Abstract
Differential diagnosis of dementia remains a challenge in neurology due to symptom overlap across etiologies, yet it is crucial for formulating early, personalized management strategies. Here, we present an AI model that harnesses a broad array of data, including demographics, individual and family medical history, medication use, neuropsychological assessments, functional evaluations, and multimodal neuroimaging, to identify the etiologies contributing to dementia in individuals. The study, drawing on 51,269 participants across 9 independent, geographically diverse datasets, facilitated the identification of 10 distinct dementia etiologies. It aligns diagnoses with similar management strategies, ensuring robust predictions even with incomplete data. Our model achieved a micro-averaged area under the receiver operating characteristic curve (AUROC) of 0.94 in classifying individuals with normal cognition, mild cognitive impairment and dementia. Also, the micro-averaged AUROC was 0.96 in differentiating the dementia etiologies. Our model demonstrated proficiency in addressing mixed dementia cases, with a mean AUROC of 0.78 for two co-occurring pathologies. In a randomly selected subset of 100 cases, the AUROC of neurologist assessments augmented by our AI model exceeded neurologist-only evaluations by 26.25%. Furthermore, our model predictions aligned with biomarker evidence and its associations with different proteinopathies were substantiated through postmortem findings. Our framework has the potential to be integrated as a screening tool for dementia in various clinical settings and drug trials, with promising implications for person-level management.
Introduction
Dementia is a complex syndrome rather than a single disease, arising from various underlying pathologies such as Alzheimer’s disease (AD), Lewy body dementia (LBD), and vascular dementia (VD). In clinical practice, these etiologies frequently overlap, leading to "mixed dementia" cases that are difficult to diagnose accurately. Current diagnostic processes face several hurdles: significant symptom overlap in early stages, a global shortage of specialized neurologists, and the inherent complexity of multimodal clinical data (e.g., MRI scans, cognitive tests, and medical history). Furthermore, real-world clinical records are often incomplete, posing a challenge for traditional diagnostic models. This study introduces a multimodal AI framework designed to address these challenges by providing a comprehensive differential diagnosis across ten distinct dementia etiologies.
Methods
The researchers developed a Transformer-based architecture capable of integrating diverse data types. The model processes demographics, family medical history, medication use, neurological test scores (such as MMSE or MoCA), and multi-sequence MRI scans (T1, T2, FLAIR).
Multimodal Integration and Architecture
Each data modality is converted into a fixed-length vector using specific embeddings. For 3D neuroimaging data, a Swin UNETR (a 3D transformer) extracts relevant features. To ensure the model remains functional in real-world scenarios where data may be missing, the authors employed random feature masking during training. This technique teaches the model to make robust predictions even in the absence of specific scans or test results.
Classification Categories
The model was designed to classify patients into three cognitive statuses and ten distinct dementia etiologies. These categories were selected to align with clinical management pathways, grouping related conditions to facilitate practical treatment decisions.
The training utilized a massive dataset of 51,269 participants from nine independent cohorts, including the National Alzheimer’s Coordinating Center (NACC), ADNI, and the Framingham Heart Study (FHS).
Results
The AI model demonstrated exceptional diagnostic performance. It achieved a micro-averaged AUROC of 0.94 for classifying cognitive status (Normal vs. MCI vs. Dementia) and 0.96 for differentiating between the ten dementia etiologies.
Resilience and Human-AI Synergy
The model proved resilient to missing data; on the ADNI dataset (which had 69% missing data relative to the training set), it maintained a weighted-average AUROC of 0.91. Crucially, the study highlighted the potential for human-AI collaboration. In a subset of 100 cases, AI-augmented neurologist assessments outperformed neurologist-only evaluations by 26.25%. For neuroradiologists, the performance boost was 17.22%.
Biological and Pathological Validation
The model's predictions showed strong alignment with biological markers and postmortem findings. Model probabilities for AD increased in correlation with Braak stages, which measure the spread of neurofibrillary tangles in the brain. Additionally, the model achieved a mean AUROC of 0.78 in identifying mixed pathology cases, a task that remains notoriously difficult for clinicians.
Discussion
While the model shows high accuracy, certain limitations remain. The dataset is heavily skewed toward Alzheimer’s disease, making the recognition of rarer subtypes like Prion disease more challenging. Furthermore, the current model groups different stages of dementia into a single category to focus on etiology rather than progression. Future work will involve integrating disease staging and validating the framework in primary care settings and across more diverse global populations. Despite these challenges, this multimodal AI framework represents a significant step toward personalized management and improved screening in clinical trials for dementia.
Summary updated