This study investigated dynamic structural changes and their implications for dynamic stability during Sit-to-Stand (STS) transitions with healthcare robot assistance. The STS movement, a fundamental activity for mobility and independence, poses challenges for older adults due to age-related declines in muscle strength and joint integrity, increasing fall risk. Assistive robots, typically categorized into upper- and lower-limb support systems, have been developed to mitigate this risk. However, conventional biomechanical assessments fail to capture the nonlinear, time-varying coordination patterns essential for maintaining dynamic stability in Human–Robot Interaction (HRI). To address this gap, we introduced nonlinear analytical techniques to detect dynamic structural changes during STS transitions under three experimental conditions: self-performed, passive robot-assisted, and proactive robot-assisted STS. Motion capture data from 14 healthy participants were analyzed, with segment synergies quantified through relative phase analysis of the trunk, knee, and ankle. A segment-synergy-based segmentation algorithm based on dynamic correlation exponents was developed to detect phase transitions in synergy dynamics. Complementary nonlinear-dynamics measures—maximum Lyapunov exponent, sample entropy, and detrended fluctuation analysis—were employed to assess dynamic stability. Results showed that passive assistance induced earlier but less synchronized transitions than self-performed STS, whereas proactive assistance preserved natural timing and enhanced intersegmental coordination stability. Temporal dissociation between knee–trunk and ankle–knee synergy transitions emerged as a potential predictor of dynamic stability outcomes. These findings underscore the importance of phase-sensitive control algorithms in assistive robotics for achieving biologically congruent HRI. By bridging biomechanical insights with advanced analytical methods, this study provides a critical step toward the design of healthcare robots that enhance mobility and reduce fall risk.

This study proposes a domain-adaptive fine-tuning method for transferring emotion estimation models trained on video viewing tasks to food consumption tasks. Only the electrocardiogram signal was used as the input, and the model estimated the valence and arousal based on Russell’s circumplex model of affect. In the evaluation experiment, both biosignals and subjective ratings were collected from nine participants during both video viewing and food consumption tasks, and the model performance was compared under different fine-tuning situations. The results demonstrated that fine-tuning specifically tailored to the food consumption task improved the estimation accuracy, indicating that effective model adaptation is feasible even with a limited amount of data. Furthermore, fine-tuning using the video-viewing task showed a certain degree of effectiveness, indicating the potential for cross-domain transferability and mitigation of individual differences.

Pattern recognition methods have been researched in the welfare and industrial fields. Embedding a pattern classifier into a system requires rapid classification and miniaturization of the calculator. Field-programmable gate array (FPGA)-implemented classifiers are useful because they can achieve these goals. However, conventional pattern recognition methods misclassify unknown data that are not assumed during learning into one of the learned patterns. To address this issue, a pattern recognition method that considers unknown data (called the open-set recognition (OSR) method) can be implemented. In this study, we propose a novel OSR method suitable for FPGA implementation and utilize our method for interface control on FPGA as an application in the welfare field. Our method uses a probabilistic neural network that incorporates two types of approximated Gaussian distributions that do not involve exponential operations in the calculation. This characteristic contributes to computational resource reduction and fast classification. The computation of our method can be executed in parallel on an FPGA. In the experiments, we conducted a performance evaluation of our method for a forearm motion estimation task. The results confirmed that our approach enables faster and safer control of myoelectric prosthetic hands on an FPGA.

In the development of myoelectric prosthetic hands for reliable determination of the user's intended motion, openset recognition methods enabling consideration of unexpected inputs help to prevent malfunction. Conversely, some abnormal inputs may encompass new motions that can be used for motion classifier evolution and automatic acquisition of anomaly cluster structures. Against such a background, this paper outlines advanced open-set recognition involving clustering of unknown classes for enhanced ease of prosthetic hand interface usage. The proposed approach involves open-set recognition using the authors' probabilistic neural network, a novel cluster detection method employing a non-parametric Bayesian model, and structural modification. Results from forearm motion recognition using electromyogram signals demonstrated the effectiveness of the technique.

We present a Go/NoGo 3D game equipped with an eye tracker that records subjects' responses and his gaze position on the monitor over time. The proposed system consists of two functions: training that allows an instructor to modify the game's parameters and make a customized test; and evaluation in which the instructor can fix the parameters to create a standardized test. During the experiment, subjects were required to respond only to Go character by pressing a spacebar. The experimental results from 59 participants demonstrated that one's response time and its variability correlated with one's gaze behavior. Subjects with higher gaze modulation tended to respond faster and more stable. We also observed that utilizing the proposed system we could monitor the improvements in an Autism Spectrum Disorder child during his rehabilitation: his gaze modulation increased and his response time became more steady. In brief, utilizing the proposed system, we could effectively measure participants' response time variability of NoGo errors and their gaze trajectory area, which previous studies found to have a strong relationship with symptoms of mental disorders.

