Liquan Guo

Automated Fugl-Meyer Assessment using SVR model

A simple, objective and quantitative unsupervised outcome measure is considered vital in the home-based rehabilitation for stroke patients. The Fugl-Meyer Assessment (FMA) scale is widely utilized in the clinical practice, while not suitable in the home settings due to its subjective and time-consuming property. In this paper, a Support Vector Regression (SVR) based evaluation model was presented to automatically estimate the FMA scores for Shoulder-Elbow movement. The estimation was obtained by analyzing accelerometer data recorded during the performance of 4 tasks from Shoulder-Elbow FMA. A combined feature selection method based on ReliefF-SVR was implemented to simplify the calculation and improve the model performance. Twenty-four subjects were involved in this study and results showed that it was possible to achieve accurate estimation of Shoulder-Elbow FMA scores using the proposed model and a cross-validation prediction error value of 2.1273 was achieved.

Upper limb motion recognition for unsupervised stroke rehabilitation based on Support Vector Machine

In order to monitor the rehabilitation training of stroke patients in unsupervised situation and provide rehabilitation advice for rehabilitation clinicians, a wireless upper limb motion recognition system has been developed using tilt sensors, to identify the complex upper limb movements such as flexion and extension of elbow, flexion of elbow and touch the head, from a stroke patients rehabilitation program. 18 different movements from a stroke patients rehabilitation training program were adopted to verify and validate this system with 12 of them in the training group and 6 of them in the testing group. After preprocessing and the feature extraction of the acquired motion data, the Support Vector Machine (SVM) recognition approach was employed to establish a small sample identification model. Finally, the data of testing group in the upper limb rehabilitation training program were used to identify the developed model. It has been found that the recognition accuracy from this developed model was 100%. This result provides a well reference for further development of an automated system for stroke patient rehabilitation motion recognition.

PID control of glucose concentration in subjects with type 1 diabetes based on a simplified model: An in silico trial

An artificial pancreas system (APS) mimics the function of a real pancreas through monitoring a diabetics blood glucose and administering the right dose of insulin via an automatic control loop. It is hailed as a promising cure of diabetes, though this technology is still years away from commercial use due to a few technological bottlenecks. The simulation model of insulin-glucose metabolism of type 1 diabetes mellitus (T1DM) is an essential part of APS. In order to simplify the parameter identification task so that the model can be implemented electronically with ease, this paper presents a simplified model based on Routh approximation model reduction method. The results show that the approximation error between the simplified model and the original model is so small that can be neglected. Based on the simplified model, a PID controller is designed to maintain normoglycemia (90 mg/dl) in subjects with T1DM. The in silico simulation results show that the glucose concentration is controlled well, the risk of hyperglycemia and hypoglycemia is reduced a lot. This suggests that the simplified model describes the insulin-glucose metabolism process accurately, and the PID control algorithm is well-suitable to guide the further development of an APS.

LVQ neural network applied for upper limb motion recognition for home-based stroke rehabilitation

To improve the rehabilitation effectiveness and reduce the hospital costs, a new upper limb motion recognition model, through which hospital based clinicians can remotely supervise home based stroke rehabilitation, is proposed in this paper. Firstly, the real time limb motion data is collected using a 3-axis accelerometer sensor which is fixed on the upper limb of a patient. Secondly, the Wavelet Transform is employed to extract the approximation coefficients of different types of rehabilitation motions. Finally, a recognition model is established based on an LVQ neural network. 2 typical rehabilitation motions, Bobath handshaking and wrist turning, were chosen to test this proposed recognition system. The experiment results indicate that the recognition accurate rate can achieve as high as 100%. This pilot forms a foundation to further develop a home based remote training and assessment system for stroke rehabilitation.

Effect of meal intake on the quality of empirical dynamic models for Type 1 Diabetes

A model-based controller for artificial pancreas requires a model that is able to predict future glucose trends precisely. To quantify the effect of meal intake on the quality of empirical dynamic models (EDM), changing meal conditions (e.g., the meal amounts and times variation, individual differences) were simulated to generate data. Both single-input single-output (SISO) and multi-input single-output (MISO) EDM were identified and evaluated via model identification technology. The prediction accuracy of these models varies significantly within a subject and between subjects due to the different variation of meal amounts, and the additional afternoon snack and meal times shift have the greatest influence on these models. The prediction accuracy of MISO models are worse than that of SISO models under the changing meal condition.