Adaptive Deep brain stimulation (DBS) systems that adjust stimulation according to brain activity have been recently suggested to optimize DBS therapy. In the treatment of Parkinson’s disorder (PD) and
obsessive-compulsive disorder (OCD), such systems may identify of signatures of physiological and pathophysiological neural activity of the basal ganglia. However, our recent studies revealed that it is
unlikely that a single neurophysiological biomarker can be used in all patients. We hypothesize that individual patients have a unique set of neurophysiological biomarkers and individualized biomarkers are needed to accommodate stimulation in a Deep Learning DBS system.
We aim to design and employ novel deep learning methods for adaptive DBS. The premise of deep learning (DL) is that it learns the best features to efficiently represent rich signals and use their
combination to predict the desired properties. In recent years, DL-based representations for computer vision, and speech recognition proved superior to manually tuned features constructed by experts. We
hypothesize that neurophysiological data from the basal ganglia are no different, and that learned representations will outperform existing hand-tuned features constructed by clinicians. However, applying state-of-the-art DL is challenging, as the algorithms require large amounts of data. We propose to tackle this using two directions. First, we will use a multiple-task paradigm, in which we learn many different individual problems, with different parameters, but with a shared structure. Second, we will apply unsupervised deep learning methods for learning embeddings of neurophysiological signals which can potentially outperform manually tuned features. We will collect a large within-patient dataset from adults diagnosed with PD and OCD.
We will test our algorithm’s clinical utility in patients by employing a personalized closed-loop DBS – An adaptive DBS system which stimulates the brain only when the aberrant neurophysiological 
features are present. Each patient will complete neurophysiological recordings during a battery of behavioral assays probing key symptoms, and cognitive- and emotional-processes. The DL algorithm
will use individual’s data to identify neurophysiological features associated with symptoms, as well as those associated with adverse side effects. We will then configure the DBS device to initiate
stimulation, only when the aberrant features are present, and alter stimulation if side effects features are detected. Based on our preliminary data we hypothesize that each patient has a unique set of
neurophysiological biomarkers, which we can extract with DL algorithms and modify using personalized DL-DBS. DL algorithms will maximize DBS efficacy while reducing side effects, patients'
inconvenience, and clinicians’ time. DL-DBS would advance treatments of brain disorders and provide future accessibility of DBS to brain disorders that currently cannot be treated by DBS.