A deep neural network model predicted the brain age of patients based on electroencephalogram (EEG) data recorded during overnight sleep studies.
The artificial intelligence (AI) model predicted brain age with a and a Pearson's r value of 0.933, surpassing the performance of prior research, reported Yoav Nygate, MS, of EnsoData in Madison, Wisconsin, at SLEEP 2021, a joint meeting of the American Academy of Sleep Medicine and the Sleep Research Society.
Brain age index -- chronological age subtracted from EEG-predicted brain age -- was associated with epilepsy and seizure disorders, stroke, elevated markers of sleep-disordered breathing (apnea-hypopnea index and arousal index), and low sleep efficiency (all P<0.05).
In addition, people with diabetes, depression, severe excessive daytime sleepiness, hypertension, or memory and concentration problems had an elevated brain age index on average compared with healthy people (all P<0.05).
"We show the power of artificial intelligence to exceed human capabilities and perform tasks that humans cannot," Nygate said. "While clinicians can only grossly estimate or quantify the age of a patient based on their EEG, this study shows an AI model can predict a patient's age with high precision."
"Since the AI model was trained to predict age -- an objective value that is not subject to label noise -- any divergence of the prediction from the target output is associated with either signal artifact in the input data or other underlying physiological conditions," he told 鶹ý.
The input to the model was a full night raw eight-channel EEG and electrooculogram (EOG) montage. The target output was the chronological age of patients.
The model was trained on 126,241 clinical sleep studies, validated on 6,638 studies, and tested on a holdout set of 1,172 studies. The holdout dataset included several categories of patient demographic and diagnoses to identify associations between brain age and various medical conditions. Analyses controlled for variables like sex and BMI.
"The first surprising result is the degree of accuracy to which the AI model was able to predict the age of a patient," Nygate observed. "A mean absolute error of 4.6 years was calculated across 1,172 patients, which is the lowest error rate we observed compared to previously published results in an exhaustive literature search."
"The second surprising finding was how many patient disorders, such as depression, diabetes, hypertension, severe excessive daytime sleepiness, and low sleep efficiency, were correlated with a shift in the predicted brain age from the chronological age of the patients," he said.
"Not only did we receive statistically significant shifts in the brain age distributions of diseased versus healthy populations, the direction of the shift was rather intuitive," he continued. "For example, we observed that diabetic patients have a higher mean predicted brain age compared to non-diabetic patients and patients with high sleep efficiency have a lower mean predicted brain age compared to patients with low sleep efficiency."
The study provides initial evidence of AI's potential to assess brain age, Nygate noted.
"Our hope is that with continued investigation, research, and clinical studies, a brain age index will one day become a diagnostic biomarker of brain health, much like high blood pressure is for risks of stroke and other cardiovascular disorders," he said.
Disclosures
The study was supported by EnsoData.
Primary Source
SLEEP
Nygate Y, et al "EEG-based deep neural network model for brain age prediction and its association with patient health conditions" SLEEP 2021; Abstract 543.