Fall-from-Bed Risk Prediction Using Physics-Based Bed Simulation.

Area of Science:

• Biomechanics
• Medical Informatics
• Patient Safety Engineering

Background:

• Fall-from-bed incidents pose significant risks in healthcare settings.
• Collecting real-world fall data is ethically challenging and data is scarce.
• Existing fall prediction methods often require dynamic monitoring.

Purpose of the Study:

• To investigate if static in-bed postures can predict fall-from-bed risk.
• To develop and validate a simulation-based method for fall risk assessment.
• To explore the feasibility of posture-based risk scoring.

Main Methods:

• Developed a physics-based simulator (MuJoCo) modeling bed-human interactions.
• Generated labeled fall data by simulating diverse initial postures and uncontrolled dynamics.
• Represented initial states as 2D skeletons and used time-to-fall for risk labels.
• Trained a multilayer-perceptron model on 50,000 simulated postures.

Main Results:

• The best model achieved high performance: AUC-ROC 0.9755, AUC-PR 0.9771, F1-score 0.9138.
• Static posture contained significant predictive information for fall risk.
• Performance was consistent across different patient positions (supine, prone, lateral).
• Pose-stratified initialization improved prediction accuracy.

Conclusions:

• Static in-bed posture is a viable predictor of fall-from-bed risk.
• Posture-based risk scoring shows promise for enhancing patient safety in controlled environments.
• Simulation offers a scalable solution for generating fall data ethically.