Deep Learning Forecast of Cognitive Workload Using fNIRS Data
DOI: 10.1109/ichms59971.2024.10555701
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This study addresses the challenge of proactively managing pilot cognitive workload in helicopter operations, aiming to shift from reactive monitoring to predictive forecasting. While current methods assess workload in real-time, they lack the foresight necessary to mitigate impending high-load scenarios. The authors propose using deep learning to forecast future cognitive states based on functional Near-Infrared Spectroscopy (fNIRS) data, which captures hemodynamic changes in the prefrontal cortex associated with cognitive demand. The research hypothesizes that deep learning models can extract meaningful temporal features from fNIRS data to forecast workload and that hybrid CNN-LSTM architectures will outperform stacked LSTM models due to their ability to capture both spatial and temporal dynamics. The experimental design involved seven participants, including three experienced pilots, performing 46 trials in a UH-60V cockpit simulator under varying flight conditions designed to induce different workload levels. fNIRS data was collected and processed to derive oxygenated and deoxygenated hemoglobin levels, which served as input features alongside real-time workload classification confidence scores. The study evaluated three sequence-to-sequence deep learning architectures: a stacked LSTM, a CNN-LSTM hybrid, and a convolutional transformer. These models were trained to forecast workload confidence levels for two future time horizons: 10 seconds and 30 seconds. The input sequence length was fixed at 120 seconds, and models were trained using an 80-20 data split, with one trial reserved for hold-out validation. Results indicated that LSTM-based architectures were superior for 10-second forecasting tasks. The CNN-LSTM model achieved 94% accuracy and the stacked LSTM achieved 93% accuracy for 10-second forecasts, while the transformer model underperformed at 88%. For 30-second forecasts, the stacked LSTM slightly outperformed the others with 87% accuracy, followed by the transformer and CNN-LSTM at 85%. Contrary to the hypothesis, the CNN-LSTM did not surpass the stacked LSTM in predictive accuracy, though it offered significantly faster computation times, roughly half that of the transformer model. The study also found that 10-second forecasting variants consistently yielded stronger classification results than 30-second variants. The findings support the feasibility of using deep learning to forecast cognitive workload from fNIRS data, validating the potential for proactive workload management in complex human-machine systems. The superior performance of LSTM models suggests limitations in transformer self-attention mechanisms for this specific application, while the speed advantage of the CNN-LSTM model highlights its potential for real-time implementation. The authors conclude that forecasting provides critical foresight for strategic resource allocation and recommend future research explore diverse time-series methods and ordinal measures of cognitive workload to better capture shifting demands.
Key finding
LSTM-based deep learning architectures outperformed transformer and CNN-LSTM hybrid models in forecasting pilot cognitive workload from fNIRS data, particularly achieving higher accuracy for 10-second prediction horizons.
Methodology
simulator
Sample size: 7
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-27.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-27 |
| archive | success | canonical_url | — | — | 1 | 2026-06-06 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | skipped | — | — | — | 3 | 2026-06-04 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.