A short window
Each estimate uses a short window of recent fMRI activity and never uses future data.
Decoding sleep from brain activity.
People often drift toward sleep during resting fMRI scans, and that shift changes the brain signals researchers use to study disease. I am developing methods to estimate sleep stage (Wake, N1, N2, or N3) from fMRI alone, so brain state can be measured and accounted for.
Drowsiness affects fMRI studies.
Sleep stages are normally scored from EEG, which most fMRI studies do not record. An fMRI-only estimate would let researchers ask how brain activity changes with drowsiness, even in scans that have already been collected.
01
Interpret brain scans
Arousal (level of wakefulness) changes brain connectivity. Without a measure of it, drowsiness during a scan can be mistaken for an effect of disease, drugs, or an intervention.
02
Learn from few labels
Few datasets record sleep stages during fMRI, so models must learn from small, imbalanced labeled sets. That makes this a natural test of when pretrained models help.
03
Build trustworthy tools
Models are evaluated on participants held out from training, and simple baselines are reported alongside more complex models.
Overall goal. A reliable, interpretable measure of brain state for sleep neuroscience and for imaging studies where drowsiness affects interpretation. This is research-stage work.
From brain activity to sleep stage.
Functional connectivity
Measure how strongly activity in each pair of brain regions rises and falls together.
Tangent-space features
Map each connectivity matrix into a space where standard classifiers work well.
Stage probabilities
Estimate how likely each sleep stage is, tested on participants the model never saw.
N1, N2, and N3 are the light-to-deep stages of non-REM sleep. I also compare this approach with neural networks trained from scratch and with pretrained fMRI foundation models.
Learning from few labels.
Sleep labels require simultaneous EEG and expert scoring, so labeled data are scarce. I am testing when pretrained fMRI models reduce the number of labeled participants needed, and which brain networks drive their predictions.
Ongoing, unpublished work.
This page is a high-level overview. Detailed methods and results will be shared with a future paper.