Sozo Inoue

In this paper, we propose an outcome-based contrastive loss for imbalanced binary classification to alert to next-day health deterioration using care records and meteorological data. Long-term care facilities maintain daily care and observation records to monitor the health of older adults. Such objective records are particularly valuable when sudden deterioration occurs, enabling timely coordination with medical institutions. Predicting deterioration one day in advance could provide care staff with an actionable window to intensify observation and adjust care plans (e.g., scheduling additional vital checks or increasing fluid intake monitoring). This could potentially reduce emergency transports and ease the burden on already understaffed care facilities. However, for such predictions to be useful in practice, false positives must be suppressed. Because deterioration events are rare, class imbalance generates an excess of false positives, causing alert fatigue and increasing the risk that actual events go unnoticed. To address these challenges, we propose an outcome-based contrastive loss that contrasts actual deteriorating samples against false alarms conditioned on mini-batch prediction outcomes. The proposed loss contracts same-label pairs to shape local structure within each ground-truth label. The loss also separates actual deteriorating samples from false alarms among samples predicted as deteriorating, thereby directly reducing unnecessary alerts. As a result, compared with random oversampling with standard cross-entropy, the proposed model improved precision from 3.97% to 12.94% (+8.97 percentage points), while limiting the F1-score decrease to 0.71 percentage points (from 7.28% to 6.57%). Pair-design ablations and UMAP projections supported this mechanism by indicating clearer separation between actually deteriorating and false-alarm samples in the learned representation space. These results suggest a viable direction for alert systems that produce fewer unnecessary alerts, reducing alert fatigue and supporting more reliable deterioration detection in care settings.

Background: Older adult activity recognition is a critical task in long-term care monitoring; yet, it remains challenging due to postural deformities and health-related variability. These factors cause different activities to appear visually similar, or the same activity to appear dissimilar, undermining the effectiveness of traditional human activity recognition models developed for the general population. Objective: This study aims to develop an improved older adult activity recognition method that integrates care assessment information with motion data to capture and understand movement variability arising from different health conditions. Methods: To achieve our objective, we propose a care-assessment-aware spatiotemporal transformer (CSTT) model that integrates body key points, heatmaps, and care level data for personalized and context-aware activity recognition. The model dynamically adjusts its attention mechanism based on care level context to improve recognition accuracy. CSTT was trained and validated on real-world older adult motion data. A total of 51 older adult participants (30 men and 21 women; age range of 64-95 years) were included in the study. Among them, 7 (13.7%) required high care assistance, 26 (51.0%) required medium care assistance, and 18 (35.3%) required low care assistance. Results: Despite data imbalance and considerable intraclass variation due to differing care needs, the proposed CSTT model achieved an F1-score and accuracy of 0.96 and area under the curve is 0.98. Analysis revealed that movement patterns differ significantly across care levels and that similar motions occur in distinct activities, highlighting the importance of care-aware modeling. Conclusions: Incorporating care level information into activity recognition models significantly enhances performance in older adult care settings. The proposed CSTT framework demonstrates the value of personalized, context-sensitive approaches for accurate and ethical monitoring in long-term care environments.

With increasing life expectancy, particularly in developed nations, the proportion of elderly individuals is rising rapidly, necessitating advanced systems for continuous monitoring and timely intervention to support independent living and enhance safety in assisted care environments. Falls are among the leading causes of hospitalizations and deaths related to injuries in this demographic, highlighting the urgent need for intelligent fall detection systems. However, most existing solutions struggle with real-world deployment due to incomplete anomaly modeling and a lack of contextual location awareness. This article introduces a novel position-aware indoor activity recognition and fall detection approach that uses spatial and motion data to detect falls with high accuracy and contextual relevance. The system integrates ultrawideband (UWB) positioning technology with a multilayer perceptron (MLP) model to achieve indoor localization. Furthermore, accelerometer and gyroscope data are used for activity monitoring, which is processed using a hybrid deep learning (DL) architecture that combines a variational autoencoder (VAE), convolutional neural networks (CNNs), and long short-term memory (LSTM) networks. This architecture takes advantage of temporal and spatial feature extraction for improved fall detection. The localization module achieves over 96% accuracy. For activity recognition, the VAE CNN-LSTM model achieves fall detection accuracy exceeding 97%. A late fusion decision layer combines spatial and activity-level insights to enable precise detection and localization of fall events within indoor environments. The proposed system is validated in a real-world smart home setting and demonstrates strong performance in terms of accuracy, scalability, and adaptability.

