A Quantum Federated LSTM Approach for Fall Detection With Wearable IoT Devices

Falls are common among older adults, but wearable fall-detection models often perform well on the dataset on which they are trained, yet lose accuracy when the user, device, or sensor location changes. This article studies a privacy-preserving approach that learns from distributed wearable data while still adapting to each person. We propose a hybrid quantum long short-term memory (QLSTM) model and train it in a three-stage workflow: a public pretraining step to learn general motion patterns, federated training to learn from multiple users without sharing raw data, and a lightweight personalization step for each client. To support this setting, we collect a new inertial dataset with multiple body locations and time-annotated fall events, and we align its activity labels with common public benchmarks to enable consistent evaluation and facilitate federated transfer learning. Results show that the QLSTM transfers better than a classical LSTM when applied to new users and unseen sensor placements, achieving nearly 90% average accuracy after federated training and personalization. Compared with the traditional machine-learning (ML) approach, the proposed quantum ML (QML) model also uses fewer trainable parameters, which reduces the size of updates exchanged during federated training.