Beyond black-box AI: Interpretable hybrid systems for dementia care

Area of Science:

• Neuroscience
• Artificial Intelligence
• Clinical Informatics

Background:

• Foundation models (FMs) are increasingly applied to Alzheimer's disease and related dementias (ADRD).
• Current FM adoption at the bedside is limited by interpretability and reliability issues, including opaque inference and hallucinations.
• Weak causal grounding in FMs hinders their clinical utility in complex neurological conditions.

Purpose of the Study:

• To analyze the gaps hindering the clinical adoption of foundation models in ADRD care.
• To propose a hybrid artificial intelligence (AI) framework integrating statistical learning with clinical knowledge and oversight.
• To outline a roadmap for developing accountable and interpretable AI tools for ADRD.

Main Methods:

• Analysis of interpretability and reliability gaps in current AI for ADRD.
• Proposal of a three-level hybrid AI integration framework: knowledge retrieval, contextualized decision support, and adaptive optimization.
• Demonstration of hybrid AI potential using clinical examples and multimodal data.

Main Results:

• Identified key challenges: opaque inference, hallucinations, and weak causal grounding.
• Proposed a hybrid AI framework pairing statistical learning with computable clinical knowledge and clinician oversight.
• Showcased potential applications in biomarker interpretation, multimodal data integration, and digital therapeutics.

Conclusions:

• Hybrid AI offers a path to overcome limitations of current FMs in ADRD.
• A structured integration framework and pragmatic evaluation are crucial for clinical AI adoption.
• The proposed roadmap aims to translate AI advancements into safe, equitable, and effective ADRD care tools.