Geriatric Care Management System Powered by the IoT and Computer Vision Techniques

**Background:** Geriatric care faces significant challenges due to aging populations and nursing staff shortages. In Europe, an estimated 30.8 million dependent people need long-term care, expected to rise to 38 million by 2050, while a shortage of 2.3 million nurses is projected by 2030. Current care processes are inefficient: manual vital sign collection takes up to 5 minutes per patient, and documentation can require up to 37 minutes per patient. Paper-based care plans and fragmented data sources lead to missed clinical changes and delayed treatment. The authors identified six key needs through stakeholder interviews in Lithuania, including automated monitoring, digital care plans, and proactive care capabilities.

**Methods:** The proposed system integrates four IoT devices: EZVIZ CS-C3TN cameras (1920×1080 resolution), Fitbit Charge 5 wristbands, Withings BPM Connect blood pressure monitor, and Withings Body+ scales. Data collection occurred from September 15, 2022 to December 28, 2022, accumulating 1.412 TB of data stored on a Dell PowerEdge R7525 server. The AI module uses YOLOv3 for object detection and AlphaPose for 17-keypoint human pose estimation. Six postures were classified: walking, standing, sitting, fallen on the ground, lying in bed, and sleeping. A pose change algorithm evaluated movement habits over 1-3 day periods using a 12.5% threshold. The decision support module used a Decision Tree with Gini impurity (max depth 3, minimum 6 samples per node), achieving 92% average classification error. Patient re-identification employed Bag of Visual Words (BOVW) with SIFT descriptors and SVM classification.

**Key Results:** The posture recognition algorithm was trained on 9,300 labeled images and tested on 3,792 images, achieving 91.63% average accuracy. Per-class F1 scores were: walking 0.9463, standing 0.8937, sitting 0.9416, fallen on the ground 0.8814, lying in bed 0.8914, and sleeping 0.8944. The macro F1 score was 0.9082 and weighted F1 score was 0.9125. AUC values ranged from 0.8790 (lying in bed) to 0.9427 (sitting). Patient re-identification accuracy was 90% (First class), 88% (Second), 91% (Third), and 87% (None/unauthorized). Real-time testing of 16 scenarios showed two prediction errors, both involving sleeping/lying in bed confusion. The system performed pose estimation with an average output time of 182 ms. The nursing care plan consisted of 61 IF-THEN rules with four possible outputs: continue current treatment, monitor, adjust, or extra situation. Combining pulse and sleep mode data from wristbands with image recognition improved sleeping and lying-in-bed detection accuracy by 15.48% and 22.06%, respectively.

**Clinical Implications:** This system addresses critical inefficiencies in geriatric care by automating vital sign monitoring and care plan adjustments. The 91.63% posture recognition accuracy enables reliable fall detection and movement monitoring, while the decision support module helps standardize care protocols. However, challenges remain: patients with dementia frequently remove wearable devices, nighttime and low-light conditions reduce detection accuracy, and patient re-identification in common areas remains problematic (best accuracy 91.4%). The authors note that full automation would require at least one year of accumulated data. The system is designed for both nursing homes and potential home care applications, with plans to develop home-based packages. The study was approved by the Kaunas Region Biomedical Research Ethics Committee (No. BE-2-24) with informed consent obtained from patients' relatives/caregivers.