Convolution Neural Network Deep Learning Model Predicts Sleep Disorder in Asthma

Of the various deep learning approaches tested, the convolution neural network (CNN) model showed the best ability to predict sleep disorder in a cohort of individuals with asthma, according to study results published in the Journal of Asthma.

This prospective study included 14,818 individuals with newly diagnosed asthma (30.52% children; 50.02% men) out of 1 million entries in the Taiwan National Health Insurance Research Database between 2002 and 2010. The individuals’ disease histories were transformed to matrices, and machine learning algorithms (support vector machine, K-nearest neighbors, and random forest) and deep learning models (CNN, gated recurrent units, long short-term memory, and recurrent neural network) were tested for predictive power. Coefficients employed included accuracy, sensitivity, specificity, and area under the curve (AUC).

Sleep disorder was diagnosed in 4469 of the 14,818 individuals with new asthma. Successful machine learning predictors of sleep disorder included the random forest (accuracy, 0.813; AUC, 0.719), support vector machine (accuracy, 0.793; AUC, 0.690), and K-nearest neighbors (accuracy, 0.798; AUC, 0.737) algorithms. Successful deep learning predictors of sleep disorder included CNN (accuracy, 0.951; AUC, 0.934), long short-term memory (accuracy, 0.815; AUC, 0.750), gated recurrent units (accuracy, 0.782; AUC, 0.732), and recurrent neural network (accuracy, 0.744; AUC, 0.658). Deep learning models showed better predictive performance than machine learning models. 

The study researchers concluded that CNN models showed the best performance in predicting sleep disorder in a cohort of individuals with asthma, with >92.3% accuracy. Further study will explore deep learning models in other serious diseases such as stroke and cancer.

Reference

Phan D-V, Yang N-P, Kuo C-Y, Chan C-L. Deep learning approaches for sleep disorder prediction in an asthma cohort [published online March 18, 2020]. J Asthma. doi:10.1080/02770903.2020.1742352