Since polysomnography recordings (PSGs) have demonstrated their richness allowing Deep Learning models to capture complex hidden patterns, some applications might require further investigation in order to achieve better performance.
This is the goal of the George B. Moody PhysioNet Challenge 2026~\cite{reyna2026challenge}, which encourages teams to explore the prediction of future cognitive impairment.
To address the challenge, a lightweight head on top of a frozen SleepFM backbone is proposed as an architecture that tries to make the most of this foundation model in the task required.
A study which covers different input data and pooling over time, among others, is developed taking into account the cross-domain difficulty of the problem. In the hidden validation set, our team, UPV Maths, had a 0.681 AUC age conditioned as its best result and that entry is ranked 115th out of 514.