A Context-Aware Neural Retrieval Framework for Fall Detection in IoT-Based Elderly Care

Abstract

Wearable Internet of Things (IoT) devices produce long, continuous streams of inertial data, and the difficult part is picking out the few moments that matter. Most fall detection systems treat each segment of data as a separate yes-or-no decision. That works in a laboratory, but it gives a caregiver no help in deciding which of 50 alerts to check first. We approach fall monitoring as an information retrieval problem instead. Accelerometer and gyroscope signals are split into overlapping windows, each window is encoded with a transformer, and the system learns to rank windows by how likely they are to contain a fall. On top of the ranking we add two context signals, one measuring motion instability and one estimating impact severity, so that the most dangerous events come out on top. On the SisFall dataset the system reaches an average precision (AP) of 0.983 with a normalized discounted cumulative gain at rank 10 (nDCG@10) of 1.0, so real fall events almost always appear at the top of the ranked list, ahead of convolutional neural network (CNN), long short-term memory (LSTM), and transformer classification baselines. The result is a monitoring system that gives ranked, interpretable output instead of a stream of binary alarms and that matches how caregivers handle events in practice.