Principal component analysis can decrease neural networks performance for incipient falls detection: A preliminary...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Gerontology

Background:

• Fall-related accidents pose significant risks to the elderly population.
• Effective fall detection systems are crucial for healthcare and personal safety.
• Blind Source Separation (BSS) is often used in fall detection preprocessing, but its impact on classification is unclear.

Purpose of the Study:

• To investigate the effects of Principal Component Analysis (PCA) and Independent Component Analysis (ICA) on artificial neural network (ANN) performance for incipient fall detection.
• To characterize how these BSS methods influence the reliability and speed of fall recognition.

Main Methods:

• Utilized 3D kinematic data from the feet and arms of subjects experiencing unexpected walking perturbations.
• Applied PCA and ICA, individually and in combination (PCA + ICA), as preprocessing steps.
• Evaluated the performance of an ANN classifier on the processed data for incipient fall detection.

Main Results:

• PCA, particularly with an 85% residual variance threshold, significantly degraded ANN classifier performance in fall detection.
• The combination of PCA followed by ICA (PCA + ICA) restored the classifier's performance.
• The effectiveness of BSS methods, like PCA, can be dataset-dependent and may adversely affect nonlinear classifiers.

Conclusions:

• The application of PCA requires careful consideration in fall detection to avoid performance degradation.
• Subsequent ICA application can mitigate the negative effects of PCA on ANN-based fall detection.
• Linear BSS techniques may interact unpredictably with nonlinear classifiers, highlighting the need for tailored preprocessing strategies.