SAS-Net: An Agitated Behavior Early Warning Model for Community-Dwelling Dementia Patients Based on Symmetric Autoencoders and Spatio-Temporal Network

Using home sensors to provide agitation warnings for community-dwelling dementia patients without ongoing clinical supervision can enable their carers to intervene early during the agitation latent period, thereby reducing unnecessary hospital admissions and harmful events. Most existing studies are based on behavioral, sleep, and physiological data from the previous 24 h to predict patients’ agitation events, which fail to fully capture patients’ recent behavioral details. In this study, monitoring data from the previous 8 × 24 h for dementia patients are used to achieve in-depth mining of patients’ living habits. Furthermore, to address the common clinical problem of extreme imbalance between agitation and normal samples (agitation samples often are extremely scarce), we designed a two-stage agitation early warning model based on symmetric autoencoders and a spatio-temporal network, dubbed SAS-Net. In the first stage, we randomly sample 80% of normal samples and employ multiple symmetric autoencoders to perform feature transformation and pseudo-label construction, and then pre-train the spatio-temporal learning network composed of convolutional networks and a gated Multilayer Perceptron. It aims to learn the patient data’s intrinsic structure, underlying patterns, and effective feature extraction methods. In the second stage, we freeze the parameters of the spatio-temporal learning network and use a balanced dataset consisting of the remaining 20% of normal samples and all agitation samples to reconstruct and fine-tune the top fully connected classifier to improve the recognition performance of agitation samples. The two-stage strategy resolves the problem of ineffective training faced by deep learning models on imbalanced datasets. The experimental results demonstrate the effectiveness of the proposed SAS-Net for agitated behavior early warning.