Thomas J Vanasse, PhD1 • Sam Rusk, BS1 • Yoav N. Nygate, MS1 • Chris R. Fernandez, MS1 • Jiaxiao M. Shi, PhD2 • Jessica Arguelles, BS2 • Matthew T. Klimper, BS2 • Emerson M. Wickwire, PhD3 • Nathaniel F. Watson, MD, MS4 • Dennis Hwang, MD, PhD2
Introduction
Machine Learning (ML) can draw upon complex patient data to predict current and future health status, but the best performing ML models are often difficult to interpret. One such ML-defined health marker is the brain age index (BAI), the difference between ML-predicted age and chronological age (CA) using electroencephalogram (EEG) during sleep. BAI is associated with multiple disease states including increased risk for all-cause mortality. Still, validation of published models has been limited, and the underlying associations between polysomnography sleep architecture metrics and BAI estimation are not well understood.
Methods
A deep neural network (DNN) model was trained to predict the brain age (BA) of patients using raw EEG signals recorded during clinical polysomnography (PSG). The DNN was trained on N=54,000 PSGs. The model was then validated on two independent datasets: a Kaiser Permanente (KP) Historical PSG Dataset (N=10,694), and a multi-site retrospective PSG dataset (N=15,158). To test association with BAI, ordinary least squares (OLS) regression was performed on these validation datasets using basic features derived from sleep staging architecture and arousal data.
Results
In the validation datasets, we observed mean absolute error (MAE) of 6.67 (6.69/6.49 F/M, KP dataset) and 5.65 (5.56/5.44 F/M, independent dataset). In the OLS analysis, consistent, negative associations were detected between both N3 and REM sleep percentage vs. BAI. Furthermore, BAI was consistently positively associated with N1 sleep percentage. Each of these effects was significant at p< 0.001.
Conclusions
This study builds upon and expands prior research by evaluating large multi-site datasets and assessing the relationship of N3/REM sleep duration with the predicted brain age. Using a DNN model, BAI was associated with significantly increased N1 and concomitant decreased N3/REM. Further research is needed to determine if BA is malleable and potentially reduced with clinical intervention and positive changes in lifestyle and wellness-related behavior.
