Optimised deep neural network model to predict asthma exacerbation based on personalised weather triggers

Area of Science:

• Computational Health Informatics
• Machine Learning in Medicine
• Environmental Health

Background:

• Existing mHealth apps for asthma self-management lack accurate prediction of exacerbations.
• Personalized predictions using weather triggers and demographics are needed for tailored user responses.
• This study addresses the gap by proposing an optimized Deep Neural Network Regression (DNNR) model.

Purpose of the Study:

• To develop and optimize a DNNR model for predicting asthma exacerbation.
• To integrate weather, demographic data, and asthma tracking into an mHealth application.
• To improve the accuracy and efficiency of asthma exacerbation prediction for self-management.

Main Methods:

• Developed an mHealth application incorporating weather, demographics, and Asthma Control Test (ACT) scores.
• Utilized a dataset of 1010 ACT scores from 10 users, including 5 weather and 5 demographic features.
• Applied an optimized DNNR model with standardisation and fragmented grid-search, using the Adam optimizer.

Main Results:

• Initial DNNR model achieved 83% accuracy with MAE=1.44 and MSE=3.62.
• Optimized DNNR model achieved 94% accuracy with MAE=0.20 and MSE=0.09, meeting acceptable loss range (<0.5).
• The optimized model demonstrated significantly higher accuracy and reduced computing time compared to existing models.

Conclusions:

• This is the first study to demonstrate the potential of DNNR in correlating asthma, weather, and demographic variables.
• The optimized DNNR model offers a highly accurate and efficient prediction tool for asthma self-management.
• The developed model can be integrated into mHealth applications to enhance user care and asthma control.