Classical machine learning on clinical and laboratory patient data

With machine learning, we were able to predict with over 80% accuracy prior to surgery, whether the stem cell treatment of post myocardial infarction patients would be successful. The identification of key features and classification of the comprehensive patient data was obtained by employing supervised and unsupervised machine learning algorithms. To identify the most important clinical features for predicting treatment outcome, we applied several supervised algorithms (AdaBoost, Support Vector Machines and Random Forest). In addition, we employed unsupervised machine learning (t-distributed stochastic neighbor embedding) which is able to cluster patients with the same treatment outcome in close proximity.

We succeeded in finding a diagnostic biomarker signature in the peripheral blood of patients by using an artificial intelligence analysis system, allowing pretreatment identification of patient responders for improved heart function. Using this new computer-aided diagnostic technology, responsive patients can be accurately identified prior to treatment with bypass surgery and stem cells.

Deep Learning on MRI images

The computer-assisted analysis for an enhanced guided interpretation of medical images have been a long standing issue. Recent advances in machine learning, especially, in the field of Deep Learning by means of Convolutional Neural Networks (CNNs), have made a big leap to help segment, identify, and quantify objects and patterns in medical images. In our project, we are using patient derived Magnetic Resonance Imaging (MRI) heart scans to evaluate the underlying cardiac functionality and the respective health status. The deep learning analyses are carried out with the TensorFlow extension Keras-Resnet (https://github.com/broadinstitute/keras-resnet).