Deep Learning Model Improves Detection of Mild Cognitive Impairment via fNIRS Data

A recent collaboration between researchers at Pusan National University, Qingdao University, and the Washington University School of Medicine in St. Louis tested a new deep learning approach to evaluate its effectiveness at diagnosing mild cognitive impairment (MCI). This study, which was published in the journal Biocybernetics and Biomedical Engineering, demonstrates how by using the full-time series of functional near-infrared spectroscopy (fNIRS) signals, researchers can bypass data segmentation and feature extraction steps to identify MCI with high accuracy (1). As a result, this new method improves clinical diagnosis of MCI and makes the process more efficient, improving patient outcomes.

What is MCI?

Mild cognitive impairment (MCI) is a disease that mostly afflicts older patients. Seen as the stage before dementia, MCI is often the cause of other brain conditions (2). Symptoms include memory lapses and a decreased mental ability that impacts the patient’s ability to process language and make sound judgments (2). However, patients with MCI, unlike those with dementia, are often aware enough to recognize their cognitive limitations, so it doesn’t always negatively impact their daily live or usual activities significantly (2).

What was the experimental procedure?

In this study, the research team developed a hybrid deep learning architecture that combines multi-scale convolutional neural networks (CNNs) with long short-term memory (LSTM) networks. This configuration was constructed because it allowed them to capture both the spatial and temporal patterns in the fNIRS data (1).

As part of the experimental procedure, the research team tested their method on a cohort of 64 participants, consisting of 37 individuals diagnosed with MCI and 27 healthy controls. Participants were evaluated while performing three cognitive tasks commonly used in neuropsychological testing: the N-back task (which measures working memory), the Stroop task (which evaluates cognitive flexibility and selective attention), and the verbal fluency test (VFT) (1). For these three cognitive tasks, the fNIRS data was collected across multiple channels. The variables the researchers measured were changes in oxyhemoglobin (HbO), deoxyhemoglobin (HbR), and total hemoglobin (HbT) levels in the brain (1).

What were the results of the study?

The researchers discovered through 10-fold cross-validation that the HbT signal was the most reliable indicator for classification. The model achieved its highest accuracy of 93.22% during the N-back task, followed by 91.14% for the Stroop task and 89.58% for the VFT (1). The researchers also noticed that HbR or HbT provided better classification results than the traditionally used HbO signal, suggesting that incorporating multiple chromophores improves its diagnostic ability (1).

What were the significant findings of the study?

The main finding was that using data from all fNIR spectroscopy channels improved classification performance. This challenges conventional assumptions that signal processing should focus on activated areas and highlights the benefits of analyzing whole-brain hemodynamic patterns (1).

The researchers also demonstrated that spectroscopic methods could help automate the clinical diagnostic process. The most commonly used methods require data segmentation and feature extraction. Here, the researchers demonstrated a way to reduce this labor-intensive work without compromising accuracy (1).

What are the implications of this study?

This study opens up avenues to explore applying this method to other mental disorders, including ADHD, bipolar disorder, and psychiatric evaluations. Furthermore, the technique is scalable to hybrid neuroimaging data sets, including combinations of fNIRS, electroencephalogram (EEG), and functional magnetic resonance imaging (fMRI) data (1).

Looking forward, the authors advocate for future fNIRS research to move beyond ROI-based analysis and instead focus on cleaning and dimensionality reduction techniques that maintain the integrity of the full-channel data (1). They also suggest developing robust, scalable prediction models that leverage deep learning’s capacity to uncover complex patterns in large, multidimensional data sets (1).

References

1. Kang, M.-K.; Hong, K.-S.; Yang, D.; Kim, H. K. Multi-scale Neural Networks Classification of Mild Cognitive Impairment Using Functional Near-infrared Spectroscopy. Bio. Biomed. Eng. 2025, 45 (1), 11–22. DOI: 10.1016/j.bbe.2024.12.001
2. Mayo Clinic, Mild Cognitive Impairment (MCI). Mayo Clinic. Available at: https://www.mayoclinic.org/diseases-conditions/mild-cognitive-impairment/symptoms-causes/syc-20354578 (accessed 2025-06-26).