A computational approach using intracranial EEG signals to predict memory performance outcomes
This project investigates how neurons in the human hippocampus and amygdala contribute to recognition memory through analyzing spatiotemporal firing patterns. Using advanced machine learning techniques applied to intracranial electroencephalography (iEEG) data (Faraut et al., 2018), I examined whether neural activity patterns could predict memory performance on both trial-by-trial and individual difference levels.
Can spatiotemporal firing patterns across neurons in the hippocampus and amygdala predict recognition memory success on a trial-by-trial basis?
Do these neural patterns contribute to memory at both encoding and retrieval stages?
Can differences in these neural firing patterns explain individual variations in recognition memory performance?
This study represents a significant advance over traditional single-unit analyses through:
Spatiotemporal pattern analysis: Rather than examining individual neurons in isolation, I analyzed distributed patterns of activity across multiple neurons over time
Machine learning approach: Employing elastic net regularization with nested cross-validation to handle sparse neural data
Dual-phase analysis: Examining both encoding and retrieval periods to understand memory processes comprehensively
The approach leverages a rare and valuable dataset of 1,576 human hippocampal and amygdala neurons recorded across 65 sessions during a recognition memory task.
The analysis revealed several important insights:
Predictive neural states: Spatiotemporal firing patterns in both the hippocampus and amygdala can predict, with significant accuracy, whether a visual stimulus will be successfully remembered
Encoding-retrieval contributions: These predictive patterns were present during both encoding and retrieval phases, suggesting these structures contribute to memory at multiple stages
Individual differences: The differentiation between successful and unsuccessful memory states correlated strongly with individuals' recognition memory performance, with better performers showing more distinct neural firing patterns
Multiple firing profiles: Data-driven clustering revealed diverse temporal firing patterns associated with memory success, with different profiles distributed unequally between high and low perform
This research provides crucial insights into the neural basis of recognition memory by:
Establishing a mechanistic link between MTL neuronal activity and memory performance
Demonstrating that memory success is represented by network-level states rather than simple firing rate increases
Suggesting that individual differences in memory ability may be linked to the quality of neural state differentiation
Providing a potential framework for understanding memory decline in aging and neurodegenerative conditions
These findings extend our understanding beyond the traditional necessity of these structures for memory, revealing how their specific activity patterns contribute to memory success or failure.
This work is currently under review for publication. The analysis was conducted on a publicly available dataset generously shared by Rutishauser and colleagues, allowing for innovative reanalysis of valuable human intracranial recording data.
While most of my research examines how external factors like music modulate memory, this project provides complementary insights into the internal neural mechanisms supporting memory formation and retrieval. Understanding these fundamental mechanisms helps contextualize how environmental factors might influence memory by modulating these neural states.