In this paper, we investigate the practical effectiveness of unimodal EEG and multimodal EMG–IMU sensing for upper-limb movement intention detection in assistive exoskeleton control. Accurate and low-latency intention decoding is essential for post-stroke rehabilitation; however, many prior studies conflate sensing modality, model architecture, and evaluation protocol, limiting fair comparison and practical interpretation. To isolate the effect of sensing configuration, we adopt a controlled top-down evaluation in which the learning algorithm, feature extraction framework, and hyperparameters are fixed, and only the biosignal modality is varied. The study follows the experimental framework defined in the MUMIDC challenge, implementing both subject-wise and cross-subject evaluation settings using reproducible stratified splits. Using the MUMID dataset, we formulate a seven-class classification task covering functional upper-limb movement primitives. A unified preprocessing and feature extraction pipeline is applied to EEG, EMG, and IMU signals, followed by classification using a fixed multi-class XGBoost model. Two sensing configurations are evaluated: EEG-only and early-fusion EMG+IMU. Performance is assessed using Accuracy and macro-averaged F1-score. Results show that EMG+IMU consistently outperforms EEG-only across both evaluation settings. In subject-wise validation, EMG+IMU achieves 89.80\% Accuracy and 89.77\% Macro-F1, compared to approximately 50\% for EEG-only, with substantially lower inter-subject variability. Under cross-subject evaluation, EMG+IMU exhibits only minor degradation, whereas EEG-only declines markedly. These findings highlight a clear trade-off between sensing complexity and robustness for practical exoskeleton control. \\Keywords: Upper-limb rehabilitation, exoskeleton control, movement intention detection, EEG, EMG, IMU, multimodal fusion, machine learning, XGBoost.

抄録全体を表示

Indoor location recognition in nursing care facilities is critical for activity monitoring and safety management. Bluetooth Low Energy (BLE) beacon signals collected in real-world environments are noisy and highly imbalanced across locations. But these noises and imbalances make reliable localization challenging. This study aims to develop and evaluate a robust BLE-based indoor location recognition framework that effectively captures short-term temporal signal patterns under realistic care-facility conditions. An established Temporal Convolutional Network (TCN) architecture is applied as part of a TCN-based framework. BLE RSSI signals are provided by the ABC 2026 Activity and Location Recognition Challenge (ALRC). The approach integrates RSSI aggregation, normalization, and masking of missing beacon observations. Then, a fixed-length temporal windowing is used to model temporal dependencies in BLE sequences. Model performance is assessed using a stratified training–validation split, with Macro F1-score adopted as the primary metric to address severe class imbalance. Experimental results demonstrate that the proposed framework achieves a validation accuracy of 0.66 and a Macro F1-score of 0.56 across 22 indoor location classes. Comparative evaluation against five baseline classifiers shows more balanced performance. Particularly, our study has improved recognition of underrepresented locations. The main contribution of this work lies in the adaptation and evaluation of temporal convolutional modeling for BLE-based indoor localization in nursing care environments. Overall, the results indicate that temporal modeling significantly enhances robustness and reliability of BLE-based indoor location recognition in care-facility scenarios.

抄録全体を表示

Indoor location recognition based on beacon signals plays a crucial role in healthcare environments, particularly in nursing homes where nurses are responsible for caring for multiple elderly residents simultaneously. This paper is written as part of the ABC 2026 Challenge (Decode The Invisible: Activity and Location Recognition Challenge in Care Facility), which focuses on evaluating indoor localization methods under realistic and challenging healthcare conditions. Reliable and accurate localization of nursing staff enables timely assistance, improves care coordination, and enhances the overall efficiency and safety of healthcare workflows. However, real-world beacon data are often noisy, unstable, misaligned, or imbalanced, creating challenges for reliable recognition. This study proposes a supervised learning–based approach for indoor nurse localization with a novel Received Signal Strength Indicator (RSSI) pattern–based preprocessing strategy. Each signal segment is represented using dominant beacon patterns to robustly characterize spatial signal distributions, while an outlier detection mechanism is applied to identify and remove abnormal RSSI patterns caused by environmental dynamics. Based on the refined dataset, a supervised learning model is trained to recognize nurses’ indoor locations in a real healthcare environment. Experimental results demonstrate that the proposed method substantially improves localization performance, achieving 79% accuracy and a 68% F1-score, compared to 58% accuracy and a 34% F1-score obtained using raw beacon signals. These results highlight the effectiveness of RSSI pattern–based preprocessing for robust indoor localization in healthcare settings.

