# Estimating cognitive workload in robot assisted surgery using time and frequency features from EEG epochs with random forest regression

## TL;DR

This paper presents a machine learning method using EEG data to accurately estimate cognitive workload during robot-assisted surgery.

## Contribution

A novel framework using time and frequency features from EEG epochs with random forest regression to estimate cognitive workload in RAS.

## Key findings

- Random Forest Regressor achieved high accuracy in predicting region-specific cognitive workload scores with R2 values up to 0.9947.
- Key predictors of cognitive workload included kurtosis, RMS, and specific power spectral bands.
- The model outperformed SVR, Linear Regression, and XGBoost in generalizability and robustness.

## Abstract

Cognitive workload (CW) refers to the mental effort required to perform a task and is critical to monitor in high-stakes environments such as robot-assisted surgery (RAS), where excessive demand can impair decision-making and performance. This study proposes a machine learning framework to estimate CW using electroencephalography (EEG) signals, focusing on four cortical regions: frontal, temporal, parietal, and occipital. EEG epochs were processed to extract both time-domain features (mean, variance, skewness, kurtosis, RMS, zero-crossings) and frequency-domain features (power spectral density across delta, theta, alpha, beta, and gamma bands). To enhance computational efficiency, data were downsampled from 500 to 128 Hz, with minimal signal degradation confirmed via topographic and spectrogram-based comparisons. Random Forest Regressor (RFR) was trained to predict region-specific EEG-derived CW scores, achieving high accuracy with R2 (coefficient of determination) values of 0.9947 (temporal), 0.9692 (parietal), 0.9635 (occipital), and 0.9329 (frontal), alongside low RMSE scores. Feature importance analysis identified kurtosis, RMS, and select power bands as key predictors. Model robustness was validated using tenfold cross-validation and statistical significance testing (p < 0.0001). Comparative evaluation with SVR, Linear Regression, and XGBoost confirmed the superior generalizability of the RFR model. Topographic EEG maps and time–frequency spectrograms visually supported region-specific activation patterns, reinforcing the effectiveness of spatially localized workload modeling. These findings demonstrate a promising, interpretable, and high-performing pipeline for EEG-based cognitive workload estimation, with broad implications for adaptive neuroergonomic systems in surgical and clinical settings.