300 Section - fNIRS-Based Early Identification of Mild Cognitive Impairment: Large-Scale Multi-Paradigm Study with Ensemble Machine Learning Models

TTo develop and validate an early screening model for mild cognitive impairment (MCI) using large-sample functional near-infrared spectroscopy (fNIRS) data, and to compare the diagnostic performance of machine-learning models built on resting-state, 1-back task, and combined fNIRS features with Montreal Cognitive Assessment (MoCA) classification.

Design:

We recruited 525 right-handed adults aged 58–87 years; after quality control, 462 participants remained (185 MCI, 277 healthy controls). MCI was diagnosed by two neurologists using Petersen’s criteria. fNIRS signals were recorded during a 5-min resting run and a 1-back working-memory task. Oxygenated hemoglobin features were extracted and reduced with Gradient Boosting; five classifiers (linear support vector machine, neural network, Naive Bayes, decision tree, k-nearest neighbor) were trained and evaluated with 10-fold cross-validation for resting-state, 1-back, and integrated datasets.

Results:

Among three datasets, the integrated resting-plus-1-back features provided the best discrimination. On the integrated dataset, the neural-network model achieved 86.49% accuracy, 94.74% sensitivity, and 77.78% specificity. For resting-state data, the k-nearest-neighbor model reached 70.27% accuracy, 72.22% sensitivity, and 68.42% specificity; for 1-back data, the decision-tree model achieved 75.68% accuracy, 78.95% sensitivity, and 72.22% specificity. MoCA-based group classification yielded 86.55% accuracy, 95.85% sensitivity, and 70.81% specificity, inferior in balanced performance to the integrated fNIRS model.


Conclusion: Large-sample, data-driven analysis of fNIRS signals integrating resting-state and task paradigms enables accurate early screening for MCI. The ensemble neural-network model based on combined fNIRS features outperformed single-paradigm models and MoCA classification, suggesting fNIRS as a promising, noninvasive and portable adjunct for identifying MCI in clinical and community settings.