Rare access to patients with intercranial implants gives some insight into how memories are formed
The work conducted by Dr Jessica Schrouff under the DecoMP_ECoG project looked into how memories are formed. The research used intracranial electrophysiological recordings from the surface of the human brain to investigate encoding, retrieval and consolidation of category-specific information. It is difficult to study how memories are formed directly. As Dr Schrouff explains: “Typically subjects see some materials, and later a researcher will test them on their knowledge of the materials. However, when exactly was the memory formed? How was a representation of the materials constructed in the brain and where? These questions have been mostly studied indirectly in humans, for example when people were remembering the materials.” “During my PhD, I investigated these questions using functional Magnetic Resonance Imaging (fMRI), but the temporal resolution was low (6-12 seconds). At Stanford, I had the chance to work with intracranial electrodes, which allows us to look at brain signals at a millisecond scale and with precise anatomical resolution.” Intracranial recordings in humans are relatively scarce as these are very invasive. In DecoMP_ECoG, undertaken with the support of the Marie Curie programme, patients were referred to hospital with drug-resistant epilepsy. Electrodes were then implanted on the surface of their brain to locate the source of the epilepsy. Patients would keep the electrodes in their heads for a few days, and the doctors would wait for seizures to happen. If possible to do so safely, the 'diseased' brain area would then be removed to try eliminating the source of the seizures. “During their stay at the hospital with the implanted electrodes (typically 7 to 10 days), some patients volunteered to undergo cognitive testing. Our team performed different cognitive tasks at the patient's bedside. For example, we investigated how numbers are perceived in the brain or, in my case, how memories are formed and ‘travel’ in different areas of the brain.” The downside of such recordings is that the population of epileptic patients was very heterogeneous: some would find the task too easy and learn 'too quickly', others would never learn the materials. In addition, the electrodes were placed in regions of interest for clinical purposes, which means that they can carry epileptic signal that heavily pollute signals elicited during learning. “While this was a great learning experience, it limited the statistical power of our results,” says Dr Schrouff. Fortunately, this limitation did not prevent the investigation of novel analysis techniques. In her work, Dr Schrouff focused on using machine learning techniques, i.e. models that learn to perform a task given some example data. The task was to predict what exact materials were presented to the patient. She says: “We displayed a sequence of images of faces, animals and words. The example data was the brain signals elicited during the presentation and the model would predict whether the patient was viewing a face, an animal or a word given a unique brain signal.” Such models have been used in neuroimaging but had been scarcely studied for intracranial recordings. In addition, the relationship between how a model analyses the example data and how the brain analyses the same data is still under debate. During the fellowship, Dr Schrouff investigated this relationship in depth. “I feel this work has contributed to the literature but also to the awareness of machine learning model users. I have also implemented my work in an open-source software PRoNTo, that will be released soon.”
Keywords
DecoMP_ECoG, cognitive function, memory, intracranial electrode recordings, machine learning, interpretability
