Dynamic multimode fiber specklegram sensor with automated training data generation for bed-exit prediction

Real-time prediction of bed-exit events represents a critical capability for preventing falls and ensuring patient safety in hospital and aged care settings. Existing bed-exit detection systems operate as reactive alarms rather than predictive tools, limiting their effectiveness for fall prevention interventions. In this paper, we present a proof-of-concept system that combines dynamic multimode fiber specklegram sensor technology with deep learning to achieve predictive bed-exit detection while preserving patient privacy. We integrate a multimode optical fiber into a hospital mattress to continuously monitor speckle pattern variations caused by patient movements. An automated labeling system analyzes these patterns using zero-mean normalized cross-correlation to identify bed entry and exit events, generating training data without manual annotation. The system employs principal component analysis for dimensionality reduction and trains a convolutional neural network to recognize temporal patterns that precede bed exits. Our test results demonstrate that the system can predict bed-exit events up to 30 s in advance with ±1 s accuracy. This supervised learning framework enables continuous adaptation of the model to individual patient behaviors and changing environmental conditions, addressing key limitations of existing healthcare monitoring technologies. The demonstrated ability to detect subtle pre-exit patterns imperceptible to conventional sensing approaches represents a significant advancement toward intelligent patient monitoring systems that enhance safety through predictive rather than reactive intervention strategies.