抄録全体を表示

Human Activity Recognition (HAR) using wearable sensors is fundamental to healthcare compliance monitoring, industrial safety, and smart environments. However, conventional HAR systems require large labeled datasets, creating a significant deployment barrier in new domains. Responding to the NHIC-ABC2026 of From Labels to Text: Next HAR Idea Challenge, we propose a zero-shot alignment framework that bridges dual-wrist accelerometer signals with natural language activity descriptions without requiring any labeled training examples. Our three-stage pipeline extracts structured semantic motion profiles from activity descriptions using a large language model, computes isomorphic sensor features from raw accelerometer data, and applies a weighted rule-based similarity function with top-K ranking (K=3, τ = 0.6) to preserve inference uncertainty. Evaluated on the NHIC-ABC2026 salad preparation dataset across eight activities, the method achieved 94.8% window-level detection. Semantically distinctive motions such as rinsing yielded higher-confidence matches (0.740) compared to kinematically similar reaching actions (0.635), which remain indistinguishable in the current feature space. An overlap-aware temporal correction reduced systematic duration over-estimation by 2.2%. These results demonstrate that semantic alignment offers an interpretable, deployable pathway for zero-shot HAR in domains where labeled sensor data is scarce.

抄録全体を表示

Accurate classification of biosignals from surface electrodes is challenging because signals are noisy and many actions have similar patterns. Reliable recognition of upper-limb actions is important for rehabilitation and human–machine interaction systems. This paper presents a crosssubject classification method for similar appearing yet distinct upper-limb actions using high-dimensional frequency-domain features extracted from EEG, EMG, and IMU signals in the Multimodal Upper-Limb Movement Dataset. Seven actions are analyzed using standardized preprocessing and modality-specific feature extraction, with models evaluated under subjectwise and cross-subject protocols to assess generalization. EMG+IMU features achieve subject-wise accuracy above 95%, showing strong and repeatable neuromuscular and motion patterns. In cross-subject testing, performance decreases for all models, but Subspace KNN remains robust with 94.64% accuracy. In contrast, EEG-only models show high variability and poor cross-subject performance because of strong inter-subject differences and weak class separability. While subject-wise EEG classification achieves accuracies up to 83.33%, cross-subject generalization remains limited, likely due to constrained dataset size, limited channel coverage, and inadequate motor cortex information. The results highlight the limitations of standalone EEG for subject-independent action classification due to dominant noise in low-amplitude surface signals, supporting multimodal approaches for reliable real-world deployment. This work is part of the Multimodal Upper-Limb Movement Intent Detection Challenge at the Activity and Behavior Computing Conference.

抄録全体を表示

Understanding caregiver behavior in nursing facilities requires utilizing real-world sensor data that is often noisy, misaligned, and imbalanced. We address this using a multimodal dataset from the Activity and Location Recognition Challenge in Care Facility, which includes BLE beacon signal collected in a working nursing home. Rather than proposing a new model architecture, we focus on a simple but effective feature representation: the count of distinct BLE beacons detected within each time window. This captures coarse spatial context without relying solely on unstable RSSI values or explicit distance estimation. Using this feature alongside RSSI, we evaluate several tree-based ensemble methods and select XGBoost as our base classifier with highest accuracy of 60% for its robustness under class imbalance. Our results demonstrate that even with minimal preprocessing and no complex fusion mechanisms, location prediction is achievable in realistic, imperfect conditions.

