Comparison of Deep Neural Networks and Regularised Cox Regression Models in the Prediction of Neurodegenerative Diseases in the General Older Population

Background: Prediction of neurodegenerative diseases (NDDs) in the older population is crucial to prevent the apparition of NDDs eventually. Using deep neural networks (DNNs) predictive models could be an alternative to Cox regression in datasets with many predictive features. We aimed to compare the performance of DNNs with regularised Cox proportional hazards models to predict NDDs in the older general population.<br><br>Methods: We included 5,433 individuals aged 60 years and older from the English Longitudinal Study of Ageing. The features were a set of 91 baseline variables. The outcome was new events of NDDs or dementia. After applying multiple imputations, we trained five models: Elastic Net regularisation (CoxEn) and selected features (CoxSf) Cox models, Feedforward, TabTransformer, and Dense Convolutional (Densenet) network models. We compared the predictive performance of the models with Uno's C statistic and the time-dependent area under the curve (td-AUC).<br><br>Results: During follow-up, 691 participants (12.7% of the participants) developed NDDs. The performance with Uno’s C statistics (coefficient (95% confidence intervals)) was 0·757 (0·702, 0·805) (TabTransformer), 0·734 (0·694, 0·772) (CoxSf), 0·732 (0·689, 0·771) (CoxEn), 0·708 (0·653, 0·754) (FeedForward), and 0·706 (0·651, 0·752) (Densenet). The highest td-AUC were observed at year eight and were 0·894 (0·874, 0·913) (CoxSf), 0·892 (0·872, 0·912) (CoxEn), 0·883 (0·838, 0·920) (TabTransformer), 0·874 (0·849, 0·898) (Densenet), and 0·870 (0·845, 0·894) (FeedForward). The best predictors with CoxSf were age, memory function, gait speed, executive function, Body Mass Index, and, vigorous physical exercise.<br><br>Interpretation: The five models successfully predicted future NDDs events in the older general population. The discriminative ability was superior using TabTransformer compared to other models, making this structure promising for heterogeneous tabular datasets. However, due to the complexity of DNNs, Cox models are more easily understandable and interpretable.<br><br>Funding: Luxembourg Institute of Health.<br><br>Declaration of Interest: The authors declare no conflict of interest.<br><br>Ethical Approval: Ethical approval was obtained from the Multicentre Research and Ethics Committee in the United Kindom, and all participants provided written informed consent.