Abstract— We present a Recurrent Neural Network using LSTM (Long Short Term Memory) that is capable of modeling and predicting Local Field Potentials. We train and test the network on real data recorded from epilepsy patients. We construct networks that predict multi-channel LFPs for 1, 10, and 100 milliseconds forward in time. Our results show that prediction using LSTM outperforms regression when predicting 10 and 100 millisecond forward in time. I. INTRODUCTION Local Field Potentials (LFPs) are recorded by implanting an electrode in the brain and measuring the voltage caused by the activations of the surrounding neurons within a few microns of an electrode. Because the resolution of LFP includes many neurons it cannot measure high frequency data such as single neuron spikes but often contains information in lower frequency bands below 500Hz. Interest in Local Field Potentials is rising because of their lower power costs to record, allowing continuous cortical recordings over long periods of time (years). LFPs have been part of studies that have investigated sensory processing [1, 2], motor planning [3], etc. There is also increasing evidence that manipulating specific features of the LFP can improve the symptoms of neuro-psychiatric disorders [4, 5]. LFPs show oscillatory behavior at 1-50Hz frequencies (periods of 20-1000ms), implying that the neural generators have a fluctuating state even without control inputs. Control of such oscillating systems often benefits from a model- predictive framework. By modeling the expected future state of the system with and without a given control input (brain stimulation), the output (LFP) can be more accurately controlled to the desired endpoint. We therefore sought to create accurate models of LFP dynamics as a framework for predictive control. Local Field Potentials have been modeled using a variety of methods in the past. They have been modeled using a recurrent network of inhibitory and excitatory *This work was supported by the Defense Advanced Research Projects Agency (DARPA), Biological Technologies Office (BTO), under contract number W911NF-14-2-0045. The opinions presented are those of the authors alone and not of DARPA, Draper, or the Massachusetts General Hospital. Akshay Rangamani and Sang Chin are in part supported by Air Force Office of Scientific Research grant (FA9550-12-1-0136) and National Science Foundation grant (NSF-DMS-1222567). 1 The Charles Stark Draper Laboratory, Inc, Cambridge, MA 02141 USA. (e-mail: {lkim, jharer, jmoran, pparks, schin}@draper.com). 2 Department of Electrical & Computer Engineering, Johns Hopkins University, Baltimore, MD 21218 USA. (e-mail: rangamani.akshay@jhu.edu). 3 Massachusetts General Hospital, Boston, MA 02114 USA. (e-mail: {awidge, eeskandar, ddougherty}@partners.org). 4 Department of Computer Science, Boston University, Boston, MA 02215 USA. (e-mail: spchin@cs.bu.edu). Table 1. Summary of datasets used. point-like integrate and fire neurons in [6]. A biophysical inverse current source density analysis is described in [7] to estimate the current sources that contribute to the LFPs. The biophysical method in [8] is used to model the spatial reach of LFPs and the factors it depends on. All three methods are based on biophysical models of LFPs and are motivated by discovering underlying properties such as source neurons, correlations between their firing, etc. Moreover, [6] and [8] evaluate their models on simulated data. Our approach is focused on statistical learning approach to modeling and predicting LFPs. Statistical learning approach was used in [9] to model LFPs using a linear autoregressive model. We made first attempt at using artificial neural networks to model LFPs [10]. In this work, we expand on our recent work to demonstrate that recurrent neural networks can successfully learn the patterns of LFPs relatively well and outperform regression model in predicting LFPs 10ms and 100ms forward in time. II. DATA The dataset we are using was collected using a Blackrock neural signal amplifier, connected to PMT epilepsy monitoring electrodes. Each high-density electrode consists of multiple probe points (recording channels) which measure LFPs. Recordings were taken from three different patients. Details are shown in Table 1. Recording time varies between 12 minutes to 40 minutes depending on patient. The locations and number of electrodes used for each patient vary, and as such networks were trained individually for each patient. Because of this variations Predicting Local Field Potentials with Recurrent Neural Networks* Louis Kim 1 , Jacob Harer 1 , Akshay Rangamani 2 , James Moran 1 , Philip D. Parks 1 , Alik Widge 3 , Emad Eskandar 3 , Darin Dougherty 3 , Sang (Peter) Chin 1,4 PATIENT 1 PATIENT 2 PATIENT 3 RECORDING DURATION ~12 minutes ~35 minutes ~39 minutes RECORDING LOCATIONS Dorsal anterior cingulate, dorsal posterior cingulate, ventral anterior cingulate (Patient 1 and 2 only), medial orbitofrontal (Patient 2 and 3 only), dorsolateral prefrontal cortex (DLPFC), ventrolateral prefrontal cortex (VLPFC), amygdala, hippocampus, superficial anterior temporal NUMBER OF RECORDING CHANNELS 82 98 89 STIMULATION LOCATION Amygdala Amygdala Dorsal posterior cingulate NUMBER OF STIMULATION PERIODS 69 260 265 DURATION OF A STIMULATION 400ms 400ms 400ms 978-1-4577-0220-4/16/$31.00 ©2016 IEEE 808