A Novel Framework Based on Deep Learning Architecture for Continuous Human Activity Recognition with Inertial Sensors

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Clinical Informatics

Background:

• Human Activity Recognition (HAR) frameworks are valuable for monitoring patient motor and functional abilities in clinical settings.
• Deep Learning (DL) models can process raw data for HAR, reducing the need for manual feature engineering.
• Existing DL-based HAR research often lacks focus on continuously executed actions, highlighting a gap in real-world applicability.

Purpose of the Study:

• To design, develop, and test a Deep Learning (DL) based workflow for Continuous Human Activity Recognition (CHAR).
• To address the limitations in current literature concerning HAR frameworks for ongoing motor actions.
• To evaluate the potential clinical utility of the proposed CHAR framework.

Main Methods:

• Development of a DL-based workflow specifically for CHAR.
• Training the model on data from ten healthy subjects.
• Testing the framework's performance on eight different subjects to assess generalization.

Main Results:

• The proposed DL framework demonstrated the capability for accurate motor action classification.
• The system achieved classification within a feasible timeframe, indicating real-time potential.
• Successful testing on subjects distinct from the training set suggests generalizability.

Conclusions:

• The developed DL workflow for CHAR is capable of accurately classifying motor actions in a timely manner.
• The framework shows promise for application in clinical scenarios, particularly for remote patient monitoring.
• Further validation with larger datasets is warranted, but initial results are encouraging for continuous activity recognition.