A stacked CNN and random forest ensemble architecture for complex nursing activity recognition and nurse...

Area of Science:

• * Computer Science
• * Healthcare Technology
• * Artificial Intelligence

Background:

• * Nursing activity recognition is crucial for smart healthcare but faces challenges like class imbalance and intra-class variability.
• * Existing methods struggle with variations in activities based on the subject and recipient.
• * Accurate recognition of both nurse actions and individual nurses is essential for effective healthcare management.

Purpose of the Study:

• * To develop a robust algorithm for simultaneous nursing activity and user recognition.
• * To overcome the limitations of class imbalance and intra-class variability in human activity recognition.
• * To enhance the accuracy and reliability of smart healthcare systems through improved activity recognition.

Main Methods:

• * A unique two-step feature extraction process incorporating 'Angle' and 'mean min max sum' features to enhance robustness.
• * Ensemble model combining a random forest classifier and a stacked convolutional neural network (S-CNN).
• * S-CNN utilizes separate input pathways for feature channels, improving robustness to intra-class variation.

Main Results:

• * Achieved highest testing accuracies for activity recognition: 70.6% (CARECOM) and 85.7% (Heiseikai).
• * Achieved highest testing accuracies for user identification: 78.2% (Dataset 1) and 92.7% (Dataset 2).
• * The algorithm automatically identifies important features within the dataset during recognition.

Conclusions:

• * The proposed method demonstrates high accuracy and robustness in simultaneous nursing activity and user recognition.
• * This approach significantly advances the capabilities of smart healthcare management systems.
• * The algorithm's ability to handle intra-class variability offers a promising solution for real-world healthcare applications.