Development of kernel deep stacking networks for improved medical diagnosis and prognosis

核深度堆积网络(KDSN)是一类有监督的深度学习方法，越来越多地被用于生物医学诊断和预测。例如分析视网膜图像以预测眼科疾病的进展，预测组织中的剪接模式，以及神经疾病的分类/预测。与许多其他深度学习方法相比，KDSN的特点是运行时间大幅减少，这是因为KDSN拟合不是基于反向传播算法，而是基于一组顺序堆叠的、容易解决的核岭回归问题。由于KDSN的效率，生物医学研究人员可以应用基于深度学习的架构，而不必依赖复杂的基于云或GPU的IT解决方案的可用性。在以前的工作中，我们在R中实现了KDSN方法，并开发并发布了一个数据驱动的KDSN调优过程。本项目的重点是KDSN的几个高度相关的扩展，以解决目前限制该方法在生物医学应用中广泛使用的问题。更具体地说，该项目的一揽子工作将包括制定、实施和分析(1)变量选择方法、(2)延长至活动的时间结果、(3)降维技术、(4)集合方法和(5)在KDSN中的辍学正规化。除模拟研究外，还将测试所有方法学发展在生物医学实践中的适用性，包括高维视网膜图像分析和纵向流行病学研究数据的分析。

Kernel deep stacking networks (KDSN) belong to the class of supervised deep learning methods, which are increasingly used for biomedical diagnosis and prognosis. Examples are the analysis of retinal images to predict disease progression in opththalmology, the prediction of splicing patterns in tissues, and the classification/prediction of neurological disorders.Compared to many other deep learning methods, KDSN are characterized by massively reduced run times, which is due to the fact that KDSN fitting is not based on the back-propagation algorithm but on a set of sequentially stacked and easy-to-solve kernel ridge regression problems. Due to the efficiency of KDSN, it is possible for biomedical researchers to apply deep-learning-based architectures without having to rely on the availability of sophisticated cloud- or GPU-based IT solutions. In previous work, we have implemented the KDSN method in R and have developed & published a data-driven procedure for KDSN tuning.The focus of this project is on several highly relevant extensions of KDSN, addressing issues that currently limit the widespread use of the method in biomedical applications. More specifically, the work packages of the project will include the development, implementation and analysis of (i) variable selection methods, (ii) extensions to time-to-event outcomes, (iii) techniques for dimension reduction, (iv) ensemble methods and (v) drop-out regularization in KDSN. In addition to simulation studies, all methodological developments will be tested with regard to their applicability in biomedical practice, including high-dimensional retinal image analysis and the analysis of longitudinal epidemiological study data.