Research

Connectivity is a popular keyword in technologically advanced societies. This public interest reflects the increased effects of computerized devices in human life. With this development, interface technology connecting between humans and computerized devices is an important application for the connectivity. Especially, the interface technology using human gestures to communicate with the devices is quite remarkable. In this manner, myoelectric interface is a methodology to connect between a human and devices by recognizing human’s gestures. There are numerous HCI applications with the myoelectric interface, i.e., controlling unmanned ground/aerial vehicles and interfacing with virtual reality and augmented reality.

In this study, I proposed a dexterous myoelectric interface using electromyogram (EMG) signals to control a 6-DOF robotic manipulator with a 1-DOF gripper via the orientation of the forearm, muscle force, and dynamic hand gestures. For reliable HCI applications, I developed an EMG-based gesture recognition technique improved in a robust way against limb position changes. Conventional techniques using machine learning approaches have shown a vulnerability when the pose of a human operator is changed from the position in which the system is trained. This is because EMG signals are highly non-stationary and depended on the status of musculoskeletal system. The proposed gesture recognition technique uses temporal information of gestures which may be relatively robust against the non-stationarity of EMG signals. To do this, an algorithm using dynamic time warping, correlation coefficient, and template matching based classifier was designed in this study to recognize dynamic gestures.

Hypertension is a major cardiovascular risk factor that is treatable. But hypertension detection and control rates are unacceptably low. Ubiquitous BP monitoring technology could improve hypertension management, but such technology has not been available. In fact, most existing non‐invasive BP monitoring techniques widely used today in healthcare and research (auscultation, oscillometry, volume clamping, and applanation tonometry) suffer from limited convenience due to the requirement for an occlusive cuff and trained operators. Hence, ultra‐convenient BP monitoring technology that can eliminate the use of an occlusive cuff and an operator has the potential to significantly improve hypertension management and control. For ubiquitous and convenient BP monitoring technology, recent effort to enable ultra-convenient BP monitoring has focused on the development of techniques to infer BP from a single wearable device. Among various approaches, ballistocardiogram (BCG), defined as the body movement induced by the heartbeat, is a promising modality for ultra-convenient BP monitoring due to its close association with cardiac functions, especially aortic BP.

From this study, wearable devices that can be worn on limb locations (e.g., wrist or upper arm) were studied for the instrumentation of BCG. This wearable BCG is likely to improve convenience compared to previous successful studies, whole-body BCG measured by bulky devices such as scale and bed. Motivated by the above rationale, the objective of this study was to investigate the association between a limb BCG and BP based on data mining. During different BP-perturbing interventions, the BCG and reference BP were measured from a wristband equipped with an accelerometer and a commercial continuous BP measurement device. Then, both timing and amplitude features in the limb BCG waveform were extracted, and significant features predictive of BP were selected and compressed by applying various machine learning techniques. The association between the predictors obtained and BP was investigated by conducting multivariate regression analysis.

Critically ill patients (in an intensive care unit) suffering from circulatory shock receive hours or days of vasopressor infusion. These medications are vasoactive agents that elevates BP by elevating vascular tone, cardiac contractility, and heart rate. Considering that extended hypotension is closely associated with end-organ damage in critically-ill and intraoperative patient populations, hypotension must be effectively treated. Therefore, suboptimal blood pressure (BP) management is able to put these patients under suffering from preventable harm. It would be ideal if hypotension could be forecasted, even before onset, compliance to BP targets during vasopressor infusion may be improved, and accordingly, preventable end-organ damage could be avoided.

In this study, I investigated two algorithms forecasting hypotension of critically ill patients: a logistic regression model and an auto-regressive model. The logistic regression that was trained to forecast a specific level of hypotension, i.e., <60 mmHg was developed. However, this has a substantial limitation, since different clinical protocols require different BP goal ranges. Moreover, the method was limited to predicting the likelihood of imminent hypotension development. It wasn’t able to forecast any additional details of the BP time series. To break through the limitations of it, the auto-regressive model, a more flexible way to forecast hypotension, was proposed. The auto-regressive model is based on first forecasting the future BP time-series, and subsequently, computing the probability that the predicted BP time-series will be below the goal range. The performance of two models was compared at two different levels of hypotension, <60 and <70 mmHg. In sum, it was confirmed that a single auto-regressive model could be used to predict hypotension for multiple different definitions of hypotension.

Exploiting hand gestures for non-verbal communication has extraordinary potential in HCI. A data glove is an apparatus widely used to recognize hand gestures. To improve the functionality of the data glove, a highly stretchable and reliable signal-to-noise ratio sensor is indispensable. To do this, the study focused on the development of soft silicone microchannel sensors using a Eutectic Gallium-Indium (EGaIn) liquid metal alloy and a hand gesture recognition system via the proposed data glove using the soft sensor. The EGaIn-silicone sensor was uniquely designed to include two sensing channels to monitor the finger joint movements and to facilitate the EGaIn alloy injection into the meander-type microchannels. We recruited 15 participants to collect hand gesture dataset investigating 12 static hand gestures. The dataset was exploited to estimate the performance of the proposed data glove in hand gesture recognition. Additionally, six traditional classification algorithms were studied. From the results, a random forest shows the highest classification accuracy of 97.3% and a linear discriminant analysis shows the lowest accuracy of 87.4%. The non-linearity of the proposed sensor deteriorated the accuracy of LDA, however, the other classifiers adequately overcame it and performed high accuracies (>90%).