DLPA: Deep-learning approaches for preprocessing and analysis of heterogeneous longitudinal patient data
Deep Learning denotes a strategy for efficient analysis and model development in high-dimensional, very large data sets by means of large neural networks consisting of multiple hidden layers. Shallow neural networks with one hidden layer only have proven to realize representations of any high dimensional model sufficiently, so they have been used since some decades for modeling of complex processes. However, their training from the data is hampered by their shallow network structure resulting in significantly non-monotonic nonlinear optimization problems restricting their efficiency to a restricted set of applications. The surprising efficiency of deep neural networks in applications like image recognition and speech recognition is primarily based on the finding that deep NN have to solve much smoother optimization problems than shallow ones. Hence, in these applications where the vast amount of data and computational power required for training of deep NN’s is available, they show surprising superiority with respect to alternative methods used in artificial intelligence. A thorough analysis, however, shows that their overwhelming superiority is restricted to specific classes of applications satisfying homogeneity conditions in the data structure. Moreover, today it is not possible to learn the structure of the underlying “real life” mechanism driving the biological process from the deep NN’s.
Unfortunately, in biomedical applications, real world data as well as medical diagnostic data do not exhibit the required conditions for superior performance of deep NN, whereas time series data from continuous monitoring devices as well as imaging data do. Hence, an appropriate combination of methods, adapted to the specific requirements of the heterogeneous data structure in medicine, is expected to satisfy unmet computational needs and will open a wide range of medical applications.