Falls are a leading cause of injury and morbidity among older adults,especially those with Alzheimer’s disease(AD),who face increased risks due to cognitive decline,gait instability,and impaired spatial awareness.Whi...Falls are a leading cause of injury and morbidity among older adults,especially those with Alzheimer’s disease(AD),who face increased risks due to cognitive decline,gait instability,and impaired spatial awareness.While wearable sensor-based fall detection systems offer promising solutions,their effectiveness is often hindered by domain shifts resulting from variations in sensor placement,sampling frequencies,and discrepancies in dataset distributions.To address these challenges,this paper proposes a novel unsupervised domain adaptation(UDA)framework specifically designed for cross-dataset fall detection in Alzheimer’s disease(AD)patients,utilizing advanced transfer learning to enhance generalizability.The proposed method incorporates a ResNet-Transformer Network(ResT)as a feature extractor,along with a novel DualAlign Loss formulation that aims to align feature distributions while maintaining class separability.Experiments on the preprocessed KFall and SisFall datasets demonstrate significant improvements in F1-score and recall,crucial metrics for reliable fall detection,outperforming existing UDA methods,including a convolutional neural network(CNN),DeepCORAL,DANN,and CDAN.By addressing domain shifts,the proposed approach enhances the practical viability of fall detection systems for AD patients,providing a scalable solution to minimize injury risks and improve caregiving outcomes in real-world environments.展开更多
基金funded by the King Salman Center for Disability Research through Research Group no.KSRG-2024-430.
文摘Falls are a leading cause of injury and morbidity among older adults,especially those with Alzheimer’s disease(AD),who face increased risks due to cognitive decline,gait instability,and impaired spatial awareness.While wearable sensor-based fall detection systems offer promising solutions,their effectiveness is often hindered by domain shifts resulting from variations in sensor placement,sampling frequencies,and discrepancies in dataset distributions.To address these challenges,this paper proposes a novel unsupervised domain adaptation(UDA)framework specifically designed for cross-dataset fall detection in Alzheimer’s disease(AD)patients,utilizing advanced transfer learning to enhance generalizability.The proposed method incorporates a ResNet-Transformer Network(ResT)as a feature extractor,along with a novel DualAlign Loss formulation that aims to align feature distributions while maintaining class separability.Experiments on the preprocessed KFall and SisFall datasets demonstrate significant improvements in F1-score and recall,crucial metrics for reliable fall detection,outperforming existing UDA methods,including a convolutional neural network(CNN),DeepCORAL,DANN,and CDAN.By addressing domain shifts,the proposed approach enhances the practical viability of fall detection systems for AD patients,providing a scalable solution to minimize injury risks and improve caregiving outcomes in real-world environments.
