In this paper, we present the design and implementation of an avatar-based interactive system that facilitates rehabilitation for people who have received total knee replacement surgeries. The system empowers patients to carry out exercises prescribed by a clinician at the home settings more effectively. Our system helps improve accountability for both patients and clinicians. The primary sensing modality is the Microsoft Kinect sensor, which is a depth camera that comes with a Software Development Kit (SDK). The SDK provides access to 3-dimensional skeleton joint positions to software developers, which significantly reduces the challenges in developing accurate motion tracking systems, especially for use at home. However, the Kinect sensor is not well-equipped to track foot orientation and its subtle movements. To overcome this issue, we augment the system with a commercial off-the-shelf Inertial Measurement Unit (IMU). The two sensing modalities are integrated where the Kinect serves as the primary sensing modality and the IMU is used for exercises where Kinect fails to produce accurate measurement. In this pilot study, we experiment with four rehabilitation exercises, namely, quad set, side-lying hip abduction, straight raise leg, and ankle pump. The Kinect is used to assess the first three exercises, and the IMU is used to assess the ankle pump exercise.

The Extreme Learning Machine (ELM) is an effective learning algorithm for a Single-Layer Feedforward Network (SLFN). It performs well in managing some problems due to its fast learning speed. However, in practical applications, its performance might be affected by the noise in the training data. To tackle the noise issue, we propose a novel heterogeneous ensemble of ELMs in this article. Specifically, the correntropy is used to achieve insensitive performance to outliers, while implementing Negative Correlation Learning (NCL) to enhance diversity among the ensemble. The proposed Heterogeneous Ensemble of ELMs (HE ${}^{\mathrm{2}}$LM) for classification has different ELM algorithms including the Regularized ELM (RELM), the Kernel ELM (KELM), and the $L_{\mathrm{2}}$-norm-optimized ELM (ELML2). The ensemble is constructed by training a randomly selected ELM classifier on a subset of the training data selected through random resampling. Then, the class label of unseen data is predicted using a maximum weighted sum approach. After splitting the training data into subsets, the proposed HE ${}^{\mathrm{2}}$LM is tested through classification and regression tasks on real-world benchmark datasets and synthetic datasets. Hence, the simulation results show that compared with other algorithms, our proposed method can achieve higher prediction accuracy, better generalization, and less sensitivity to outliers.