Neural Noise

Deep learning for mild cognitive impairment detection via fNIRS

2 min read

  • A deep learning model improved detection of mild cognitive impairment (MCI) using fNIRS data.1
  • fNIRS–machine learning is viable for cognitive screening and aligns with portable neurotech and neural data analysis.1 1

Weekly enrichment (2026-07-20)

  • The headline result refers to a multi-scale CNN combined with LSTM layers applied to the full fNIRS time series (Pusan National University, Qingdao University and Washington University School of Medicine in St. Louis; Biocybernetics and Biomedical Engineering), which skips manual feature extraction and signal segmentation.2 3
  • The study enrolled 64 participants (37 MCI patients and 27 healthy controls) performing three tasks—N-back, Stroop and verbal fluency—evaluated with 10-fold cross-validation.2 3
  • Highest accuracies were 93.22% (N-back), 91.14% (Stroop) and 89.58% (verbal fluency), all obtained using total hemoglobin (HbT).2 3
  • Deoxy- (HbR) or total hemoglobin (HbT) outperformed the conventionally used oxyhemoglobin (HbO), and using all channels beat region-of-interest-only analysis, challenging the assumption that only activated areas matter.2 3
  • For context, an earlier four-layer CNN on fNIRS temporal-feature maps (15 MCI, 9 controls) reached average accuracies of 89.5% (N-back), 87.8% (Stroop) and 90.4% (VFT), peaking at 98.61% for the HbO slope map over the 20-60 s window of the N-back task.4
  • A CNN on HbO2 slope biomarkers in a larger cohort (82 MCI, 148 controls) reached 96.09% (left prefrontal cortex, 20-60 s window), outperforming the Korean version of the MoCA screening.5
  • A spatio-temporal framework integrating resting-state and task-state fNIRS across 104 participants achieved 90.91% accuracy with ~91% feature-dimensionality reduction, addressing small-sample overfitting.6
  • Clinically, MCI is a precursor to Alzheimer’s disease, and portable, motion-tolerant fNIRS with deep learning is positioned as a low-cost, online-capable screening tool—though sample sizes remain small (tens of participants).3 6

Footnotes

  1. https://news.google.com/rss/articles/CBMivwFBVV95cUxOSndNRFdYak1CaDhWTjd3Z0pYM0FmWmI3cmdycmVXTTZRYTNLSWk3dWpabnQ1RGhpTFlMSzJJTzJFeFhTSDdObFd3RWd6eS1sVEpUVnZvTTZEMHRFa3dCRUlxR2xYaWJNOGFkaGtWSC04R2s1anFOcXdwN1RkLTlIRk8xeERLZFVDUlBfamoyTUNlZ3I0Uk10bHV3ejIyNFFxaXdEc3c0UG9OTUlwN0EwRUhDWURjY3E0M1NtTHF6Yw?oc=5 ↩ ↩2 ↩3

  2. https://www.spectroscopyonline.com/view/deep-learning-model-improves-detection-of-mild-cognitive-impairment-via-fnirs-data ↩ ↩2 ↩3 ↩4

  3. https://www.sciencedirect.com/science/article/abs/pii/S0208521624000895 ↩ ↩2 ↩3 ↩4 ↩5

  4. https://www.frontiersin.org/articles/10.3389/fnagi.2020.00141/full ↩

  5. https://www.psychiatryinvestigation.org/journal/view.php?number=1720 ↩

  6. https://pmc.ncbi.nlm.nih.gov/articles/PMC12947967/ ↩ ↩2

Graph View