Using Artificial Intelligence to Detect Early Signs of Alzheimer's Disease in People With Memory Concerns

Overview

AHEAD is a prospective, longitudinal, risk-stratified single-arm interventional study enrolling 300 patients with Subjective Cognitive Decline (SCD) at IRCCS San Raffaele Hospital, Milan, Italy. The study uses artificial intelligence (AI) to integrate multimodal data - including MRI, EEG, Optical Coherence Tomography (OCT), neuropsychological assessments, and plasma biomarkers - to identify individuals with underlying Alzheimer's disease (AD) biology and predict cognitive progression. Only participants found to be AD plasma biomarker positive (SCD+) undergo longitudinal follow-up at 12 and 24 months. Participants classified as high risk additionally receive a 6-month personalized multidisciplinary intervention combining high-frequency transcranial magnetic stimulation (TMS), digital cognitive training, structured physical exercise, and targeted management of modifiable vascular and behavioral risk factors.

Subjective Cognitive Decline (SCD) refers to the self-perception of worsening cognitive abilities despite normal performance on standardized neuropsychological testing. It affects approximately 10% of the general population and 20-35% of patients attending memory clinics. Although the majority of individuals with SCD do not progress to clinical forms of Alzheimer's disease (AD), they show a higher prevalence of AD-related pathological biomarkers compared with individuals without subjective cognitive complaints, with rates of cognitive decline estimated at approximately 20% per 1,000 person-years in memory clinic patients. Plasma biomarkers for AD represent minimally invasive and easily accessible diagnostic tools; however, their large-scale implementation in the broad SCD population is neither economically nor ethically sustainable because of costs, the risk of overdiagnosis, and the associated psychological burden. Artificial intelligence (AI) may represent a transformative tool for addressing the complexity of SCD management. By integrating multimodal data including cognitive assessments, MRI, EEG, and OCT, AI may help identify those individuals with SCD most likely to benefit from further diagnostic investigations, including plasma biomarker assessment. At baseline (T0), all participants undergo a minimum assessment dataset including clinical evaluation, standard neuropsychological assessment, structural MRI, and blood sampling. A subset additionally undergoes a comprehensive risk assessment, extended neuropsychological evaluation including digital cognitive testing and the Preclinical Alzheimer Cognitive Composite (PACC), resting-state EEG, and retinal imaging through Optical Coherence Tomography (OCT). Only patients found to be AD plasma biomarker positive (SCD+) undergo longitudinal follow-up visits at 12 months (M12) and 24 months (M24), including clinical evaluation, neuropsychological assessments, and blood sampling to monitor cognitive and biological progression. Participants stratified as high risk - defined as plasma p-tau217 greater than 0.1325 pg/mL, and/or APOE epsilon4 carrier, and/or elevated CAIDE Dementia Risk Score - enter a 6-month single-arm multidisciplinary intervention comprising: (1) targeted management of modifiable vascular and behavioral risk factors with monthly remote follow-up; (2) high-frequency TMS during the first 4 weeks (2-3 sessions per week); (3) home-based digital cognitive training, 2 sessions per week of 30 minutes each over 5 months; (4) structured physical exercise (walking, cycling, resistance training), 2 sessions per week of 30 minutes each. A retrospective SCD cohort (rSCD), comprising patients who underwent the minimum assessment dataset within one year prior to enrollment and were found to be AD plasma biomarker positive, undergoes follow-up at M12 and M24 according to the same longitudinal protocol, with the intervention starting at M12. AI models will integrate multimodal baseline data using machine learning (logistic regression, random forest), deep learning (CNNs for MRI/OCT, RNNs/Transformers for EEG), and survival analysis (Cox proportional hazards, DeepSurv). All models will be validated using k-fold cross-validation with performance metrics including AUC, sensitivity, specificity, balanced accuracy, and positive and negative predictive values.