Prefrontal fNIRS Functional Connectivity-Based Detection of Asymptomatic Alzheimer's Disease Using Feature Engineering and Machine Learning

Abstract

Alzheimer’s disease （AD） is a biological continuum, and its asymptomatic stage, where the amyloid pathology present while cognition remains intact, is the longest and largest segment of this continuum and the key window for intervention. Confirming amyloid pathology, however, currently lies on costly and invasive amyloid PET or cerebrospinal fluid analysis, which are unsuited to large-scale screening. Functional connectivity （FC） is disrupted before symptoms appear, and among the modalities used to measure it, functional near-infrared spectroscopy （fNIRS） is distinctively portable, low-cost, and suitable for elderly and patient populations. Yet no FC-based classifier has been developed for the asymptomatic amyloid-positive stage, FC is rarely assessed across combined resting-state and task conditions, and prior work has relied on static, single-condition features. Using a portable 4-channel prefrontal fNIRS device, paired resting-state and verbal fluency task recordings were obtained from 156 participants （81 healthy controls, HC; 75 asymptomatic at-risk AD, aAD, confirmed amyloid-PET-positive）. Three biologically motivated features were engineered to capture distinct phases of the prefrontal response to cognitive demand: Task Evoked Activation （Δ1; change in FC from rest to the first phonemic trial）, Network Reconfiguration （Δ2; change at the phonemic-to-semantic transition）, and Task Tolerance （β; the trajectory of FC across trials）. Group differences were tested statistically, and nine machine-learning classifiers were evaluated under cross-validation, complemented by ablation and SHAP analyses. Eleven of the 18 features survived FDR correction, with the interhemispheric pairs ch13 and ch14 showing the largest effects; sex as not a significant covariate. An SVM-RBF classifier reached an accuracy of 0.705, sensitivity of 0.760 and AUC of 0.710, and SHAP rankings closely matched the statistical findings. To our knowledge, this is the first FC classifier for the asymptomatic amyloid- positive stage of AD and the first to combine resting-state and task conditions in features capturing the temporal trajectory of the prefrontal response. Although the cross-sectional design and moderate sample size warrant cautious interpretation, the findings support 20231209 Seyedeh Mohadeseh Shadizadeh. Prefrontal fNIRS Functional Connectivity-Based Detection of Asymptomatic Alzheimer's Disease Using Feature Engineering and Machine Learning. 전전두엽 fNIRS 기능적 연결성을 이용한 전임상 알츠하이머병 선별 기계학습모델 개발. College of Life Sciences and Medical Engineering. Department of Biomedical Science and Engineering. 2026. 41p. Prof. Jae Gwan Kim portable fNIRS connectivity as a promising, non-invasive candidate for asymptomatic AD screening. ©2026 Seyedeh Mohadeseh Shadizadeh ALL RIGHTS RSERVED