Genome‐wide association study‐based deep learning for survival prediction

Abstract

Informative and accurate survival prediction with individualized dynamic risk profiles over time is critical for personalized disease prevention and clinical management. The massive genetic data, such as SNPs from genome‐wide association studies (GWAS), together with well‐characterized time‐to‐event phenotypes provide unprecedented opportunities for developing effective survival prediction models. Recent advances in deep learning have made extraordinary achievements in establishing powerful prediction models in the biomedical field. However, the applications of deep learning approaches in survival prediction are limited, especially with utilizing the wealthy GWAS data. Motivated by developing powerful prediction models for the progression of an eye disease, age‐related macular degeneration (AMD), we develop and implement a multilayer deep neural network (DNN) survival model to effectively extract features and make accurate and interpretable predictions. Various simulation studies are performed to compare the prediction performance of the DNN survival model with several other machine learning‐based survival models. Finally, using the GWAS data from two large‐scale randomized clinical trials in AMD with over 7800 observations, we show that the DNN survival model not only outperforms several existing survival prediction models in terms of prediction accuracy (eg, c‐index =0.76), but also successfully detects clinically meaningful risk subgroups by effectively learning the complex structures among genetic variants. Moreover, we obtain a subject‐specific importance measure for each predictor from the DNN survival model, which provides valuable insights into the personalized early prevention and clinical management for this disease.

Deep Learning-Based Survival Analysis for High-Dimensional Survival Data

“…Thus, the neural network algorithm can be extremely effective even if it consists of a very simple architecture such as just one single hidden layer. The DNNSurv model [7] was built by combining the DNN survival model and the regular Cox PH model, and it can be applied to the HD or ultra-HD survival datasets. The DNNSurv model was constructed as follows.…”

“…In other words, it is the extension of the linear predictor in the regular Cox PH model, and it becomes the Cox model when g(x; β) = x T β. As a result, the DNNSurv model can be used for various nonlinear covariate structures [7]. Furthermore, because of the presence of tied events, which means that more than one event occurs from different individuals at the same time, the DNNSurv model applies Efron's approach [27] to approximate the partial log-likelihood (β; x).…”

“…The multilayer deep neural network (DNN) model has recently made remarkable achievements for complex and HD cases with complete data [4][5][6]. Nevertheless, the application of deep learning to survival analysis for censored data is still limited because the existing DNN survival modeling approaches only use a single hidden layer [7]. Faraggi and Simon [8] proposed a single-hidden-layer feedforward neural network, which is usually regarded as a nonlinear extension of the Cox proportional hazards (PHs) model.…”

“…It was demonstrated that the neural network survival model performed similar to or better than the regular Cox model, the penalized Cox model, and the random survival forest model using TCGA (The Cancer Genomic Atlas) cancer data with high-throughput gene expressions. To overcome such a restriction, very recently, Sun et al [7] developed a multi-hidden-layer Cox-based DNN survival model (DNNSurv model) to predict the progression of age-related macular degeneration (AMD) disease and compared it with other survival models based on machine learning. It was shown that, in a study of AMD progression, the DNNSurv model not only outperformed several other survival models (e.g., penalized Cox model and random survival forest model) in terms of the evaluation metrics (e.g., C-index), but also successfully obtained the patient-specific predictor importance measures using the local interpretable model-agnostic explanation (LIME) method [15].…”

“…Keras contains numerous commonly used building blocks for neural networks, such as layers, activation functions, and optimizers, which makes the work much easier in terms of writing the deep neural network code. The DNNSurv model [7] is compatible with both GPUs and CPUs, via Keras and TensorFlow.…”