Strategies to optimise machine learning classification performance when using biomechanical features.

Area of Science:

• Biomechanics
• Machine Learning
• Orthopedics

Background:

• Developing prediction models with biomechanical features often requires large sample sizes, which are logistically challenging to collect.
• Limited sample sizes pose a significant hurdle in creating accurate biomechanical prediction models for healthcare applications.

Purpose of the Study:

• To investigate the efficacy of modern machine learning algorithms in overcoming sample size limitations for biomechanical prediction models.
• To compare the performance of various machine learning algorithms on biomechanical datasets with differing sample sizes.

Main Methods:

• Secondary analysis of two biomechanical datasets: a walking dataset (2295 participants) and a countermovement jump dataset (31 participants).
• Input features included three-dimensional ground reaction forces (GRFs) of the lower limbs.
• Algorithms compared: multinomial/LASSO regression, XGBoost, deep learning time-series algorithms (with data augmentation and transfer learning).

Main Results:

• For the walking dataset, top models included InceptionTime with x12 augmentation (AUC 0.810), XGBoost (AUC 0.804), and multinomial logistic regression (AUC 0.800).
• For the jump dataset, top models were LASSO (AUC 1.00), InceptionTime with x8 augmentation (AUC 0.750), and transfer learning (AUC 0.653).
• Machine learning strategies effectively addressed limited sample sizes in biomechanical data analysis.

Conclusions:

• Machine learning, particularly deep learning time-series algorithms with data augmentation and transfer learning, can effectively develop prediction models despite limited sample sizes in biomechanics.
• These ML-based strategies offer a viable approach for creating alternative prediction models in healthcare, especially when dealing with time-series biomechanical data.