Citation

Haque, Radiah and Ho, Sin Ban and Chai, Ian and Abdullah, Adina (2023) Improved Adam-based Feedforward Deep Neural Network Model for Personalized Asthma Predictions. Journal of System and Management Sciences, 13 (2). pp. 241-257. ISSN 1816-6075, 1818-0523

Text Vol.13.2.17.pdf - Published Version Restricted to Repository staff only Download (402kB)

Abstract

A Feedforward Deep Neural Network (FDNN) model contains densely connected layers where backpropagation is applied to calculate the loss function gradients. Optimising the network weights is important to minimise the loss value, hence decreasing the prediction errors and increasing the accuracy rate. Optimisers are used to update the weight values or the learning rate for each weight. Recent studies show that, although Adaptive Moment Estimation (Adam) produces better results in terms of optimising the parameters of the FDNN model, it might lead to poor generalisation performance. Therefore, in this paper, an improved Adam-based FDNN model was built for personalised predictions of asthma. Data transformation techniques (standardisation and normalisation) and regularisation techniques (dropout and max-norm constraint) were applied. Several experimental models were trained, and their prediction performance were compared. The empirical findings reveal that the best prediction results with low loss value can be obtained when the model is trained with standardised inputs and normalised outputs. Moreover, applying dropout (p=0.1) with max-norm (c=3) minimises the generalisation error of the model effectively. The results also show that the improved Adam-based FDNN model (with 2 hidden layers and 50 hidden nodes) produces better performance results with lower prediction loss (Mean Absolute Error (MAE)=0.0409, Mean Squared Error (MSE)=0.0038, and Root Mean Squared Error (RMSE)=0.0618) and higher accuracy rate (96%) than Stochastic Gradient Descent (SGD), Root Mean Squared Propagation (RMSProp), and Adaptive Gradient Descent (AdaGrad). Consequently, the proposed model can be used for personalised asthma predictions based on demography and weather.

Item Type:	Article
Uncontrolled Keywords:	Feedforward Deep Neural Network, Adam Optimiser, Data Transformation, Regularisation.
Subjects:	Q Science > QP Physiology > QP351 Neurophysiology and Neuropsychology
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	04 Jul 2023 03:23
Last Modified:	04 Jul 2023 03:23
URII:	http://shdl.mmu.edu.my/id/eprint/11526

Downloads

Edit (login required)