A long-standing problem at the frontier of biomechanical studies is to develop fast methods capable of estimating material properties from clinical data. In this paper, we have studied three surrogate models based on machine learning (ML) methods for fast parameter estimation of left ventricular (LV) myocardium. We use three ML methods named K-nearest neighbour (KNN), XGBoost and multi-layer perceptron (MLP) to emulate the relationships between pressure and volume strains during the diastolic filling. Firstly, to train the surrogate models, a forward finite-element simulator of LV diastolic filling is used. Then the training data are projected in a low-dimensional parametrized space. Next, three ML models are trained to learn the relationships of pressure–volume and pressure–strain. Finally, an inverse parameter estimation problem is formulated by using those trained surrogate models. Our results show that the three ML models can learn the relationships of pressure–volume and pressure–strain very well, and the parameter inference using the surrogate models can be carried out in minutes. Estimated parameters from both the XGBoost and MLP models have much less uncertainties compared with the KNN model. Our results further suggest that the XGBoost model is better for predicting the LV diastolic dynamics and estimating passive parameters than other two surrogate models. Further studies are warranted to investigate how XGBoost can be used for emulating cardiac pump function in a multi-physics and multi-scale framework.

Progress has been made enabling expensive, high-end inertial measurement units (IMUs) to be used as tracking sensors. However, the cost of these IMUs is prohibitive to their widespread use, and hence the potential of low-cost IMUs is investigated in this study. A wearable low-cost sensing system consisting of IMUs and ultrasound sensors was developed. Core to this system is an extended Kalman filter (EKF), which provides both zero-velocity updates (ZUPTs) and Heuristic Drift Reduction (HDR). The IMU data was combined with ultrasound range measurements to improve accuracy. When a map of the environment was available, a particle filter was used to impose constraints on the possible user motions. The system was therefore composed of three subsystems: IMUs, ultrasound sensors, and a particle filter. A Vicon motion capture system was used to provide ground truth information, enabling validation of the sensing system. Using only the IMU, the system showed loop misclosure errors of 1% with a maximum error of 4–5% during walking. The addition of the ultrasound sensors resulted in a 15% reduction in the total accumulated error. Lastly, the particle filter was capable of providing noticeable corrections, which could keep the tracking error below 2% after the first few steps.

No abstract

Smart wearable robotic system, such as exoskeleton assist device and powered lower limb prostheses can rapidly and accurately realize man–machine interaction through locomotion mode recognition system. However, previous locomotion mode recognition studies usually adopted more sensors for higher accuracy and effective intelligent algorithms to recognize multiple locomotion modes simultaneously. To reduce the burden of sensors on users and recognize more locomotion modes, we design a novel decision tree structure (DTS) based on using an improved backpropagation neural network (IBPNN) as judgment nodes named IBPNN-DTS, after analyzing the experimental locomotion mode data using the original values with a 200-ms time window for a single inertial measurement unit to hierarchically identify nine common locomotion modes (level walking at three kinds of speeds, ramp ascent/descent, stair ascent/descent, Sit, and Stand). In addition, we reduce the number of parameters in the IBPNN for structure optimization and adopted the artificial bee colony (ABC) algorithm to perform global search for initial weight and threshold value to eliminate system uncertainty because randomly generated initial values tend to result in a failure to converge or falling into local optima. Experimental results demonstrate that recognition accuracy of the IBPNN-DTS with ABC optimization (ABC-IBPNN-DTS) was up to 96.71% (97.29% for the IBPNN-DTS). Compared to IBPNN-DTS without optimization, the number of parameters in ABC-IBPNN-DTS shrank by 66% with only a 0.58% reduction in accuracy while the classification model kept high robustness.