Appropriate data segmentation improves speech encoding models: Analysis and simulation of electrophysiological...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Speech Processing

Background:

• Modeling neural processing of naturalistic speech offers insights into brain representations.
• Common linear encoder models assume data stationarity, which may be violated by long neural recordings.
• Non-stationary trends in neural data can impair the performance of speech encoding models.

Purpose of the Study:

• To examine the impact of non-stationary trends in continuous neural recordings on linear speech encoding model performance.
• To investigate whether data segmentation can mitigate performance impairments caused by non-stationarity.

Main Methods:

• Utilized temporal response functions (TRFs) to predict neural responses to speech.
• Segmented continuous neural data into varying lengths before model fitting.
• Tested hypotheses using both simulated stationary/non-stationary data and actual human neural recordings during story listening.

Main Results:

• For stationary data, increasing segmentation decreased model performance.
• For non-stationary data, segmentation initially improved model performance.
• Intermediate segment lengths (5-15s) optimized performance for actual neural recordings, outperforming very short (1-2s) or very long (30-120s) segments.

Conclusions:

• Data segmentation enhances the performance of encoding models for both simulated and real neural data.
• Segmentation improves performance by making shorter data segments more stationary.
• Applying encoding models to segmented neural recordings is recommended over using long, continuous segments.