Deep learning for precision medicine

Esteban, Cristóbal (2018): Deep learning for precision medicine. Dissertation, LMU München: Fakultät für Mathematik, Informatik und Statistik

Abstract

As a result of the recent trend towards digitization, an increasing amount of information is recorded in clinics and hospitals, and this increasingly overwhelms the human decision maker. This issue is one of the main reasons why Machine Learning (ML) is gaining attention in the medical domain, since ML algorithms can make use of all the available information to predict the most likely future events that will occur to each individual patient. Physicians can include these predictions in their decision processes which can lead to improved outcomes. Eventually ML can also be the basis for a decision support system that provides personalized recommendations for each individual patient. It is also worth noticing that medical datasets are becoming both longer (i.e. we have more samples collected through time) and wider (i.e. we store more variables). There- fore we need to use ML algorithms capable of modelling complex relationships among a big number of time-evolving variables. A kind of models that can capture very complex relationships are Deep Neural Networks, which have proven to be successful in other areas of ML, like for example Language Modelling, which is a use case that has some some similarities with the medical use case. However, the medical domain has a set of characteristics that make it an almost unique scenario: multiple events can occur at the same time, there are multiple sequences (i.e. multiple patients), each sequence has an associated set of static variables, both inputs and outputs can be a combination of different data types, etc. For these reasons we need to develop approaches specifically designed for the medical use case. In this work we design and develop different kind of models based on Neural Networks that are suitable for modelling medical datasets. Besides, we tackle different medical tasks and datasets, showing which models work best in each case. The first dataset we use is one collected from patients that suffered from kidney failure. The data was collected in the Charité hospital in Berlin and it is the largest data collection of its kind in Europe. Once the kidney has failed, patients face a lifelong treatment and periodic visits to the clinic for the rest of their lives. Until the hospital finds a new kidney for the patient, he or she must attend to the clinic multiple times per week in order to receive dialysis, which is a treatment that replaces many of the functions of the kidney. After the transplant has been performed, the patient receives immunosuppressive therapy to avoid the rejection of the transplanted kidney. Patients must be periodically controlled to check the status of the kidney, adjust the treatment and take care of associated diseases, such as those that arise due to the immunosuppressive therapy. This dataset started being recorded more than 30 years ago and it is composed of more than 4000 patients that underwent a renal transplantation or are waiting for it. The database has been the basis for many studies in the past. Our first goal with the nephrology dataset is to develop a system to predict the next events that will be recorded in the electronic medical record of each patient, and thus to develop the basis for a future clinical decision support system. Specifically, we model three aspects of the patient evolution: medication prescriptions, laboratory tests ordered and laboratory test results. Besides, there are a set of endpoints that can happen after a transplantation and it would be very valuable for the physicians to be able to know beforehand when one of these is going to happen. Specifically, we also predict whether the patient will die, the transplant will be rejected, or the transplant will be lost. For each visit that a patient makes to the clinic, we anticipate which of those three events (if any) will occur both within 6 months and 12 months after the visit. The second dataset that we use in this thesis is the one collected by the MEmind Wellness Tracker, which contains information related to psychiatric patients. Suicide is the second leading cause of death in the 15-29 years age group, and its prevention is one of the top public health priorities. Traditionally, psychiatric patients have been assessed by self-reports, but these su↵er from recall bias. To improve data quantity and quality, the MEmind Wellness Tracker provides a mobile application that enables patients to send daily reports about their status. Thus, this application enables physicians to get information about patients in their natural environments. Therefore this dataset contains sequential information generated by the MEmind application, sequential information generated during medical visits and static information of each patient. Our goal with this dataset is to predict the suicidal ideation value that each patient will report next. In order to model both datasets, we have developed a set of predictive Machine Learning models based on Neural Networks capable of integrating multiple sequences of data withthe background information of each patient. We compare the performance achieved by these approaches with the ones obtained with classical ML algorithms. For the task of predicting the next events that will be observed in the nephrology dataset, we obtained the best performance with a Feedforward Neural Network containing a representation layer. On the other hand, for the tasks of endpoint prediction in nephrology patients and the task of suicidal ideation prediction, we obtained the best performance with a model that combines a Feedforward Neural Network with one or multiple Recurrent Neural Networks (RNNs) using Gated Recurrent Units. We hypothesize that this kind of models that include RNNs provide the best performance when the dataset contains long-term dependencies. To our knowledge, our work is the first one that develops these kind of deep networks that combine both static and several sources of dynamic information. These models can be useful in many other medical datasets and even in datasets within other domains. We show some examples where our approach is successfully applied to non-medical datasets that also present multiple variables evolving in time. Besides, we installed the endpoints prediction model as a standalone system in the Charit ́e hospital in Berlin. For this purpose, we developed a web based user interface that the physicians can use, and an API interface that can be used to connect our predictive system with other IT systems in the hospital. These systems can be seen as a recommender system, however they do not necessarily generate valid prescriptions. For example, for certain patient, a system can predict very high probabilities for all antibiotics in the dataset. Obviously, this patient should not take all antibiotics, but only one of them. Therefore, we need a human decision maker on top of our recommender system. In order to model this decision process, we used an architecture based on a Generative Adversarial Network (GAN). GANs are systems based on Neural Networks that make better generative models than regular Neural Networks. Thus we trained one GAN that works on top of a regular Neural Network and show how the quality of the prescriptions gets improved. We run this experiment with a synthetic dataset that we created for this purpose. The architectures that we developed, are specially designed for modelling medical data, but they can be also useful in other use cases. We run experiments showing how we train them for modelling the readings of a sensor network and also to train a movie recommendation engine.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Keywords: Machine Learning, Deep Learning, Personalized Precision, Medicine Neural Networks
• Themengebiete: 000 Allgemeines, Informatik, Informationswissenschaft ; 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik
• Fakultäten: Fakultät für Mathematik, Informatik und Statistik
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 4. Oktober 2018
• 1. Bericht­erstatter:in: Tresp, Volker