Stratifying heterogeneity in mild cognitive impairment: A multimodal subtype and stage inference approach.
Authors
Affiliations (7)
Affiliations (7)
- Innovation Health Center, Einstein Israeli Hospital, São Paulo, Brazil. Electronic address: [email protected].
- InBrain Lab, Department of Physics, FFCLRP, University of São Paulo, Ribeirão Preto, SP, Brazil; Graduate Program in Neurology, FMRP, University of São Paulo, Ribeirão Preto, SP, Brazil. Electronic address: [email protected].
- UCL Hawkes Institute and Department of Computer Science, University College London, London, UK. Electronic address: [email protected].
- UCL Hawkes Institute and Department of Computer Science, University College London, London, UK. Electronic address: [email protected].
- UCL Hawkes Institute and Department of Computer Science, University College London, London, UK. Electronic address: [email protected].
- UCL Hawkes Institute and Department of Computer Science, University College London, London, UK. Electronic address: [email protected].
- InBrain Lab, Department of Physics, FFCLRP, University of São Paulo, Ribeirão Preto, SP, Brazil; Graduate Program in Neurology, FMRP, University of São Paulo, Ribeirão Preto, SP, Brazil; Department of Medical Imaging, Hematology and Clinical Oncology, FMRP, University of São Paulo, Ribeirão Preto, SP, Brazil. Electronic address: [email protected].
Abstract
This study aims to identify distinct clinical and biological subtypes of Mild Cognitive Impairment (MCI) and their progression stages using a data-driven multimodal approach, aimed at improving the stratification and prediction of conversion to Alzheimer's disease (AD). We applied the Subtype and Stage Inference (SuStaIn) algorithm to a multimodal dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI) to identify distinct MCI subtypes and their progression stages on 558 patients classified as MCI amyloid-positive. Our model incorporated T1-weighted structural MRI (measuring cortical, subcortical, and hippocampal volumes), CSF biomarkers (Aβ<sub>1-42</sub> and p-TAU), APOE-ε4 status, and a comprehensive set of neuropsychological scores. We delimited four subtypes, which two exhibited higher risk for conversion to AD. The remaining two were characterized by prominent ventricular enlargement and a significant association with geriatric depression, respectively. Predictive performance was driven mainly by the class of input features: a multimodal model using individual biomarkers alone achieved the best discrimination (AUC 0.865; accuracy 0.810), and adding SuStaIn subtype and stage did not improve it. However, subtype and stage in isolation (two variables rather than sixteen) reached an AUC of 0.811 with the highest recall observed (0.867), statistically indistinguishable from the full model. Thus, applying SuStaIn to multidomain data successfully disentangles the phenotypic heterogeneity of MCI. By separating temporal disease progression (staging) from inherent trait differences (subtyping), this framework provides a more refined diagnostic tool for patient stratification and early intervention in the AD continuum.