抄録全体を表示

Conventional human activity recognition (HAR) primarily maps wearable sensor time series to predefined action labels. While effective for well-defined actions, this paradigm is limited for open-ended activities in which neither a single correct sequence nor a unique outcome can be specified, and in which the manner of engagement has intrinsic value. Responding to the NHIC theme ``From Labels to Text,'' we propose a process-oriented HAR framework that represents how an activity progresses and describes its characteristics in natural language without action classification or outcome evaluation. Using dominant-arm wrist-worn IMU data, we compute acceleration magnitude, segment it into short windows, and derive low-granularity activity states (movement/rest). From state transitions and their temporal structure, the pipeline extracts meta-level process features: rhythm stability, trial-and-error patterns (short pauses followed by resumption), self-defined boundaries (relatively long pauses), and adaptive switching between early and late phases. These features are mapped to concise post-hoc positive feedback (PFB) statements that are descriptive and non-judgmental; a language model is used only to paraphrase templates. To examine generality, we apply the same pipeline to two contrasting datasets: an instruction-guided salad preparation task provided by NHIC and an open-ended collaborative paper-tower construction task from our prior studies. Across both tasks, we identify common process structures, suggesting that process-oriented, text-based feedback provides an alternative alignment between sensor time series and natural language beyond label-based HAR.

抄録全体を表示

Indoor localization in nursing homes, addressed through the ABC 2026 'Decode the Invisible' Challenge, faces significant obstacles because of a strong class imbalance, where minority rooms are underrepresented in the training data and high-traffic areas predominate, leading to inadequate location coverage and poor recognition performance. The study suggests an Ultimate inference framework that combines ensemble voting, minority boosting, and Random Forest classification to achieve 100\% prediction coverage while preserving accuracy across all room classes. The technique processes 4,107 CSV files into 40-second intervals using 33-dimensional feature vectors using Bluetooth Low Energy (BLE) beacon signals from 25 stationary transmitters. The scarcity of minority classes is addressed by augmenting data using Gaussian noise injection ($\sigma = 1.0$ dB). The suggested method successfully recovers (1) rooms 503 and 510 from zero recall to 0.800 and 1.000 F1-scores, respectively; (2) detects 20 out of 22 room classes compared to 18 in baseline methods; and (3) achieves a weighted F1-score of 0.7492 and Macro F1-score of 0.6241 with full coverage. While conventional approaches sacrifice coverage for accuracy, our ensemble-based minority recovery approach maintains macro-level fairness without compromising majority class performance. For healthcare settings with limited resources, this provides a consistent solution.

抄録全体を表示

Indoor location recognition is important for understanding caregiving activities and improving operational efficiency in nursing care facilities. This study investigates a room-level localization approach for caregivers under real-world conditions characterized by RSSI instability and severe class imbalance. A hierarchical classification framework is adopted to decompose the localization task into two stages: coarse-grained classification between patient rooms and non-patient areas, followed by conditional fine-grained recognition of individual patient rooms. RSSI data collected from a real-world nursing care facility are segmented into fixed-length temporal windows and represented using a unified feature set. Conventional machine learning models and temporal models, including MLP, LSTM, and GRU, are evaluated using accuracy, macro-averaged F1 score, and training dynamics analysis. Experimental results show that Level 1 classification achieves stable accuracy above 0.87 across all models, confirming reliable separation of patient and non-patient areas. For fine-grained room recognition, hierarchical classification improves Level 2 macro F1 scores when evaluation is restricted to samples correctly classified at Level 1, highlighting the impact of error propagation. XGBoost achieves the highest test macro F1 score of 0.77, while feedforward and other tree-based models show moderate improvements. In contrast, LSTM and GRU exhibit stronger overfitting, with noticeable performance degradation from validation to test data. These findings demonstrate that hierarchical modeling effectively mitigates class imbalance and reduces room-level confusion in RSSI-based indoor localization. This work was conducted as part of the ABC 2026 Challenge, ALRC: Decode the Invisible: Activity and Location Recognition Challenge in Care Facility.

