XGBoost versus Convolutional Neural Network Transfer Learning for Predicting Future Cognitive Impairment from Polysomnography

Topi Niemi1, Esko Toivonen1, Mateusz Solinski2, Venla Koivunen1, Marjaana Nurmo1, Noora Poutanen1, Teemu Pukkila1, Esa Rasanen1
1Tampere University, 2Computational Physics Laboratory, Tampere University


Abstract

As part of the George B. Moody PhysioNet Challenge 2026, team Better Call Sandman developed two approaches for predicting future cognitive impairment from polysomnography. The first combined engineered physiological and sleep-related features with XGBoost, while the second used a convolutional neural network (CNN) with foundation-model pretraining and transfer learning. XGBoost achieved a mean cross-validation age-conditioned area under the receiver operating characteristic curve (AUROC) of 0.870 on the public training data. On the official hidden validation set, XGBoost achieved an age-conditioned AUROC of 0.636, compared with 0.574 for the CNN, ranking 191st of 514 submissions. Based on this performance, XGBoost was selected as our final model for evaluation on the hidden test set. Although the feature-based approach performed better in the Challenge, our CNN experiments indicate that transfer learning remains a promising direction, particularly with richer multimodal inputs and improved generalization.