What dystonia can teach us about the normal function of the basal ganglia

Lecturer: Prof. Terence D. Sanger, MD PhD, Lee/Ramo Chair in Health Science and Technology, University of Southern California

Abstract: Advances in Deep Brain Stimulation (DBS) techniques now allow us to record simultaneously from up to 120 microelectrode channels in GPi and motor thalamus in awake unrestrained children. Recordings in 20 children with secondary dystonia reveal very low neural firing rates at rest (<1hz) that increase with movement (5-10hz) in both GPi and thalamus. Furthermore, pairs of neurons in pallidum show highly significant spike coincidences with sub-millisecond precision even between neurons that can be millimeters apart.  While this pattern likely reflects widespread diffuse injury to basal ganglia, it also provides an opportunity to examine the behavior of a sparse interconnected network of surviving neurons. 

If cells fire rarely but with precise relative spike timing, this permits an effective digital code, in which spatial patterns rather than individual cells transmit information.  Given an estimated 100 million neurons in GPi for instance, it is theoretically possible to allocate 10,000 unique two-neuron patterns to every millisecond of a human life.

This suggests a new model of basal ganglia function, in which the basal ganglia-thalamus circuit memorizes every state-output mapping that has ever occurred in the person’s life.  Dopamine, acetylcholine, oxytocin, substance P, endorphins, and other neuromodulators determine whether each motor memory is rewarding, risky, important, unimportant, social, avoidant, painful, or pleasurable.  Future behavior is modified by selection and replay of motor memories based on current state and desired outcome.  The output of the basal ganglia circuit modulates cortical excitability in order to selectively enhance the desired sensory-motor dynamics.  The computational model is closest to a radial basis-function network with nodes allocated at each datapoint, augmented by assigned valence from neuromodulators. 

This model explains important behavioral observations that have been difficult to reconcile with standard computational models based on recursive learning rules:  (1) one-shot learning, (2) lack of interference between different motor memories, (3) lack of motor forgetting, (4) poor generalization of complex motor skills outside the conditions in which they were learned.  Through simulations, the effectiveness of this surprisingly elegant brute-force learning algorithm will be demonstrated.  The result of injury will also be demonstrated and the effect on previously learned skills and the ability to acquire new skills is consistent with psychophysical results from human patients with dystonia.

CENTRO CONGRESSI