Beyond Load: Understanding Cognitive Effort through Neural Efficiency and Involvement using fNIRS and Machine Learning DOI: XXXXXXX.XXXXXXX Conference: CHI; 13–17 April, 2026; Barcelona ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Human-centered computing Human computer interaction (HCI) CCS: Applied comput

Shayla Sharmin, Roghayeh Leila Barmaki · 2025

Read Beyond Load: Understanding Cognitive Effort through Neural Efficiency and Involvement using fNIRS and Machine Learning DOI: XXXXXXX.XXXXXXX Conference: CHI; 13–17 April, 2026; Barcelona ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Human-centered computing Human computer interaction (HCI) CCS: Applied comput by Shayla Sharmin, Roghayeh Leila Barmaki (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-13952.txt.

Introduction

Sharmin and co-authors estimate mental effort during digital tasks from brain signals and machine learning. Effort joins how hard a task feels with how well a person performs. In clinical training, excess load ties to errors and burnout. Sixteen people did a four-part digital task with rests between parts. A headband method tracked blood flow in the front of the brain. Effort was scored through neural efficiency and involvement measures.

Those scores mix brain activity with task performance at each segment. Models then tried to guess effort for each person. The study asks if effort shifts across task phases in a readable way. A working gauge could warn trainees before overload hits.

Catalog record

CREATORShayla Sharmin, Roghayeh Leila Barmaki
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-13952

