Parkinson’s disease (PD) is a heterogeneous neurodegenerative disorder that presents a wide spectrum of clinical phenotypes, posing a fundamental challenge for early and accurate subtyping. This requires a multimodal assessment to identify disease patterns. Robust integration of multi-omics data, brain MRI, clinical biomarkers and cognitive assessments remains challenging, especially when data are missing or incomplete. Existing tools often lack domain-driven reasoning and offer limited interpretability, undermining their clinical utility. This paper presents SPARROW, a multimodal framework that unifies genomic, imaging, clinical and cognitive data in a common semantic knowledge space for PD subtyping. Within SPARROW, specialist modules for omics and MRI analysis provide structured ontology-driven outputs that a large language model-based reasoner then interprets via chain-of-thought reasoning. This approach achieves transparency in subtype decisions by highlighting how each data source contributes to the final classification. Applied to the Parkinson’s Progression Markers Initiative (PPMI) dataset, SPARROW achieved superior performance on classification of all subtypes using the baseline visit data in a zero-shot setting. Our findings underscore the potential of SPARROW for accurate and interpretable PD subtyping in clinical settings
Reyes et al. (Wed,) studied this question.