Classification algorithms trained on simple (symmetric) lifting data perform poorly in predicting hand loads during...

Area of Science:

• Biomechanics
• Ergonomics
• Machine Learning

Background:

• Machine learning (ML) algorithms are increasingly used for assessing lifting risks.
• Current ML models often lack real-world applicability due to training on limited, controlled datasets.
• Variations in lifting scenarios and postures significantly impact ML performance.

Purpose of the Study:

• To investigate the impact of different lifting scenarios on ML algorithm performance.
• To evaluate ML models trained on surface electromyography (sEMG) data for classifying hand-load levels.
• To compare the effectiveness of various training datasets and sEMG feature types.

Main Methods:

• Twelve participants performed symmetric, asymmetric, and free-dynamic lifting tasks.
• ML algorithms were trained on diverse datasets (symmetric, asymmetric, combined, free-dynamic) using sEMG data.
• Algorithms were tested using the free-dynamic dataset, simulating unconstrained lifts.

Main Results:

• Models trained on constrained datasets (symmetric lifts) showed significantly lower accuracy and sensitivity compared to those trained on naturalistic lifts.
• Frequency-domain sEMG features yielded higher accuracy, precision, and sensitivity than time-domain features.
• ML models trained on controlled lifts performed poorly on dynamic, unconstrained lifting tasks.

Conclusions:

• ML algorithms trained on controlled lifting scenarios demonstrate limited effectiveness in real-world applications.
• Training ML models with diverse, naturalistic lifting data is crucial for accurate risk assessment.
• Frequency-domain sEMG features enhance the performance of ML models for lifting analysis.