In this paper, we aim to develop a stress detection model for caregivers using data collected from a real-world care facility. The importance of stress management has grown, and many physiological datasets and stress detection studies exist for nurses. Caregivers experience various types of stress, including interpersonal issues and assisting with toileting, and are physically active throughout the day, moving around the floor and performing tasks like transfer assistance. These diverse stressors and physical activity can significantly affect stress detection using wearable sensors. However, sufficient data collection in caregiving settings is not being conducted compared to other fields (e.g., nursing, doctor). In this study, we conduct a data collection experiment and evaluate the results of stress detection machine learning using that data. In the data collection, we create 8-day dataset from four caregivers that includes care record data, wearable sensor data, and self-reported stress labels. Next, we extracted statistical features from the time-series data and performed stress detection using a Random Forest model. As a result, we achieved a maximum classification accuracy of 72%. While data augmentation improved the detection of minority classes such as High Stress, it also lowered the overall classification performance, revealing a trade-off that remains a challenge. Nevertheless, this dataset makes it possible to analyze caregiver stress in relation to specific care activities, representing an important step toward stress-aware support in the caregiving domain.

抄録全体を表示

Modeling fuel consumption based on driving states is essential in developing eco-driving and route selection strategies for advanced driver assistance systems. However, accurately predicting vehicle fuel consumption is challenging due to dynamic factors like road slope, vehicle conditions, and rapid changes in driving behavior. Developed based on ideal road and vehicle conditions, traditional empirical models often fail to adapt to these conditions, leading to inaccurate estimations. This study explores incorporating throttle position, a critical indicator of driving activity regarding motion control, acceleration, and speed, to enhance predictive accuracy, overcoming the unknown slope effects. Since the throttle position indicating the applied torque to the engine has a complicated relationship associated with engine and vehicle speeds, gear, and road slope, machine learning methods are employed to develop the model instead of typical empirical models, like VT-Micro, for standard driving. The results show that the Random Tree performed best on the training dataset (RMSE= 0.2396, R2 = 0.9859). They are further evaluated on a cross-dataset, which is not used for training, where Neural Network outperformed all models (RMSE = 0.4462, R2 = 0.9545). In contrast, VT-Micro consistently exhibited poor accuracy, particularly in transient driving conditions (RMSE = 2.0282, R2 = 0.0591). These findings underscore the advantages of using activity information with machine learning for modeling complex, nonlinear vehicle dynamics, providing more accurate and adaptive fuel consumption predictions for sustainable transportation planning and energy management.

抄録全体を表示

This study examined the attitudes of individuals toward texts generated by large language models (LLMs), including social networking service posts and news comments. Recently, the number of people viewing texts generated by LLMs has increased. Because an LLM can generate natural texts that are almost indistinguishable from those written by humans, there is concern that generating such natural texts may cause problems, such as maliciously influencing public opinion. To evaluate the reception of LLM-generated texts, we conducted an experiment based on the hypothesis that the knowledge that a text was generated by an LLM would influence user acceptance. In the experiment, participants were shown news comments that included AI-generated comments. We controlled whether the user was aware that the text had been generated by an LLM, and assessed their viewpoints from four perspectives: perceived friendliness, trustworthiness, empathy, and reference. The results showed that a generated comment imitating the opinion of an expert increased in rank when it was disclosed that the LLM generated the comment. In particular, “reliability” and “informative” were sensitive to this disclosure, whereas “familiar” and “empathy” were not. This result suggests that expert labeling significantly enhances perceived reliability, and the finding raises concerns about the potential for news viewers to be implicitly guided toward a particular opinion.

