Hybrid GCN-LSTM for Privacy-Preserving Fall Detection in Human Pose-Based Elderly Monitoring Systems

Falls are a significant health risk among elderly populations, necessitating effective detection systems. Vision-based methods, particularly those using human pose estimation combined with Long Short-Term Memory (LSTM) networks, have shown promising accuracy but can struggle to capture complex spatial relationships in human motion and raise privacy concerns due to the use of sensitive video data. This paper proposes a privacy-preserving fall detection system that leverages skeletal pose data and a hybrid Graph Convolutional Network-Long Short-Term Memory (GCN-LSTM) architecture to improve recognition of falls while using only anonymized pose inputs. We evaluate our approach on a publicly available fall video dataset, comparing a standard bidirectional LSTM baseline against our proposed GCN-LSTM model. Our results indicate that the inclusion of graph-based spatial modeling yields a measurable performance improvement: the GCN-LSTM achieves higher accuracy, precision, recall, and F1-score than the LSTM-only model, with particularly notable gains in detecting falls (the minority class). In our experiments, the GCN-LSTM reached approximately 94% accuracy (versus 92.5% for the LSTM), and it substantially reduced false alarms, thereby improving the fall class precision from about 48% to 65%. These findings demonstrate that advanced spatio-temporal modeling can enhance fall detection from privacy-safe pose data, offering a robust solution for sensitive environments such as elderly care facilities where maintaining both privacy and high reliability is crucial.