An increasing frequency of fall accidents associated with demographic aging in Japan has given rise to a need for effective fall prevention methods. Elderly people often use support devices such as canes and walking frames to reduce the risk of falling, however such solutions may be inappropriate for certain environments. Previous research has shown that unsteadiness can be mitigated via light touch contact (LTC) with a force of up to around 1 N with curtains or similar (Jeka, 1994), and the authors also previously proposed a virtual light touch contact (VLTC) approach based on LTC (Shima et al., 2013). VLTC supports standing stability based on a surrounding virtual partition connected to a vibrotactile fingertip stimulator. Here, it is known that the VLTC effect is not achieved via simple constant fingertip stimulation. Thus, vibrotactile stimulation is in VLTC needs to be controlled based on fingertip motion characteristics such as acceleration. However, LTC effect can be achieved via constant contact fingertip with a piece of paper or similar without fingertip movement. Assuming that reaction force from a fixed point fluctuates with body sway or psychological tremors in LTC, the LTC effect may be achievable by reproducing such fluctuations via vibrational stimulation. In this study, the authors proposed a novel VLTC method involving the use of vibration stimulation control to reproduce fluctuations in contact reaction force caused by the individual's movement based on fingertip acceleration data. Verification of the method indicated the proposed method can reduce body sway and reproduce the LTC effect. This suggests that reduction may be associated with slight improving fingertip positional sense.

Reduced attentional capacity caused by increased mental workload (MWL) among drivers can cause traffic accidents. Against this background, a method for effective MWL evaluation is required. The proposed method is used to map multidimensional biological signals into a probabilistic space based on a mixed normal distribution model. The load state of the driver is evaluated, and the cognitive load can be determined from a posteriori probability. In the experiments reported here, tasks performed to examine tracking of the vehicle in front and N-Back evaluation helped to clarify the cognitive load of the three subjects, for whom multidimensional biosignal monitoring was performed. The results demonstrated a high correlation between evaluation and NASA-TLX values. In other experiments, increased evaluation values were observed in satnav operation. Accordingly, the proposed method is considered suitable for real-time driver MWL evaluation.

Most methods of multi-view human activity recognition can be classified as conventional computer vision approaches. Those approaches separate feature descriptor and discriminator. Hence, the feature extractor cannot learn from the mistakes made by the classifier. In this paper, a deep neural network (DNN) model for human activity estimation using multi-view sequences of raw images is presented. This approach incorporates features extractor and discriminator into a single model. The model comprises three parts, a convolutional neural network (CNN) block, MSLSTMRes, and a dense layer. This method enables discrimination of human activity such as “walk” and “sit down” by merely using sequences of raw images. Experimental results on IXMAS dataset using one-subject cross validation demonstrates high prediction rate that is comparable to other methods in the literature, which utilized preprocessed images such as silhouette and volumetric data and sophisticated feature extractor.

This paper proposes a probabilistic neural network (NN) developed on the basis of time-series discriminant component analysis (TSDCA) that can be used to classify high-dimensional time-series patterns. TSDCA involves the compression of high-dimensional time series into a lower dimensional space using a set of orthogonal transformations and the calculation of posterior probabilities based on a continuous-density hidden Markov model with a Gaussian mixture model expressed in the reduced-dimensional space. The analysis can be incorporated into an NN, which is named a time-series discriminant component network (TSDCN), so that parameters of dimensionality reduction and classification can be obtained simultaneously as network coefficients according to a backpropagation through time-based learning algorithm with the Lagrange multiplier method. The TSDCN is considered to enable high-accuracy classification of high-dimensional time-series patterns and to reduce the computation time taken for network training. The validity of the TSDCN is demonstrated for high-dimensional artificial data and electroencephalogram signals in the experiments conducted during the study.

This paper proposes a novel method of estimating posteriori probability for learned and unlearned classes based on a Gaussian mixture model (GMM). With prior distributions of learned and unlearned classes defined as a novel GMM incorporating a one-versus-the-rest classifier, any defined/undefined class can be classified through training of the classifier using given training samples. This method can be used for bioelectric signal discrimination in various applications such as human-machine interfaces and diagnosis support systems. In the experiments reported here, artificial data generated from Gaussian distributions and electromyogram (EMG) patterns measured from the forearm muscles of a volunteer were classified to demonstrate the capabilities of the proposed method for learned and unlearned class discrimination. The results showed that the approach produces high performance for classification of learned (artificial data: 100%; EMG patterns: 95.6%) and unlearned (artificial data: 93.4%; EMG patterns: 70.4%) classes based on simple neural network comparison, and indicated that the proposed method is applicable to human-machine interfaces such as prosthetic hand control systems.

This paper proposes a control method for a human-assisting manipulator using acceleration sensors. The technique involves an arm control part (ACP) and a hand-and-wrist control part (HWCP); the ACP controls the manipulator's shoulder and elbow joints using acceleration signals, while the HWCP controls the corresponding joints using mechanomyogram (MMG) signals measured from the human operator. A distinctive feature of the proposed method is its estimation of information on force and motion from measured acceleration signals using MMG processing and a probabilistic neural network. Experiments demonstrated that the MMG patterns seen during hand and wrist motion can be classified sufficiently (average rate: 94.3%), and that a prosthetic manipulator can be controlled using the acceleration signals measured. Such manipulators are expected to prove useful as assistive devices for people with physical disabilities.

This paper proposes a novel human interface that can be used to operate domestic appliances and game machines. In this system, finger tapping movement is measured by magnetic sensors, and is evaluated by computing five features, such as the finger tapping interval, on a real-time basis. These factors are then discriminated through a probabilistic neural network and allocated as machine operation commands. The user can thus voluntarily operate various machines using finger tapping movements. Experimental results showed that the prototype system developed can learn and classify user movements with a high degree of accuracy (average rates: 98.56plusmn1.15 [%]), and that it can be used to smoothly control appliances and game machines. The possibility of the system supporting finger movement training was also confirmed with three subjects through game operation.