Abstract
Introduction: Impaired sleep quality, quantity and timing are associated with neurological and mental health disorders, potentially through disruption of functional and anatomical neuronal pathways. Sleep state oscillations likely encode brain health, therefore, sleep may provide accessible biomarkers for estimating brain health. One potential biomarker is sleep electroencephalogram (EEG)-based brain age, which if elevated above chronological age, could indicate abnormal or accelerated aging. Another potential method to build a biomarker is to directly estimate cognitive function from sleep data. Here, we investigate how well age and neuropsychological test scores can be predicted from EEG using an artificial deep neural network.
Materials and Methods: We used data from 735 participants of the Framingham Heart Study (FHS). Each participant had at a minimum one polysomnographic recording (PSG) and one neuropsychiatric evaluation for fluid intelligence (Wechsler Adult Intelligence Scale, NPS). Additionally, age at evaluation was available for each participant, resulting in 1244 PSG-age-NPS triplets. We solely used EEG data from C3 electrodes for this analysis. The EEG signal, recorded in 125 Hz, was bandpass-filtered between 0-20Hz and transformed into the time-frequency domain with the multitaper method (2 second window length, 1 second step size). The resulting spectrograms were all zero-padded to 11 hours, harmonizing the numerical dimensions of the sleep representations across subjects.
We developed an artificial deep neural network with a “U-Net”-like architecture, i.e., consisting of an encoder and decoder part. The network’s input was a subject’s EEG spectrogram, and its tasks were to predict sleep stages for 30-second epochs, the subject’s age, and the subject’s neuropsychological test score (NPS). The model mainly consisted of convolution-batch normalization-max pooling/upsampling layers, with a total of 11 million parameters. Fully connected layers for the age and NPS prediction tasks consisted of 15,600 parameters. The loss functions used were cross entropy for the sleep staging task and mean squared error for the age and NPS tasks. We evaluated the model’s performance with 10-fold cross-validation.
Results: Sleep staging: The accuracies (means and standard deviations across folds) per sleep stage were Wake 0.64 (0.02), N2 0.61 (0.02), N3 0.59 (0.04), NREM combined 0.79 (0.01), and REM 0.53 (0.06). Cohen’s Kappa was k=0.54 (0.02). Predicting age: Pearson correlation between chronological age and predicted age was r= 0.60 (0.06). Predicting neuropsychological test score: Pearson correlation between the NPS and the predicted cognitive score was r= 0.25 (0.07).
Conclusions: A multi-task deep learning network was able to predict sleep stages, age, and cognitive scores from sleep EEG with varying performance precision. While there was a strong correlation between predicted age and chronological age, predicted cognition scores correlated weakly to moderately with the actual scores. It is likely that prediction performance for all tasks can be increased with larger and more variable datasets. The sleep-based, subject-specific estimates for age and cognition may serve as biomarkers for brain health, a hypothesis we will test in future studies.
Acknowledgements: Research was supported by the National Institute of Neurological Disorders and Stroke of the National Institutes of Health, 1R01AG062989.
Ganglberger, W