AI Model Revolutionizes Dementia Diagnosis with High Accuracy Across Multiple Data Sources

A recent study published in the journal Nature MedicineResearchers have developed and validated an artificial intelligence (AI) model that uses multimodal data to accurately distinguish between different etiologies of dementia (significant cognitive decline) and improve earlier and more personalized management.

study: AI-based differential diagnosis of dementia etiology based on multimodal dataImage credit: PopTika / Shutterstock

background

Dementia affects nearly 10 million people every year and represents a major clinical and socio-economic challenge. Accurate diagnosis is essential for effective treatment, but overlapping symptoms across different types make it difficult to diagnose. As the population ages and the demand for accurate diagnosis in drug trials increases, improved tools are urgently needed. A shortage of experts exacerbates this problem, highlighting the need for scalable solutions. Further research is needed to evaluate the impact of AI models on healthcare outcomes and their integration into clinical practice.

About the Research

The study enrolled 51,269 participants from nine cohorts and collected comprehensive data including demographics, medical history, laboratory results, physical and neurological examinations, medications, neuropsychological testing, functional assessments, and multi-sequence magnetic resonance imaging (MRI) scans. Participants or their informants provided written informed consent, and the protocol was approved by the institution's ethical review board. Cohorts included individuals with normal cognition (NC) (healthy brain function, 19,849 individuals), mild cognitive impairment (MCI) (mild cognitive decline, 9,357 individuals), and dementia (22,063 individuals).

OneThe dementia differential diagnosis model was developed using diverse data modalities, including individual-level demographics, health history, neurological examination, physical/neurological examination, and multi-sequence MRI scans. These data sources were aggregated, whenever available, from nine independent cohorts: 4RTNI, ADNI, AIBL, FHS, LBDSU, NACC, NIFD, OASIS, and PPMI (Table 1 and S1The model was trained using combined data from NACC, AIBL, PPMI, NIFD, LBDSU, OASIS, and 4RTNI. A subset of the NACC dataset was used for internal testing. External validation was performed using the ADNI and FHS cohorts. bThe Transformer served as the scaffolding for the model. Each feature was processed into a fixed-length vector using a modality-specific embedding (emb.) strategy and fed into the Transformer as input. A linear layer was used to connect the Transformer and the output prediction layer. cA subset of the NACC test dataset was randomly selected to perform a comparative analysis of neurologists' performance augmented with the AI ​​model versus performance without AI assistance. Similarly, a comparative evaluation was performed with practicing neuroradiologists who were provided with a randomly selected sample of confirmed dementia cases from the NACC test cohort to evaluate the impact of AI augmentation on diagnostic performance. In both of these evaluations, the model and clinicians had access to the same multimodal data set. Finally, we evaluated our model predictions against biomarker profiles and pathology grades available from the NACC, ADNI, and FHS cohorts.

Dementia cases were further classified into Alzheimer's disease (AD) (amnesic dementia, 17,346), Lewy body dementia (hallucinations and movement disorder) and Parkinson's disease (movement disorder with dementia) (LBD, 2,003), vascular dementia (VD) (cognitive decline due to reduced cerebral blood flow, 2,032), prion diseases (PRD) (rapid neurodegenerative disease, 114), frontotemporal dementia (FTD) (deterioration of personality and language function, 3,076), normal pressure hydrocephalus (NPH) (dementia-like symptoms due to fluid accumulation, 138), systemic and external causes of dementia (SEF, 808), psychiatric illness (PSY, 2,700), traumatic brain injury (TBI, 265), and other causes (ODE, 1,234).

The study used data from the National Alzheimer's Coordinating Center (NACC), Alzheimer's Disease Neuroimaging Initiative (ADNI), Frontotemporal Dementia (FTD) Neuroimaging Initiative (NIFD), Parkinson's Progression Markers Initiative (PPMI), Australian Imaging, Biomarkers, and Lifestyle Ageing Flagship Study (AIBL), Open Access Series of Imaging Studies-3 (OASIS), 4-Repeat Tauopathy Neuroimaging Initiative (4RTNI), Stanford Lewy Body Dementia Center of Excellence (LBDSU), and the Framingham Heart Study (FHS). Eligibility criteria were having a diagnosis of NC, MCI, or dementia, with NACC data as baseline. Data from other cohorts were standardized using the Uniform Data Set (UDS) dictionary. An innovative model training approach addressed missing features and labels, ensured robust data utilization, and maximized sample size.

research result

This study leverages multimodal data to strictly classify dementia into 13 neurologist-defined diagnostic categories along clinical management pathways. LBD and Parkinson's disease dementia are classified under LBD due to their similar treatment pathways, while VD includes stroke symptom cases managed by stroke specialists. Psychiatric disorders such as schizophrenia and depression are classified under PSY.

The model performed well on NC, MCI, and dementia test cases, achieving a micro-averaged Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.94 and Area Under the Precision Recall Curve (AUPR) of 0.90. The model outperformed CatBoost on the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Framingham Heart Study (FHS) datasets, highlighting its superior diagnostic accuracy.

Shapley analysis identified key features influencing the diagnostic decision: cognitive status, Montreal Cognitive Assessment (MoCA) score, and memory task performance for predicting NC, memory-related features, functional impairment, and T1-weighted MRI for predicting MCI, and functional impairment, poor Mini-Mental State Examination (MMSE) score, and apolipoprotein E4 (APOE4) allele for predicting dementia.

The model demonstrated tolerance to incomplete data and maintained reliable scores even with missing features. Validation on external datasets such as ADNI and FHS showed good performance despite significant missing data, with weighted average AUROC and AUPR scores of 0.91 and 0.86 for ADNI and 0.68 and 0.53 for FHS, respectively.

When assessing consistency with prodromal Alzheimer's disease (AD), the model consistently attributed high AD probability to MCI cases associated with AD, reinforcing its utility in early detection of the disease. Comparison with Clinical Dementia Rating (CDR) across NACC, ADNI, and FHS datasets was strongly correlated with CDR scores, highlighting the model's sensitivity to staged clinical dementia ratings.

The model demonstrated strong diagnostic ability across 10 different dementia etiologies, with micro-average AUROC and AUPR values ​​of 0.96 and 0.70, respectively. Although variability in AUPR scores indicated challenges in identifying less common and complex dementias, the model demonstrated robust performance across demographic subgroups.

When the model predicted probabilities were matched to AD, FTD, and LBD biomarkers, the model showed clear differentiation between biomarker negative and positive groups, validating its validity in capturing the pathophysiology of dementia. Validation with postmortem data further supported the model's ability to match probability scores to neuropathological markers.

AI-assisted clinician evaluation significantly improved diagnostic performance, with improved AUROC and AUPR scores across all categories, demonstrating the model's potential to enhance clinical dementia diagnosis.

Conclusion

In this study, we present an AI model for the differential diagnosis of dementia using multimodal data. Unlike previous models, our model distinguishes between different dementia etiologies, including AD, VD, and LBD, which is essential for personalized treatment strategies. Validated in different cohorts, the model's predictions were supported by biomarker and postmortem data. We highlight that combining the model's predictions with neurologists' assessments outperforms neurologists' assessments alone, potentially increasing diagnostic accuracy. Our model addresses mixed dementias by providing probability scores for each etiology, improving clinical decision-making.