Personalized AI Brain Digital Twin for Alzheimer’s Disease Detection

Background: Digital twins are personalized, virtual representations of patients that can replicate patient-specific activity. These have been extensively used in clinical applications for drug discovery, simulating patient trajectories, and modeling various pathologies such as Alzheimer’s disease (AD). EEG, or electroencephalography, is a cost-effective and non-invasive neuroimaging tool that captures real-time brain electrical activity. Advancements in artificial intelligence (AI) have driven the development of digital brain twins by utilizing biomarkers as parameters derived from EEG, functional Magnetic Resonance Imaging (fMRI), and other data modalities. With more data modalities, the mechanisms and accuracy of patient-specific digital twins can be reliably improved with increased subject-level identifiability. Multimodal data have been explored in digital twins, as well as unimodal approaches (e.g. EEG-only). However, EEG, despite being comparatively more accessible than fMRI and other modalities, lacks the spatial resolution that fMRI possesses while being much more susceptible to noise. Naturally, this opens up avenues to explore trends of identifiability, data modalities, and cost-effective approaches to EEG-only digital twins of AD. Prior research on digital brain twins has focused on the reliability of classification, rather than accurate identifiability of subjects from their digital twin parameters, which has been recently established as essential. We intend to identify key trends in digital twins for memory disorders such as AD, comparing EEG-only approaches and alternative digital twin architectures. Our research aims to inform future approaches to digital twins of Alzheimer’s disease by proposing new approaches to digital twin data modalities and methods of validation.

Methods: We conducted an extensive literature review on digital brain twins for Alzheimer’s detection and other pathologies. We also explore other architectures, frameworks and models, including neural mass models, latent space state models, and agentic pipelines.

Results: We identify three types of digital brain twins. Mechanistic models are biophysically grounded and are highly interpretable, while data-driven models employing deep learning models capture temporal/spatial structure of EEG, being comparatively cheaper but can struggle with mechanistic interpretation. Hybrid approaches utilize multiple imaging modalities or pairing biophysically plausible models with a learned data-driven component, but have the most computational burden. One modality (like fMRI), may be used to define the model's brain network structure, while the second modality (like EEG) fine-tunes fast, cortical activity. The result is interpretable models which have digital biomarkers anchored by more than one source of evidence, shown in cases to outperform standard EEG/MEG metrics. EEG is comparatively inexpensive, portable, and accessible with high temporal resolution, but with high susceptibility to noise and low spatial resolution. fMRI offers substantially higher spatial resolution and can capture hemodynamic responses reflecting subcortical activity that EEG cannot resolve, but it is expensive, immobile, and far less accessible for large-scale or repeated-measures studies. EEG+fMRI digital twins have been explored in recent literature, yet functional Near-Infrared Spectroscopy (fNIRS) is another alternative hemodynamic technique demonstrated to trade spatial resolution and cost for similar temporal resolution and increased portability. Furthermore, recent work in digital twins for AD tends to reflect group-level statistical comparisons by using digital biomarkers to test discrimination accuracy, rather than testing the reliability and subject-specific identifiability of those parameters themselves.

Conclusion: Our findings demonstrate that future approaches for AD digital twins, specifically neural mass models, lie in addressing identifiability. In addition, investigating relatively unexplored combinations of data modalities such as EEG+fNIRS for cost-effective, accessible clinical alternatives and utilizing latent-state space models for denoising EEG into lower-dimensional representations may serve to constrain the digital twin and increase identifiability among subjects. Future phases include employing identifiability analyses to existing digital brain twin AD models, and measuring whether additional constraints (EEG+fNIRS, latent-state) indeed increase identifiability.