抄録全体を表示

This paper proposes PatchEchoClassifier, an energy-efficient classifier for time-series data that leverages a reservoir-based mechanism known as the Echo State Network (ESN). Designed for human activity recognition (HAR) using one-dimensional sensor signals, the model employs a tokenizer to extract patch-level representations. To train the model effectively, we introduce a knowledge distillation framework that transfers knowledge from a high-capacity MLP-Mixer teacher to a lightweight reservoir-based student. Experimental results on multiple HAR datasets demonstrate that our model achieves over 80% accuracy while substantially reducing computational cost. Notably, PatchEchoClassifier requires only about one-sixth of the floating-point operations (FLOPs) compared to DeepConvLSTM, a commonly used convolutional baseline. These findings indicate that PatchEchoClassifier is a promising solution for real-time, energy-efficient human activity recognition in edge computing environments. An open-source implementation of our method is provided. It is available at https://github.com/Okita-Laboratory/PatchEchoClassifier/.

抄録全体を表示

In recent years, the widespread adoption of wearable devices has high-lighted the growing importance of behavior analysis using IMU. While applications span diverse fields such as healthcare and robotics, recent studies have increasingly focused on multimodal analysis, in addition to unimodal analysis. Several studies have proposed multimodal foundation models that incorporate first-person video and text data; however, these models still fall short in providing a detailed analysis of full-body human activity. To address this limitation, we propose Activity Understanding and Representations Alignment - Multimodal Foundation Model (AURA-MFM), a foundation model that integrates four modalities: third-person video, motion capture, IMU, and text. By incorporating third-person video and motion capture data, the model enables a detailed and multidimensional understanding of human activity, which first-person video alone fails to capture. Additionally, a Transformer-based IMU encoder is employed to enhance the model’s overall performance. Experimental evaluations on retrieval and activity recognition tasks demonstrate that our model surpasses existing methods. Notably, in the zero-shot classification for action recognition, our method achieved significantly higher performance, with an F1-score of 0.6226 and an accuracy of 0.7320, whereas the existing method recorded an F1-score of 0.0747 and an accuracy of 0.1961. The code is available at https://github.com/Okita-Laboratory/AURA-MFM.

抄録全体を表示

Fear of falling remains a significant concern among individuals with Parkinson’s Disease (PD). Elevated scores on the Fall Efficacy Scale-International (FES-I) are associated with recurrent falls and anxiety that limits daily activities, ultimately increasing fall risk. While machine learning presents opportunities for fall risk prediction, the integration of psychological indicators and advanced feature engineering techniques, particularly for gait data, has not been thoroughly investigated. This study proposes quantum state probability values as novel features to predict fear of falling. Utilizing a motion capture dataset from the Federal University of ABC, Brazil, the model incorporates demographic data, L-Dopa dosage, and gait features. Among four dataset configurations, the model using only quantum state probability features achieved 93% accuracy with an SVM-RBF classifier. High Cohen’s Kappa and MCC values confirmed the strong predictive alignment with true labels. Visualization using PCA and t-SNE demonstrated less overlapping class separation in quantum state probability as feature.

抄録全体を表示

Parkinson’s disease tremor classification from wearable-sensor time series remains challenging due to small datasets, class imbalance, and signal noise. We propose a hybrid model that fuses a Transformer encoder for contextualized feature extraction with a deep Gaussian process (DGP) for probabilistic classification, further accelerated by variational inference (VGP). On the Tremor Challenge benchmark, our method achieves a Test F1 score of 0.4336 using a seven-layer Transformer trained for 3000 epochs and a spectral-mixture kernel. We find that moderate Transformer depth balances representation power against overfitting, and that kernel choice critically interacts with DGP depth: while increasing DGP layers often degrades performance in limited data regimes, spectral-mixture kernels can harness additional hierarchy to improve accuracy. Extended training benefits deeper Transformer variants, suggesting that parameter- rich models require longer optimization. We analyze limitations including dataset size, class imbalance, computational cost, and the omission of uncertainty quantification. We further identify opportunities for improvement via learnable time embeddings (e.g., Time2Vec) and cross-dataset validation. Our results demonstrate the viability of combining self-attention with Bayesian nonparametrics for robust biomedical time-series classification, offering a promising direction for clinical AI applications.

抄録全体を表示