INVESTIGATING THE BEST PERFORMING TASK CONDITIONS OF A MULTI-TASKING LEARNING MODEL IN HEALTHCARE USING CONVOLUTIONAL NEURAL NETWORKS: EVIDENCE FROM A PARKINSON’S DISEASE DATABASE Aggeliki Vlachostergiou *1 , Athanasios Tagaris *1 , Andreas Stafylopatis 1 , Stefanos Kollias 2 1 School of Electrical and Computer Engineering, National Technical University of Athens, Athens, Greece 2 School of Computer Science, University of Lincoln, UK aggelikivl@image.ntua.gr, thanos@islab.ntua.gr, andreas@cs.ntua.gr, skollias@lincoln.ac.uk ABSTRACT This paper presents three conditions of Multi-Task Learn- ing (MTL) model architectures based on Deep Neural Net- works (DNNs) to predict the Parkinson’s Disease (PD) from brain images. It also demonstrates the usefulness of incorpo- rating additional patients’ contextual epidemiological infor- mation (i.e. their age and sex). Our aim is to investigate which are the patients’ clinical data, that when joined with the pri- mary task could perform best and thus provide an improved computational PD prediction model. Our proposed model ar- chitectures are evaluated on a new medical dataset, which is presently under development. Our preliminary results sug- gest the robustness of our proposed systems to analyze and provide an accurate estimate of the status of the disease. Fi- nally, we discuss the lessons learned from our experimental settings with respect to addressing several research questions such as the importance of selecting the auxiliary task(s) re- spectively, which is the best performing task combination to achieve such improved Parkinson’s disease prediction, as well as whether all auxiliary tasks are equally effective. Index Terms— Convolutional Neural Networks, Health- care domain, Medical Image Analysis, Multi-task learning, context 1. INTRODUCTION Multi-Task Learning (MTL) has recently started receiving in- creasing interest due to its potential to reduce the need for labeled data and to enable the induction of more robust mod- els. One of the main drivers has been its ability to impose the relatedness of tasks to incorporate the insight from task inter- connections in joint learning [1]. Particularly, it presumes that all tasks are related and that the knowledge in each task can be shared through Deep Neural Networks (DNNs) with all other tasks. Under this view, this type of Multi-Task Learning is *Contributed equally to this work The Titan X Pascal used for this research was donated by the NVIDIA Corporation. currently considered as one of the state-of-the-art approaches in various tasks, ranging from Image Processing [2] and Com- puter Vision [3] to Speech Analysis [4] and Natural Language Processing [5]. Multi-Task Learning approaches have been also lately proposed as an appealing approach in the Healthcare domain [6, 7] due to their ability to involve many auxiliary tasks that are interrelated and complementary. For example, ex- amining the prediction of Parkinson’s Disease (PD) as a case study 1 , clinicians are required to manage multiple patients’ data at once in environments characterized by time pressure [8]. Under this context, it would be useful for them, to have a computational model that could identify the best perform- ing attributes of the patients’ clinical data, in terms of disease prediction. Therefore, it is expected that such a generated sys- tem will provide significant support to clinicians and medical doctors for early prediction of Parkinson’s. By simultane- ously learning all the tasks, the MTL architecture performs inductive knowledge transfer among tasks to improve the generalization performance of all tasks involved [1]. How- ever, to date, its task relations and ability to guarantee or make gains are poorly understood and remain unclear. Motivated by these observations, we attempt to shed light into these issues by discussing our proposed MTL architec- tures for PD prediction, (Figure 2a), in terms of why and how MTL works with DNNs. While the superiority of MTL sys- tems over the Single-Task Learning (STL) ones is evident by our experimental results, the scope of this work is to inves- tigate the impact of the auxiliary task selection on the sys- tem’s overall performance. Our model is therefore evaluated on four different conditions (1 STL and 3 MTLs), separated by task selection, where the imaging data constantly serves as our primary/main task. Unpacking the associations between 1 The importance of predicting as early as possible Parkinson’s disease (PD) is a main challenge in the field of PD therapeutics due to the fact that affects as many as one million Americans and more than 10 million individ- uals worldwide. It is also expressed through a number of symptoms such as shaking, rigidity, slowness of movement and postural instability, which how- ever begin to occur in very late stages, making very hard the prevention or the reverse of disease progress. 2047 978-1-4799-7061-2/18/$31.00 ©2018 IEEE ICIP 2018