Benchmarking Time-Series Artificial Intelligence Architectures for Wearable Sensor-Based Fall Prediction: A Synthetic...

Area of Science:

• Gerontology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Falls in older adults are a significant cause of injury and loss of independence.
• Current fall detection systems often lack early warning capabilities, detecting falls only after they occur.

Purpose of the Study:

• To develop and present a synthetic benchmarking framework for early fall-risk prediction.
• To compare the performance of classical and temporal machine learning architectures for early fall detection.

Main Methods:

• Generated a synthetic dataset of 1000 sequences simulating normal activity, slips, and pre-fall instability using biomechanical, physiological, and contextual data.
• Trained and evaluated eight baseline models and two augmented temporal variants using subject-wise splits to prevent data leakage.
• Assessed model performance using a realistic evaluation protocol, focusing on early-warning capabilities and alert triggering.

Main Results:

• Classical machine learning baselines achieved superior macro-F1 scores compared to temporal models.
• Early-warning performance varied significantly; some models failed to trigger alerts, while others increased false alarms with higher pre-fall trigger rates.
• Model performance was highly dependent on the data partitioning strategy, calibration methods, and alert design.

Conclusions:

• The proposed synthetic benchmarking framework offers a reproducible method for evaluating early-warning fall-risk prediction models.
• Findings highlight the critical role of calibration and alert design in the operational effectiveness of fall prediction systems.
• Further validation with real-world data and deployment-oriented strategies is recommended.