Digital twins support cross-modal and cross-centric classification of mild cognitive impairment

Background Neural recordings capture crucial pathophysiological processes along the dementia continuum. However, cross-center variability in recording techniques and paradigms limit their generalizability and diagnostic power, preventing clinical use. We here propose a computational approach enabling cross-center classification even in the presence of completely different clinical pipelines. Methods We leveraged a digital twin model to derive digital biomarkers linking neurodegeneration mechanisms to alterations in neural activity across multiple recording modalities. We tested the generalizability of digital biomarkers through cross-center classification of Mild Cognitive Impairment (MCI) and healthy subjects in two independent clinics. The two datasets presented different recording techniques (EEG and MEG), preprocessing modalities, recruitment criteria and diagnostic guidelines. Digital biomarkers derived from one clinic were tested for classifying patients in the other clinic and vice versa employing a transfer learning approach. Results Digital biomarkers outperform standard biomarkers in the MCI vs healthy classification in both separate datasets (83% vs 58% for EEG dataset and 75% vs 68% for MEG dataset). Moreover, they achieve accurate and consistent cross-center classification (77–78% accuracy), while standard biomarkers perform poorly in the generalization attempt (56–65%). Additionally, digital biomarkers reliably predict global cognitive status across clinics across both datasets ( p < 0.01), while standard biomarkers present no correlation. Conclusions Digital biomarkers generalize across recording techniques and datasets, enabling a cross-modal and cross-center classification of a patient’s condition. These biomarkers offer a robust measure of patient-specific neurodegeneration, mapping neural recordings anomalies into a common framework of underlying structural alterations. The vast differences between the two datasets support the applicability of this approach also in the presence of high inter-center variability.