An AI-first framework for multimodal data in Alzheimer's disease and related dementias

Advancing the understanding and management of Alzheimer's disease and related dementias requires integrating and analyzing diverse data modalities. Traditional diagnostic tools, like neuroimaging, provide valuable insights but are limited by accessibility and infrastructure demands. Meanwhile, emerging modalities, including wearable sensors and speech analysis, enable less invasive and more continuous data collection but introduce challenges related to standardization and privacy. The coexistence of these heterogeneous data streams complicates multimodal integration across cohorts, populations, and clinical settings. Current analytical approaches typically require modality-specific preprocessing pipelines and harmonization methods that were not designed to accommodate modern AI-based capabilities, such as multimodal fusion. In this perspective, we propose an "AI-first" strategy for multimodal data integration that aligns data structuring, harmonization, and modeling within a unified set of guiding principles to optimize modern AI development, while remaining flexible enough to support classical analytical approaches. HIGHLIGHTS: Understanding and managing ADRD requires integrating biological, cognitive, and behavioral data across multiple modalities. Incorporating multiple modalities requires new standards for harmonization and interoperability. Current data platforms are not necessarily built to support multimodal fusion or generalizable AI models across diverse ADRD populations. Modern AI models are capable of learning from messy, multimodal, and incomplete data but require infrastructure designed for this purpose. We propose rethinking ADRD data systems to prioritize AI compatibility, enabling scalable tools for early diagnosis and longitudinal care.