Neural Noise

fNIRS feature extraction for workload estimation

2 min read

  • Functional near-infrared spectroscopy (fNIRS) feature extraction has been applied to workload estimation. 1
  • The approach is transferable to BCI, neurofeedback, and physiological time-series applications. 1
  • Work is from Microsoft with a credible implementation path; tier-1. 1 1

Weekly enrichment (2026-07-20)

  • The primary work is “Functional Near-Infrared Spectroscopy Feature Extraction with Application in Workload Estimation” by Elisabeth R. M. Heremans, David Johnston, Dimitra Emmanouilidou, Andre Golard, Ivan Tashev, and Ryen White (Microsoft Research), presented at IEEE ICASSP in April 2025.2 3
  • fNIRS is a non-invasive optical technique that estimates cortical neuronal activity by measuring blood oxygenation (changes in oxygenated hemoglobin HbO and deoxygenated hemoglobin HbR), and is more robust to movement artifacts than EEG, which matters for real-time monitoring in dynamic settings such as aviation.3
  • The Microsoft study developed an extensive fNIRS feature set and combined it with respiration and heartbeat (ECG) signals, building a subject- and session-independent workload estimator validated in a virtual flight simulator where workload was objectively defined by task performance.2 4
  • Their best fNIRS-only regression model reached a correlation of 0.3188 with objective workload labels, improving to 0.3268 when breathing signals were incorporated; these are modest correlations, reflecting the difficulty of subject-independent estimation from noisy hemodynamic signals.2
  • The experimental protocol had participants perform flight maneuvers of varying complexity while fNIRS, breathing, and ECG were recorded, with workload quantified by both objective task-performance metrics and subjective self-report.4
  • A common, simple, and effective feature in this literature is the slope of a line fit to each channel’s HbO and HbR signal over a window (e.g., 16 features from 8 channels); other frameworks add features such as hemodynamic slope plus a “deep contribution ratio,” which raised three-level workload classification accuracy to ~80.6% versus ~59.8% for a single conventional feature.5 6
  • Deep-learning approaches on fNIRS workload data commonly use CNNs and hybrid CNN-LSTM/CNN-GRU architectures to capture spatial and temporal structure; reported subject-specific accuracies include roughly 83-96% for CNNs and about 92.35% for a CNN-LSTM binary low-vs-high workload classifier, illustrating the gap between subject-specific benchmarks and the harder subject-independent regression that Microsoft targeted.7 8

Footnotes

  1. https://news.google.com/rss/articles/CBMi5AFBVV95cUxPX1FvaTFnTjVULUcwVjViSHlMZ0RrcDYzQlA2dzNCX0FDZ0VEWU9WNDl0ZTUxOUptQ3JLNjE2OG11OGhReEhtZjFiaGF2cGU1RnBLNjhuMkpLWk5OMENBRTRabXJ6NF9IcWg0RDBkUktzSnBhZFVBb2k0bjBsVGJqNXphVXZTclRPcXNRNHVGM0ZENGl6VnlRR0sycGtDSTloWUtCR3ZtbUdyUE1sU29fNXpndXRXbWV2WnhTRzFPOHo1V3JhalhGUE00U2lCM3U0VmFwZXNJWnItUzZJek9mbF9jRlI?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.microsoft.com/en-us/research/publication/functional-near-infrared-spectroscopy-feature-extraction-with-application-in-workload-estimation/ ↩ ↩2 ↩3

  3. https://ieeexplore.ieee.org/document/10888530 ↩ ↩2

  4. https://www.microsoft.com/en-us/research/video/shining-light-on-the-learning-brain-estimating-mental-workload-in-a-simulated-flight-task-using-optical-f-nirs-signals/ ↩ ↩2

  5. https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2013.00935/full ↩

  6. https://doi.org/10.1109/tnsre.2020.3026991 ↩

  7. https://doi.org/10.1186/s12883-023-03504-z ↩

  8. https://doi.org/10.1109/ichora65333.2025.11017192 ↩

Graph View