Artificial intelligence in presymptomatic neurological diseases: Bridging normal variation and prodromal signatures

Presymptomatic neurological diseases are marked by early pathological changes that occur before overt clinical symptoms. These stages, which include prodromes such as REM sleep behavior disorder in Parkinson's or mild cognitive impairment in Alzheimer's, offer critical opportunities for early intervention. However, their detection remains challenging due to the subtlety of changes and the overlap with normal interindividual variability. Artificial intelligence (AI), especially machine learning (ML) and deep learning (DL), offers new tools to uncover hidden signatures in complex biomedical data. First, we explore how supervised ML models can detect known prodromal patterns across diverse modalities, including EEG, cognitive scores, and structural imaging. Depending on the input, various model types - such as tree-based algorithms for structured data and convolutional or transformer networks for images and signals - can extract predictive features of early neurodegeneration. These approaches have demonstrated success in identifying at-risk individuals before clinical thresholds are reached. Yet, detecting only known patterns limits the scope of early intervention. Many individuals who will go on to develop neurological disease may not yet exhibit any recognized prodromal syndrome. Bridging this gap requires moving beyond predefined labels toward models capable of identifying subtle, unknown anomalies in individuals still considered healthy. Second, we address the detection of latent anomalies among individuals not yet considered at risk without identifiable known prodromal patterns. By mining clinical records, free-text medical notes, and population-level health databases (e.g., UK Biobank, EDS-AP-HP), and by analyzing sensor data from smartphones or wearables, AI can flag deviations from healthy patterns long before symptom onset or formal diagnosis. This approach holds promise for scalable, low-burden, ecological screening. Finally, we introduce the concept of pseudo-healthy twins - synthetic, personalized baselines generated from structural data such as MRI, to improve anomaly detection. These models predict a patient's expected healthy signal in another modality, such as PET, enabling the subtraction of normal anatomical and physiological variability to isolate disease-specific effects. Generative models like GANs and VAEs have shown promise in producing these cross-modal references, enhancing early anomaly detection in diseases like Alzheimer's and multiple sclerosis. Together, these approaches show how AI can bridge the gap between normal variation and early pathology, enabling more sensitive, personalized, and population-scalable detection of presymptomatic neurological disease.