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Beyond Load: Understanding Cognitive Effort through Neural Efficiency and Involvement using fNIRS and Machine Learning DOI: XXXXXXX.XXXXXXX Conference: CHI; 13–17 April, 2026; Barcelona ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Human-centered computing Human computer interaction (HCI) CCS: Applied computing Interactive learning environments CCS: Applied computing Learning management systems CCS: Applied computing Computer-assisted instruction Shayla Sharmin, Roghayeh Leila Barmaki Abstract . The estimation of cognitive effort could potentially help educators to modify material to enhance learning effectiveness and student engagement. Where cognitive load refers how much work the brain is doing while someone is learning or doing a task cognitive effort consider both load and behavioral performance. Cognitive effort can be captured by measuring oxygen flow and behavioral performance during a task. This study infers cognitive effort metrics using machine learning models based on oxygenated hemoglobin collected by using functional near-infrared spectroscopy from the prefrontal cortex during an educational gameplay. In our study, sixteen participants responded to sixteen questions in an in-house Unity-based educational game. The quiz was divided into two sessions, each session consisting of two task segments. We extracted temporal statistical and functional connectivity features from collected oxygenated hemoglobin and analyzed their correlation with quiz performance. We trained multiple machine learning models to predict quiz performance from oxygenated hemoglobin features and achieved accuracies ranging from 58%–67% accuracy. These predictions were used to calculate cognitive effort via relative neural involvement and efficiency, which consider both brain activation and behavioral performance. Although quiz score predictions achieved moderate accuracy, the derived relative neural efficiency and involvement values remained robust. Since both metrics are based on the relative positions of standardized brain activation and performance scores, even small misclassifications in predicted scores preserved the overall cognitive effort trends observed during gameplay. Figure 1 . The overview of the proposed system (a) fNIRS data collected during quiz gameplay; (b) signals pre-processed and \Delta HbO extracted; (c) statistical and connectivity features derived; Temporal feature engineering applied (d) ML models predict quiz score; (e) predicted score and \Delta HbO used to estimate cognitive effort via relative neural efficiency and involvement. 1. Introduction Cognitive effort (CE) refers to the level of cognitive load which is an individual experiences while learning ( Getchell and Shewokis 2023 ; Koiler et al. 2022 ) . Cognitive load refers to how much the brain is working during a task ( Getchell and Shewokis 2023 ; Shewokis et al. 2015 ; Hasan et al. 2023 ) . In many cases, we see that cognitive load becomes high, but this increase alone does not clearly tell us whether the person is deeply involved in the task or struggling with it. High brain activity can occur in both situations ( Getchell and Shewokis 2023 ; Shewokis et al. 2015 ; Hasan et al. 2023 ) . This is where the idea of cognitive effort becomes useful. Unlike cognitive load, which only shows how much the brain is working, cognitive effort considers both brain activation and performance together ( Getchell and Shewokis 2023 ; Koiler et al. 2022 ; Sharmin et al. 2024c ; Sharmin et al. 2025 ) . When both brain activity is high and performance is also good, we interpret that as involvement. But if the brain is working hard and performance is still poor, it likely reflects burden. In other words, CE helps us differentiate between productive effort and mental overload. Even when performance looks efficient, CE can reveal whether that efficiency required excessive mental strain. By combining brain load and performance, we get a clearer picture of whether the effort was manageable or overwhelming ( Sharmin et al. 2025 ; Sharmin et al. 2024c ; Getchell and Shewokis 2023 ; Koiler et al. 2022 ) . CE is associated with increased neural activity during problem-solving, decision-making, and learning ( Getchell and Shewokis 2023 ; Koiler et al. 2022 ) . When students are learning new or complex tasks, measuring CE helps us understand how hard their brains work and how to improve the design of the educational content. Relative Neural Efficiency (RNE) and Relative Neural Involvement (RNI) are two key metrics that can be used to explain CE with behavioral performance ( Getchell and Shewokis 2023 ; Koiler et al. 2022 ; Reddy et al. 2022 ) . These metrics are derived from hemodynamic responses in the prefrontal cortex (PFC), via functional Near-Infrared Spectroscopy (fNIRS), alongside with behavioral performance. The PFC is responsible for executive functions such as decision-making, and problem-solving, which are critical in educational contexts. RNE indicates how efficiently an individual processes information and performs a task with minimal CE. Higher RNE indicates low effort and high performance, and lower RNE suggests high effort and low performance. On the other hand, RNI measures motivation and engagement. High RNI indicates the user is more engaged and motivated during a task ( Paas et al. 2003 ; Shewokis et al. 2015 ; Getchell and Shewokis 2023 ; Koiler et al. 2022 ) . A balanced RNE and RNI represents individual’s engagement and efficiency during a task. Thus, understanding RNE and RNI may help to optimize learning strategies, improve engagement, and enhance cognitive performance. Traditional methods of evaluating student efficiency and involvement rely on test assessments, self-reports, or reaction times, which are often subjective and inconsistent ( De La Cruz et al. 2022 ) . A few recent studies used hemodynamic responses to understand learning efficiency along with subjective measures and cognitive load ( Pan et al. 2020 ; Jeun et al. 2022 ; Oku and Sato 2021 ) . However, CE prediction using fNIRS and machine learning models remains an open research challenge. By applying machine learning models to estimate performance from fNIRS signals, we evaluate the possibility of estimating RNE and RNI, allowing the detection of efficient learning. Our work introduces a novel dataset, that captures both fNIRS hemodynamic responses and quiz performance score during an educational quiz game from sixteen participants. We predicted performance score using various AI models. Based on this and fNIRS data we observed cognitive effort by estimating relative neural efficiency (RNE) and involvement (RNI). Instead of optimizing only for prediction accuracy, this study emphasizes interpreting brain signal patterns to better understand cognitive effort. Our research questions (RQs) are as follows: RQ_{1} Can machine learning models accurately predict performance scores from fNIRS-based features during educational gameplay? RQ_{2} Can predicted performance scores from ML model reflect cognitive effort meaningfully through derived RNE and RNI metrics? The paper is organized as follows. section 2 reviews relevant literature. section 3 outlines our experimental methodology, detailing participants, apparatus, study procedure, cognitive effort measurement methods, data acquisition procedures, pre-processing techniques, and feature extraction frame work. section 4 presents results. section 5 discusses findings about the research questions, addressing practical implications, limitations, and possible improvements. Finally, section 6 summarizes the study’s findings and highlights its educational contributions. 2. Related Works fNIRS and Machine Learning in Brain Signal Analysis: Machine learning models have been widely used on fNIRS-derived hemodynamic responses to support brain computer interface (BCI) tasks such as speech decoding ( Cooney et al. 2022 ) , motor imagery ( Chiarelli et al. 2018 ; Ma et al. 2021 ) , and finger movement classification ( Khan et al. 2021 ; Ortega and Faisal 2021

Read the complete text