Deep Learning

Our neuronal statistics

We use highly sophisticated methods of deep learning and neuronal network statistics. These methods allow an unsupervised machine-learning that can produce relevant results independent of interference of the controlling supervisor. The results are evaluated using various classification algorithms. Classification procedures are methods and criteria for classifying objects (in our case patients) into classes (here e.g. types and subtypes of prediabetics or diabetics).

The high effectiveness of our approach has been shown by our statistical results, where outcomes regularly achieve p-values below 0.0001. Areas under the ROC curve (AUC) typically are up to 0.95 and higher.

Methods

Classification is based on self-organising maps (SOMs), i.e., nonlinear, parallel, robust, fault tolerant neural networks. Self-organizing maps (SOM) are types of artificial neural networks with an unsupervised learning method with the aim of achieving a topological representation of the input space (in our case patient data). The learning algorithm automatically generates classifiers, according to which it divides the input pattern into previously unknown clusters. SOMs are used for clustering, visualizing complex relationships, prediction, evaluation, modeling and data exploration.

To further validate the models and to calculate the importance of the biomarkers for the different classification models, we use RBF networks and other predictive models like MLPs, logistic regression and Quinlan’s decision tree. Where necessary and feasible, we use adaptive boosting based on the methods of Schapire and Freund.