Deep Learning Model Predicts Cognitive Status With High Accuracy From Sleep EEG Data


Researchers presented an innovative multi-task learning paradigm that leveraged deep learning, night sleep EEG data, sleep stage labels, and covariates to simultaneously predict cognitive performance in an older patient cohort.

Joyita Dutta, PhD  (Credit: X)

Joyita Dutta, PhD

(Credit: X)

In a new study, a deep learning, multi-task framework using whole-night sleep electroencephalography (EEG) outperformed a comparison model in predicting cognitive status. According to study investigators, this was the first successful deep learning model for jointly predicting cognitive status based on 2 different testing instruments from a patient’s sleep EEG data.

The presented model leverages deep learning to simultaneously predict cognitive performance measured from two distinct cognitive tests, Cognitive Abilities Screening Instrument (CASI) and Digit Symbol Coding Test (DSCT), using whole-night sleep electroencephalography (EEG) data, sleep stage labels, and covariates.

Presented at the 2024 SLEEP Annual Meeting, held June 1-5, in Houston, Texas, the transformer model achieved an overall accuracy of 70.21% and 70.85%, respectively, for CASI and DSCT. In addition, it achieved an F1-score of 71.43% and 69.35%, respectively, for the 2 cognitive tests. In contrast, a neural network trained using 132 handcrafted sleep features as inputs had an overall accuracy of 59.57% with an F1-score of 60.08% for CASI and an overall accuracy of 60.64% with an F1-score of 60.04% for DSCT. The overall accuracy and F1-score margins were 10.64% and 11.35% for CASI and 10.21% and 9.31% for DSCT, respectively.

"We show that our feature learning approach, which relies on a transformer model, outperforms traditional feature handcrafting by a sizable margin in terms of an array of performance evaluation metrics," senior author Joyita Dutta, PhD, associate professor in the department of biomedical engineering at the University of Massachusetts Amherst, and colleagues, wrote.1 "Sleep disruptions have been found to have a strong association not only with normal cognitive decline that occurs with age, but also with dementia caused by neurodegenerative diseases such as Alzheimer disease. Predictive techniques that can automatically detect cognitive impairment from an individual’s sleep patterns have broad clinical and biological significance."

READ MORE: Age-Dependent Risk for Neurodegenerative Disorder Identified in Obstructive Sleep Apnea

Top Clinical Takeaways

  • A deep multi-task learning framework outperformed a method trained on handcrafted features in predicting cognitive status using sleep EEG data.
  • The transformer model achieved significantly higher accuracy and F1-scores for both Cognitive Abilities Screening Instrument and Digit Symbol Coding Test.
  • This study presented the first deep learning model that can jointly predict cognitive status using 2 different testing instruments from a patient's sleep EEG data.

In the study, investigators utilized data obtained from the MultiEthnic Study of Atherosclerosis participants who were randomly assigned to 2 cohorts, validation (n = 470) and training (n = 1110). Authors used a multi-task transformer architecture in their approach to learn complex patterns of sleep architecture for binary cognitive status prediction. The ground truth binary cognitive status labels (low and high score groups) were generated using the dataset's median CASI and DSCT values as thresholds. Researchers used sleep EEG, sleep stage labels, and covariates as inputs and binary CASI and DSCT cognitive score classes as targets to train and validate their model.

"Although the MESA dataset used in this study has good representation for mildly impaired individuals, very low CASI scores are under-represented. This leads to some data imbalance that may affect the underlying model which is set up for a regression task," Dutta et al noted.1 "Owing to this limitation, we report only binary classification results for this model. The threshold of 90 is consistent with a clinical MCI diagnosis based on CASI. Since there is good overall representation of each binary class for this threshold, the model’s performance for this classification task is robust."

The study was limited by the fact that results were based on only 1 night of data from each patient. Since patients tend to have a high degree of night-to-night variability in sleep patterns, researchers noted that a single night’s data may not capture their sleep habits. Given the cost and complexity of these types of studies, they are rarely performed for more than 1 or 2 nights for a single patient. Investigators noted that improved sleep monitoring accuracy for easy-to-use wearable EEG devices2 could enable multi-night data acquisition and capture a longer-term as well as a more complete picture of a patient's sleep habits. Longer-term assessments could potentially be more meaningful for sleep-cognition mapping efforts.

"In our current setup, we use attention from the second transformer encoder layer at the start point which combines information from all the epochs from the first encoder layer. As future work, we plan to utilize the first encoder layer to see which epochs are important and receive more attention," Dutta et al noted.1 "This will allow us to go further and examine which features each of those epochs carry and thus conduct an epoch-by-epoch interpretation."

Click here for more coverage of SLEEP 2024.

1. Song TA, Malekzadeh M, Saxena R, Purcell SM, Dutta J. A Transformer Model for Predicting Cognitive Impairment From Sleep. Presented at: 2024 SLEEP Annual Meeting; June 1-5; Houston, Texas. Abstract 0059.
2. Camargos EF, Louzada FM, Nóbrega OT. Wrist actigraphy for measuring sleep in intervention studies with Alzheimer's disease patients: application, usefulness, and challenges. Sleep Med Rev. 2013;17(6):475-488. doi:10.1016/j.smrv.2013.01.006
Recent Videos
Martin Tolar, MD, PhD
Patricia K. Coyle, MD
Aliza Ben-Zacharia, PhD, DNP, ANP-BC, FAAN
4 KOLs are featured in this series.
4 KOLs are featured in this series.
4 KOLs are featured in this series.
4 KOLs are featured in this series.
5 KOLs are featured in this series.
5 KOLs are featured in this series.
Martin Tolar, MD, PhD
© 2024 MJH Life Sciences

All rights reserved.