An application for classifying perceptions on my health bank in Taiwan using convolutional neural networks and...

Area of Science:

• Health Informatics
• Machine Learning Applications
• Survey Methodology

Background:

• Classifying online user opinions (positive/negative) efficiently is crucial for practical applications.
• My Health Bank (MHB) user perceptions in Taiwan were assessed using a survey integrated with Convolutional Neural Networks (CNNs) and web-based Computerized Adaptive Testing (CAT).
• This research aimed to develop an online module for accurate and efficient classification of user perceptions using CNNs and web-based CAT.

Purpose of the Study:

• To accurately and efficiently classify user perceptions of My Health Bank (MHB) into positive and negative classes.
• To integrate Convolutional Neural Networks (CNNs) with web-based Computerized Adaptive Testing (CAT) for perception classification.
• To reduce respondent burden by optimizing survey length while maintaining accuracy.

Main Methods:

• 640 MHB users (20-70 years) completed a 26-item questionnaire (PMHB26) on MHB perceptions.
• CNNs and k-means clustering were employed to classify respondents into satisfied/unsatisfied groups and build a predictive model.
• Exploratory factor analysis, Rasch model, and descriptive statistics were used to refine the model for CNN and Rasch Multidimensional CAT (MCAT).

Main Results:

• Three construct factors were identified from the PMHB26 questionnaire, demonstrating high reliability (Cronbach's alpha > 0.94).
• The CNN model achieved a high accuracy rate of 0.98 (AUC: 0.98) for classifying MHB perceptions.
• Rasch MCAT demonstrated that approximately one-third of the items were sufficient, significantly reducing respondent burden.

Conclusions:

• The combined PMHB26 CNN model and Rasch online MCAT enhance the accuracy and efficiency of classifying user perceptions of MHB.
• The developed application facilitates self-assessment of MHB value co-creation.
• This approach is adaptable for future applications in diverse survey contexts.