Komeyer, V. | Search bioRxiv

bioRxiv2024

Can we predict sleep health based on brain features? A large-scale machine learning study

BackgroundsSeveral correlational or group comparison evidence highlighted robust associations between sleep health and macro-scale brain organization. However, inter-individual variability is critical in such interplay. Therefore, in this study, we aimed to investigate the role of brain imaging features in predicting diverse sleep health-related characteristics at the individual subject level using the Machine Learning (ML) approach. MethodsA sample of 28,088 participants from the UK Biobank was employed to calculate 4677 structural and functional neuroimaging markers. Then, we employed them to predict self-reported insomnia symptoms, sleep duration, easiness of getting up in the morning, chronotype, daily nap, daytime sleepiness, and snoring. To assess the predictability of brain features, we built seven different linear and nonlinear ML models for each sleep health-related characteristic. ResultsWe performed extensive ML analyses that involved more than 19 years of compute time. We observed relatively low performance in predicting all sleep health-related characteristics from brain images (e.g., balanced accuracy ranging between 0.50-0.59). Across all models, the best performance achieved was 0.59, using a linear ML model to predict the ease of getting up in the morning. In fact, a similar performance was achieved with models trained solely on age and sex, indicating that these demographic factors might be the ones driving the predictions. ConclusionsThe low capability of multimodal neuroimaging markers in predicting sleep health-related characteristics, even under extensive ML optimization in a large population sample, suggests a complex relationship between sleep health and brain organization.

neuroscience↗

bioRxiv2023

Sex classification from functional brain connectivity: Generalization to multiple datasets

Machine learning (ML) approaches are increasingly being applied to neuroimaging data. Studies in neuroscience typically have to rely on a limited set of training data which may impair the generalizability of ML models. However, it is still unclear which kind of training sample is best suited to optimize generalization performance. In the present study, we systematically investigated the generalization performance of sex classification models trained on the parcelwise connectivity profile of either single samples or a compound sample containing data from four different datasets. Generalization performance was quantified in terms of mean across-sample classification accuracy and spatial consistency of accurately classifying parcels. Our results indicate that generalization performance of pwCs trained on single dataset samples is dependent on the specific test samples. Certain datasets seem to "match" in the sense that classifiers trained on a sample from one dataset achieved a high accuracy when tested on the respected other one and vice versa. The pwC trained on the compound sample demonstrated overall highest generalization performance for all test samples, including one derived from a dataset not included in building the training samples. Thus, our results indicate that a big and heterogenous training sample comprising data of multiple datasets is best suited to achieve generalizable results.

neuroscience↗