SPARROW: subtyping Parkinson's disease with agentic reasoning and robust omics workflow

Area of Science:

• Neuroscience
• Computational Biology
• Medical Informatics

Background:

• Parkinson's disease (PD) is a complex neurodegenerative disorder with varied clinical presentations, making early and precise subtyping difficult.
• Integrating multimodal data (genomics, imaging, clinical, cognitive) for PD subtyping is challenging due to data complexity and missing values.
• Current tools lack domain reasoning and interpretability, limiting their clinical usefulness.

Purpose of the Study:

• To present SPARROW, a multimodal framework designed for accurate and interpretable Parkinson's disease subtyping.
• To unify diverse data sources including omics, MRI, clinical, and cognitive assessments into a semantic knowledge space.
• To enhance the clinical utility of PD subtyping through transparent decision-making.

Main Methods:

• SPARROW employs specialist modules for omics and MRI analysis, generating ontology-driven outputs.
• A large language model-based reasoner utilizes chain-of-thought reasoning to interpret these structured outputs.
• The framework integrates genomic, imaging, clinical, and cognitive data within a unified semantic knowledge space.

Main Results:

• SPARROW demonstrated superior performance in classifying all Parkinson's disease subtypes using baseline data from the Parkinson's Progression Markers Initiative (PPMI) dataset.
• The framework achieved high accuracy in a zero-shot setting, indicating robust generalization capabilities.
• The interpretability of SPARROW allows for transparent subtype decisions, showing the contribution of each data source.

Conclusions:

• SPARROW offers a promising approach for accurate and interpretable subtyping of Parkinson's disease.
• The framework's ability to integrate and reason over multimodal data enhances clinical decision-making for PD.
• Future applications of SPARROW could significantly improve personalized medicine strategies for Parkinson's disease patients.