# fNIRS-based early identification of mild cognitive impairment: a large-scale multi-paradigm study with ensemble machine learning models

## TL;DR

This study uses fNIRS and machine learning to improve early detection of mild cognitive impairment by combining data from multiple brain activity paradigms.

## Contribution

The novel approach integrates resting-state and 1-back task fNIRS data with ensemble machine learning for more accurate MCI screening.

## Key findings

- An integrated fNIRS dataset with machine learning achieved 86.49% accuracy in identifying MCI.
- Combined resting-state and 1-back task features outperformed single-paradigm models in diagnostic performance.
- The Neural Network model on the integrated dataset showed high sensitivity (94.74%) and AUC (93.49%).

## Abstract

Early and accurate identification of mild cognitive impairment (MCI) is crucial for timely intervention and preventing further cognitive decline. Functional near-infrared spectroscopy (fNIRS) is a non-invasive, portable tool for clinical screening, but its diagnostic accuracy is often constrained by single-paradigm approaches and small sample sizes. To address this limitation, this study aimed to develop and validate an efficient early MCI screening model by integrating large-sample fNIRS data from resting-state and 1-back task paradigms using ensemble machine learning, thereby enhancing the accuracy and reliability of early MCI diagnosis.

A total of 462 right-handed participants (185 MCI patients and 277 healthy controls, aged 58 -87 years) were included in the final analysis after screening, with MCI diagnosis jointly determined by two experienced neurologists based on Petersen’s criteria. fNIRS signals were collected during resting-state and 1-back task sessions; after preprocessing in MATLAB, features were extracted from oxygenated hemoglobin (HbO) signals of both paradigms.

Feature selection was performed via a gradient boosting classifier based on feature importance scores, resulting in 108 selected features. Five classifiers were trained and evaluated using 10-fold cross-validation. The integrated dataset combining resting-state and 1-back task features outperformed the single-paradigm datasets: the Neural Network model on this integrated dataset achieved an accuracy of 86.49%, sensitivity of 94.74%, specificity of 77.78%, and Area Under the Curve (AUC) of 93.49%. In contrast, the Nearest Neighbor model on the resting-state dataset and the Decision Tree model on the 1-back task dataset yielded accuracies of 70.27% and 75.68%, respectively. Group classification using MoCA scores achieved an accuracy of 86.55%, which was comparable to single-paradigm machine learning models but inferior to the integrated model.

This study demonstrates the value of a large-sample, data-driven approach and multi-paradigm feature integration in fNIRS-based MCI screening, providing an efficient diagnostic model for clinical application.

https://www.chictr.org.cn/showprojEN.html?proj=192047.