module is trained to predict an angle profile from its given latent representation. The proposed model successfully predicted knee joint angle profiles at three walking speeds (0.8, 1, and 1.2 m/s) and nine AOI ranging from −10° to 10°. MoveNet obtained root mean squared error of 3.24 ± 1.19° and mean absolute error of 2.66 ± 1.00° from tenfold cross validation. These results suggest that artificial intelligence can predict subject-specific knee joint angle profiles for variable slopes at a given walking speed from the data of knee joint angle profiles recorded at a flat surface.

This paper presents a new, simple, and efficient strategy for locomotion classification using a force myography (FMG) system developed in-house. The FMG signals were acquired from eight calibrated force sensors. These sensors were wrapped around the thigh and captured volumetric changes in thigh muscles during locomotion. The data were collected from eight able-bodied subjects and two unilateral transfemoral amputees while walking in different terrains, i.e., level walk, ramp ascent, ramp descent, stair ascent, and stair descent (SD). The repeatability of FMG signals within all the five locomotion modes was expressed in terms of variance ratio (VR). For all locomotion modes, except SD, the VR ranges from 0.1 to 0.3, showing high repeatability within the locomotion modes for able-bodied and amputee subjects. Furthermore, these locomotion modes are characterized by computationally simplified time-domain features of FMG signals and classified using linear discriminant analysis (LDA). The subject-dependent LDA model gave an average classification accuracy of 99.5% ( ${\text{SD}}\,= \,{\text{0.5}}$ ) and 96.1% ( ${\text{SD}}\,= \,{\text{1.9}}$ ) for able-bodied and amputee subjects, respectively. These results show the potential of the FMG system in a real-time classification of locomotion modes for lower-limb prosthetic control.