Aims: This study aims to evaluate deep learning architectures applied to polysomnography recordings (PSG) for predicting the future cognitive im-pairment (CI) to evaluate whether it is possible to use automated tools for early detection based on sleep biomarkers. Methods: A balanced dataset of 611 patients (after excluding 5 subjects due to missing .edf files or incomplete channels) was used. Only algorithmic annotations of sleep stages and clinical events were used as input data. While a 1D Convolutional Neural Network (CNN) was the first approach, a hybrid CNN and Long Short-Term Memory (LSTM) architecture is being designed to effectively capture both local features and temporal dynamics. In order to train, optimize and evaluate models a 5-fold balanced cross-validation framework was implemented on 85% of the data, with an isolated 15% test set. Then, after tuning the hyperparameters, a 5-fold unbalanced cross-validation (prevalences between 5-15%) was used to verify the best models' robustness against real scenarios. Hyperparameter optimization was con-ducted via Bayesian inference (Optuna), using a 3-epoch average validation AUC to ensure the selection of stable models. Results: The CNN architecture achieved a modest generalization with a test AUC of ~0.60, however, performance decreased to 0.527 on the unoffi-cial test set (UPV Maths team). While first attempts on CNN+LSTM have being made, the next natural step will be to add raw signals and the human annotations as inputs, as well as to implement a transformer architecture to study the performance of the self-attention mechanism, as feature extrac-tion techniques to enrich the input data. Conclusion: Deep learning applied to PSG has been tested as a predictor of the future onset of CI. Preliminary results show a modest degree of suc-cess, highlighting that while CNN+LSTM architectures are highly capable, more techniques and other architectures might be essential to achieve better performance.