AI-driven neuroimaging for Alzheimer’s disease: Addressing genetic and ethnic bias in diagnostic modeling Authors: Hamna binte Iqbal, Ayesha Janjua, Raghabendra kumar Mahato Published: 2026 To the editor, Alzheimer’s disease is the most prevalent cause of dementia, accounting for over 60% of cases, with its incidence steadily increasing due to global population aging and improved life expectancy. With clinical advancements, Artificial Intelligence (AI) enhances early and more accurate detection by identifying subtle structural and functional brain changes that may not be apparent on conventional MRI- and PET-based evaluation used for detection of preclinical Alzheimer’s disease (AD). AI-driven models combine information from multiple sources such as brain imaging, genetic data, biological markers, and behavioral assessments and capitalize on the key pathological hallmarks of AD to support earlier and more accurate diagnosis while identifying patterns indicative of future disease progression. 1 Despite the promising performance of AI-based neuroimaging tools in AD, a substantial proportion of these datasets, particularly the Alzheimer’s disease neuroimaging initiative (ADNI), contains individuals of European ancestry, which results in limited representation of global populations. These models combine information from multiple sources such as brain imaging, genetic data, biologic markers, and behavioral assessments. This may compromise the reliability and generalizability of these models across diverse populations. Within these data sets, APOE ε4-associated genetic variability is present that significantly influences neuroimaging biomarkers across the Alzheimer’s disease continuum, exhibiting significant variation in allele frequency and disease impact among various ethnic groups. 2 Evidence supporting the genetic and ethnic bias in AI neuroimaging of Alzheimer’s includes a cohort of 1,730 participants in which sensitivity was lower in Hispanic individuals than in non-Hispanic individuals, with true positive rate differences ranging from 20.9% to 27.8% for progression to mild cognitive impairment (MCI) and 24.1% to 48.2% for progression to Alzheimer’s disease (AD). In Black and Asian participants compared to White participants, sensitivity was also lower, with differences of up to 28.4% for AD prediction and up to 25.6% for MCI prediction across different models, which did not meet the present fairness criteria. 3 Furthermore, in a multicenter cohort consisting of 334 AD and 520 cognitively normal individuals, ADNI-based analyses, SVM (Support Vector Machine) and CNN (Convolutional Neural Network) classifiers achieved high internal AUCs (area under the curve) for Alzheimer’s disease classification of approximately 0.94 and 0.93, which decreased to approximately 0.90 and 0.88 on external validation using independent multi-center non-ADNI cohorts. 4 This evidence suggests that AI neuroimaging shows reduced generalizability when applied across different cohorts, mostly due to reliance on non-representative training datasets. This is further indicated by evidence that delayed diagnosis of Alzheimer’s disease affects underrepresented and socioeconomically disadvantaged populations, resulting in clinical delays for timely intervention. Despite advancements in multi-model neuroimaging, the insufficient evaluation of models across diverse demographic subgroups and failure to address ethnic and genetic biases continue to perpetuate healthcare inequities and undermine confidence in AI-enabled clinical decision-making. 5 To conclude, these limitations suggest that more efforts may be needed to improve the evaluation and development of Alzheimer’s AI models. It might be beneficial to reassess existing systems with more diverse populations, including those from underrepresented regions. The use of more inclusive training data and fairness-focused approaches could reduce bias and improve accuracy. ORCID iD Raghabendra kumar Mahato https://orcid.org/0009-0009-9434-1650 References