The rising demand for elderly care amid ongoing nursing shortages has highlighted the limitations of conventional alert systems, which frequently generate excessive alerts and contribute to alarm fatigue. The objective of this study is to develop a hybrid, context-aware nurse alerting framework for long-term care (LTC) facilities that minimizes redundant alarms, reduces alarm fatigue, and enhances patient safety and caregiving balance during multi-person care scenarios such as mealtimes. To do so, we aimed to intelligently suppress, delay, and validate alerts by integrating rule-based logic with Large Language Model (LLM)-driven semantic reasoning. We conducted an experimental study in a real-world LTC environment involving 28 elderly residents (6 high, 8 medium, and 14 low care levels) and four nurses across three rooms over seven days. The proposed system utilizes video-derived skeletal motion, care-level annotations, and dynamic nurse–elderly proximity for decision making. Statistical analyses were performed using F1 score, accuracy, false positive rate (FPR), and false negative rate (FNR) to evaluate performance improvements. Compared to the baseline where all nurses were notified (100% alarm load), the proposed method reduced average alarm load to 27.5%, achieving a 72.5% reduction, with suppression rates reaching 100% in some rooms for some nurses. Performance metrics further validate the system’s effectiveness: the macro F1 score improved from 0.18 (baseline) to 0.97, while accuracy rose from 0.21 (baseline) to 0.98. Compared to the baseline error rates (FPR 0.20, FNR 0.79), the proposed method achieved drastically lower values (FPR 0.005, FNR 0.023). Across both spatial (room-level) and temporal (day-level) validations, the proposed approach consistently outperformed baseline and purely rule-based methods. These findings demonstrate that the proposed approach effectively minimizes false alarms while maintaining strong operational efficiency. By integrating rule-based mechanisms with LLM-based contextual reasoning, the framework significantly enhances alert accuracy, mitigates alarm fatigue, and promotes safer, more sustainable, and human-centered care practices, making it suitable for practical deployment within real-world long-term care environments.

BACKGROUND Older adult activity recognition is a critical task in long-term care monitoring; yet, it remains challenging due to postural deformities and health-related variability. These factors cause different activities to appear visually similar, or the same activity to appear dissimilar, undermining the effectiveness of traditional human activity recognition models developed for the general population. OBJECTIVE This study aims to develop an improved older adult activity recognition method that integrates care assessment information with motion data to capture and understand movement variability arising from different health conditions. METHODS To achieve our objective, we propose a care-assessment-aware spatiotemporal transformer (CSTT) model that integrates body key points, heatmaps, and care level data for personalized and context-aware activity recognition. The model dynamically adjusts its attention mechanism based on care level context to improve recognition accuracy. CSTT was trained and validated on real-world older adult motion data. A total of 51 older adult participants (30 men and 21 women; age range of 64-95 years) were included in the study. Among them, 7 (13.7%) required high care assistance, 26 (51.0%) required medium care assistance, and 18 (35.3%) required low care assistance. RESULTS Despite data imbalance and considerable intraclass variation due to differing care needs, the proposed CSTT model achieved an F1-score and accuracy of 0.96 and area under the curve is 0.98. Analysis revealed that movement patterns differ significantly across care levels and that similar motions occur in distinct activities, highlighting the importance of care-aware modeling. CONCLUSIONS Incorporating care level information into activity recognition models significantly enhances performance in older adult care settings. The proposed CSTT framework demonstrates the value of personalized, context-sensitive approaches for accurate and ethical monitoring in long-term care environments. CLINICALTRIAL

In this paper, we propose a novel few-shot optimization with Hybrid Euclidean Distance with Large Language Models (HED-LM) to improve example selection for sensor-based classification tasks. While few-shot prompting enables efficient inference with limited labeled data, its performance largely depends on the quality of selected examples. HED-LM addresses this challenge through a hybrid selection pipeline that filters candidate examples based on Euclidean distance and re-ranks them using contextual relevance scored by large language models (LLMs). To validate its effectiveness, we apply HED-LM to a fatigue detection task using accelerometer data characterized by overlapping patterns and high inter-subject variability. Unlike simpler tasks such as activity recognition, fatigue detection demands more nuanced example selection due to subtle differences in physiological signals. Our experiments show that HED-LM achieves a mean macro F1-score of 69.13 ± 10.71%, outperforming both random selection (59.30 ± 10.13%) and distance-only filtering (67.61 ± 11.39%). These represent relative improvements of 16.6% and 2.3%, respectively. The results confirm that combining numerical similarity with contextual relevance improves the robustness of few-shot prompting. Overall, HED-LM offers a practical solution to improve performance in real-world sensor-based learning tasks and shows potential for broader applications in healthcare monitoring, human activity recognition, and industrial safety scenarios.