Subject-Independent Unsupervised Fall Detection via Masked Transformers and Energy–Reconstruction Scoring

Анотація

We present a subject-independent unsupervised fall detector for wearable inertial data. The method uses a masked Transformer encoder trained on activities of daily living only, with two scores at inference time: masked reconstruction error and a latent energy from the class token. Scores are standardized on validation ADL and fused as Sα; an Extreme Value Theory model fitted on the top 5% tail yields an operating threshold for a fixed false-alarm rate. Light temporal post-processing (median smoothing, hysteresis, refractory interval) converts scores to events. Experiments on a 29-subject dataset (8,953 sessions, 100 Hz; 6 motion channels) under leave-one-subject-out report pooled window-level ROC–AUC 0.985 ± 0.008 and PR–AUC 0.955 ± 0.012 for the fused score, exceeding reconstruction-only and energy-only variants. At a target 0.5 false alarms per hour, event sensitivity reaches 95.8%±2.6% with a mean detection delay of 1.72±0.28 s. Per-subject analysis shows compact dispersion around the FA/h target and higher sensitivity than single-cue baselines. Reconstruction MSE remains low on ADL while rising on falls, confirming that the training objective preserves signal recovery and supports anomaly scoring. The pipeline requires no fall labels for training or calibration and is suited to deployment where annotation is scarce.