抄録全体を表示

In nursing care facilities, practical challenges with Bluetooth Low Energy (BLE)-based indoor localization include extreme RSSI noise, temporal sparsity, and class imbalance between rooms. This study does not introduce a new localization algorithm; instead, it proposes a floor plan aware feature engineering framework that transforms raw RSSI signals into spatially structured representations by embedding prior knowledge of beacon placement, room adjacency, and temporal activity patterns. RSSI statistics, zone based aggregation, inter zone interactions, proximity based transformations, and temporal behavioral indicators that represent routine movement dynamics inside the facility are all included into the designed features. XGBoost obtains 73.35% accuracy under stratified evaluation (weighted F1: 0.7282). The accuracy of Leave-One-Day-Out cross validation (LODO-CV), which is used to evaluate resilience against temporal bias, is 54.1% (weighted F1: 0.5230) under day level generalization. While temporal indicators provide moderate increases under rigorous validation, ablation analysis shows that spatial feature groupings contribute cooperatively. The majority of errors, according to confusion matrix analysis, happen between geographically neighboring rooms, indicating coherent spatial learning rather than random signal correlation. These findings show that feature design that incorporates floor plan and behavioral context enhances resilience under practical deployment constraints without the need for more sensors or calibration.

抄録全体を表示

The ABC TICO-ALRC 2026 presents a highly challenging indoor localization problem in nursing care facilities, characterized by extreme class imbalance in room visit frequencies. Minority patient rooms are severely underrepresented due to natural caregiving workflows and technical constraints in Bluetooth Low Energy (BLE) signal collection. To address this issue, this study proposes a two-stage, signal-pattern-guided relabeling-based data augmentation framework for BLE-based indoor localization. The method leverages statistical similarity between rooms using KL/JS divergence for broad initial augmentation, followed by Wasserstein distance for targeted refinement of remaining problematic classes. Importantly, the approach reuses only real signal measurements and incorporates spatial neighbor constraints to preserve physical plausibility. Experiments are conducted on the ABC TICO-ALRC 2026 dataset using Leave-One-Day-Out (LODO) cross-validation and Random Forest classification. Results show that the proposed two-stage approach achieves a mean Macro-F1 score of 67.9%, outperforming the baseline (37.4%), single-stage KL/JS relabeling (58.4%), and single-stage Wasserstein-based relabeling (64.5%). These findings indicate that leveraging signal pattern similarity through iterative and physically grounded relabeling provides a practical and robust solution to extreme class imbalance in real-world nursing home localization scenarios.

抄録全体を表示

Accurate indoor location recognition is critical for monitoring and operational analysis in nursing care facilities. However, Bluetooth Low Energy (BLE) signal data collected in real-world environments are typically noisy, sparse, and highly imbalanced, limiting the reliability of conventional localization methods. This paper is submitted to the ALRC: Decode the Invisible Activity and Location Recognition Challenge in Care Facility organized by Kyushu Institute of Technology at ABC 2026. We propose a data-centric machine learning framework for BLE-based indoor location recognition that prioritizes systematic preprocessing and multi-stage feature engineering over additional sensor deployment or complex modeling architectures. Raw RSSI signals are processed through temporal windowing and transformed into a comprehensive multi-dimensional representation integrating statistical descriptors, spatial relationships among beacons, temporal dynamics, and distance-based features derived from a log-distance path-loss model. An Extreme Gradient Boosting (XGBoost) classifier is trained using randomized cross-validation with F1-micro optimization to address class imbalance. Experiments on a large-scale real-world nursing facility dataset containing 1,928,175 samples and 230 engineered features achieve test accuracy above 80% with stable generalization across location classes. Feature importance and ablation analyses demonstrate that engineered spatial and distance-based representations contribute more significantly than raw RSSI signals alone. The results confirm that systematic data-centric feature optimization enables robust and scalable BLE-based indoor localization under realistic deployment constraints without requiring additional infrastructure.

抄録全体を表示