Behavioral Change Prediction from Physiological Signals Using Deep Learned Features.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Multivariate time series analysis is crucial for predicting changes in various fields, including medicine and engineering.
• Multisensory stimulation therapy aims to modify behavioral states in dementia patients, such as transitioning from agitation to relaxation.

Purpose of the Study:

• To predict changes in patient behavioral states using physiological and neurovegetative parameters.
• To support therapists by providing timely insights during multisensory stimulation sessions.
• To compare handcrafted and learned features for predicting behavioral state changes.

Main Methods:

• Handcrafted features were derived from the CATCH22 feature collection.
• Learned features were extracted using a temporal convolutional network.
• Behavioral state prediction was performed using a joint bidirectional long short-term memory auto-encoder.

Main Results:

• The learned features-based approach demonstrated superior performance compared to handcrafted features and state-of-the-art methods.
• Accuracy rates reached up to 99.42% with a 70-second time window.
• High accuracy of 98.44% was achieved with a shorter 10-second time window.

Conclusions:

• Learned features extracted via temporal convolutional networks are highly effective for predicting behavioral state changes.
• The developed model offers a promising tool for real-time support in dementia patient therapy.
• Accurate prediction of behavioral states can enhance the efficacy of multisensory stimulation interventions.