Explorative Analysis and Data Mining in Big Neural Data

Abstract: Most motor actions are carried out by many neurons acting together in different areas of the brain. Recently, techniques have been developed and increasingly used that allow simultaneous recordings of the activity from a large number of neurons over time ranges as long as several weeks. At Neuronano Research Center (NRC) at Lund University, neural activity in rodents has been recorded using chronically implanted 128-channel electrodes. The targets are within areas that play a prominent role in the planning, monitoring and execution of movements and, consequently, are strongly affected by motor diseases such as Parkinson’s disease. The experiments conducted at NRC naturally produces a large amount of neural data. This master thesis work aims to perform feature extraction from that data, i.e., find a way to mathematically describe the disease states of the rodents without a priori knowledge of the states. This is done by extracting features from the recordings, whose values later are used to analyse and cluster the data. The results show that a clear majority of the neurons exhibit a significant difference in feature values between the disease states. This means that it is possible to mathematically describe the different disease states. Moreover, the results show that different neurons behave differently: Some, e.g., exhibit increased activity going from one state to another, while others exhibit decreased activity. Adding to this, some does not exhibit any change while others exhibit a significant change. That it is possible to mathematically describe the different disease states, with a timescale of hours, indicate that the same may be possible for states with smaller timescales. These states, with a smaller timescale, do not have to be connected to Parkinson’s disease but could rather be normal, healthy states such as locomotion or reaching.