fNIRS Skill Assessment Foundation Model

Deciding whether a trainee surgeon or medic is competent is still largely a matter of an expert watching and forming a judgment. That is slow, subjective, and offers no window into the cognitive and motor processes that separate a proficient operator from a struggling one. Functional near-infrared spectroscopy — which measures hemodynamic activity in the prefrontal cortex with a wearable optode array — has been shown to carry that signal, but the models built on it have been narrowly task-specific, dependent on elaborate preprocessing pipelines, and unable to survive a change of procedure or optode montage without full retraining.

This model, developed by researchers at Rensselaer Polytechnic Institute and the FAMU-FSU College of Engineering with clinical collaborators at the University at Buffalo and the U.S. Army DEVCOM Soldier Center, applies the foundation-model recipe to that problem. A transformer encoder is pretrained by self-supervised masked segment reconstruction on minimally processed fNIRS time series from laparoscopic suturing and endotracheal intubation, then frozen. New procedures are reached not by retraining but by inserting a small adapter, and the model was validated on cricothyrotomy — an emergency airway procedure with a different optode montage and a completely different task structure.

Where EEG foundation models such as LaBraM are pretrained on thousands of hours of public recordings, fNIRS has no comparable public corpus; the contribution here is showing that a compact model trained on a few carefully collected human-subject studies can still transfer across procedures. The work was published in ACM Transactions on Computing for Healthcare.

#Key Features

• Raw-signal input: Preprocessing is limited to optical-density conversion and a 0.01–0.5 Hz band-pass filter, avoiding the multi-stage pipelines that discard subtle features and block real-time deployment.
• Masked segment pretraining: A BERT-style objective adapted to time series masks contiguous segments of 3 to 11 samples across the full channel set and reconstructs them under mean-squared-error loss, requiring no proficiency labels.
• Frozen-backbone few-shot transfer: Adaptation to a new procedure trains only a two-layer dense adapter of under 3,000 parameters, with encoder and decoder weights and adapter biases all held fixed.
• Interpretable channel attention: A channel-axis attention module learns which cortical channels carry discriminative information; ablation confirms that removing the top-attended channels collapses performance while removing the rest barely affects it.
• Temporal attention aligned to subtasks: Attention over time concentrates on hierarchically defined subtask boundaries and tracks stress-induced changes in neural variability.

#Technical Details

The architecture is a compact two-layer transformer encoder with an 80-dimensional embedding and five attention heads, paired with a decoder that queries the encoder through cross-attention using one learned token per label category — here, a single "task proficiency" token producing a sigmoid-activated pass/fail logit. The channel attention module uses 16 heads with 0.5 dropout and a residual connection back to the input. Variable-length trials are zero-padded and masked out through an attention padding mask. Nine encoders were pretrained from different initializations and results averaged, since temporal attention maps proved sensitive to initialization even though output logits were not.

Data were collected with a NIRSport2 continuous-wave spectrometer at 5.09 Hz using 16 sources and 15 detectors at 3 cm separation plus 8 short-separation detectors, at 760 nm and 850 nm. Classification accuracy exceeds 88% on all pretraining tasks, with Matthews correlation coefficient above 0.91 on endotracheal intubation. On the held-out cricothyrotomy task the adapter reaches an AUC above 85% from as few as 24 labeled samples, and leave-one-subject-out cross-validation over the full out-of-distribution set gives a mean accuracy of 87.7%.

#Applications

The target use is objective, automated credentialing and training feedback for high-stakes procedural skills — laparoscopic surgery, airway management, and military combat casualty care. Because the backbone stays frozen, a training program can extend coverage to a new procedure by collecting a few dozen labeled trials rather than mounting a new study. The interpretability layer is not incidental: identifying which prefrontal sub-networks and which subtasks discriminate proficiency points toward targeted remediation rather than a pass/fail verdict alone.

#Impact

This is an early demonstration that the pretrain-freeze-adapt pattern works on fNIRS, a modality where dataset scale is a fraction of what EEG offers. The evidence base remains modest: cohorts are small, all data come from a single research consortium, and the model addresses binary proficiency rather than graded skill. Reproduction is currently not possible from public artifacts — the code repository is private and available only on request, and the human-subject datasets are stated as pending release after the underlying studies conclude.