Data-driven subtyping of Parkinson’s disease using MRI: current insights, challenges, and future directions

Parkinson’s disease (PD) is a heterogeneous neurodegenerative disorder marked by diverse motor and non-motor symptom profiles. Traditional symptom-based subtyping shows limited stability and lacks clear biological grounding. Integrating magnetic resonance imaging (MRI) with machine learning (ML) offers a promising avenue for defining biologically informed PD subtypes. This narrative review synthesizes evidence from MRI-based subtyping studies that used structural (T1-weighted), diffusion, functional, or multimodal MRI features as primary inputs for unsupervised or hybrid ML approaches to derive PD subtypes and outlines key methodological challenges and future translational needs. T1-weighted MRI studies consistently identify two to three subtypes characterized by distinct patterns of cortical and subcortical atrophy associated with variation in motor and non-motor symptoms. Although fewer in number, diffusion MRI studies have identified microstructural heterogeneity in PD. However, the findings remain heterogeneous and preliminary, and a stable subtyping framework has yet to be established. Multimodal MRI approaches show that combining modalities provides complementary insights into the neurobiology underlying PD heterogeneity but require further validation. Collectively, MRI-based subtyping shows promise for mapping clinical variability onto neuroanatomical patterns. At present, these subtypes are best viewed as research constructs that illuminate disease variability rather than clinical diagnostic tools. Translation into clinical practice will require addressing critical methodological gaps to achieve the reproducibility and prognostic utility necessary for precision medicine.