Extended Data Fig. 6: Ablation studies for assessing the benefit of multi-task learning and transfer learning.

a, PD diagnosis performance on breathing belt data (n = 6,660 nights from 5,652 subjects) and wireless data (n = 2,601 nights from 53 subjects), for the model with all of its components and the model without the qEEG auxiliary task (that is, without multi-task learning) and without transfer learning. Each bar and its error bar indicate the mean and standard deviation across 5 independent runs. The graphs indicate that the qEEG auxiliary task is essential for good performance and eliminating it reduces the AUC by almost 40%. Transfer learning also boosts performance for both breathing belt data (7.8% improvements) and wireless data (8.3% improvements), yet is not as essential as multi-task learning. b, Pearson correlation of the PD severity prediction and MDS-UPDRS. The correlation is computed for subjects in the wireless datasets (n = 53 subjects) since their MDS-UPDRS scores are available. Each bar and its error bar indicate the mean and standard deviation across 5 independent runs. The results indicate that transfer learning is useful, but multi-task learning (that is, the qEEG auxiliary task) is essential for good performance.