What question did this study set out to answer?

The aim is to create multimodal models for classifying Alzheimer’s disease stages and predicting conversion from mild cognitive impairment to Alzheimer’s disease.

May 3, 2026Open Access

Classification and prediction of Alzheimer’s disease stages and conversion from mild cognitive impairment based on multimodal data fusion

Key Points

The aim is to create multimodal models for classifying Alzheimer’s disease stages and predicting conversion from mild cognitive impairment to Alzheimer’s disease.
Single-center retrospective cohort study with 658 individuals aged ≥50 years
Collected data included demographic info, neurocognitive assessments, MRI features, and blood-based biomarkers
Used logistic regression for classification and prediction, evaluated performance with AUC and calibration curves
Significant differences noted across disease stages in cognitive and biomarker assessments (all p < 0.05)
Multimodal models achieved high classification performance (CN vs. AD: AUC = 0.96; MCI vs. AD: AUC = 0.86)
Conversion prediction model showed good discrimination (AUC = 0.87) with improved performance using hippocampal volume (AUC = 0.88)

Abstract

ObjectiveThis study aimed to develop multimodal prediction models based on real-world clinical data for classifying different stages of Alzheimer’s disease (AD) and for predicting the conversion from mild cognitive impairment (MCI) to AD.MethodsA single-center retrospective real-world cohort study was conducted. A total of 658 individuals aged ≥50 years were included and classified into cognitively normal (CN), MCI, and AD groups. Demographic characteristics, neurocognitive assessment results, conventional magnetic resonance imaging (MRI) features, and blood-based biomarkers were collected. Logistic regression was used to construct pairwise classification models for disease stages and prediction models for MCI-to-AD conversion. Model performance was evaluated through stepwise integration of multimodal features. Discrimination ability was assessed using the area under the receiver operating characteristic curve (AUC), together with calibration curves and decision curve analysis. In a sub-cohort with thin-slice MRI data, the additional value of hippocampal volume was further examined.ResultsSignificant differences were observed among disease stages in cognitive function, imaging markers, and blood biomarkers (all p < 0.05). Multimodal fusion models achieved the best performance in disease stage classification (CN vs. AD: AUC = 0.96 ± 0.01; MCI vs. AD: AUC = 0.86 ± 0.03). The conversion prediction model integrating multimodal features showed good discrimination (AUC = 0.87) and satisfactory calibration. In the thin-slice MRI sub-cohort, inclusion of hippocampal volume increased the AUC to 0.88.ConclusionMultimodal prediction models based on real-world clinical data improved the accuracy of AD stage classification and the prediction of MCI-to-AD conversion risk. These models demonstrated good clinical feasibility. High-resolution structural imaging markers further enhanced predictive performance in selected populations.

KI fragen

Bookmark

View Full Paper