PromptDx: Differentiable Prompt Tuning for Multimodal In-Context Alzheimer's Diagnosis

Researchers introduce PromptDx, a novel AI framework that combines differentiable prompt tuning with multimodal learning to diagnose Alzheimer's Disease using MRI and biomarker data. The method achieves competitive performance using only 1% of context samples compared to 30% in standard approaches, demonstrating significant data efficiency gains for medical imaging applications.

PromptDx addresses a fundamental challenge in applying in-context learning to medical imaging: the disconnect between how AI systems and clinicians approach diagnosis. While traditional deep learning models function as parametric memory systems that apply fixed knowledge learned during training, physicians employ analogical reasoning by comparing new cases to similar historical examples. This research bridges that gap by adapting TabPFN, an in-context learning framework originally designed for tabular data, to handle multimodal medical data through a novel differentiable prompt tuning mechanism.

The technical innovation centers on solving the 'manifold mismatch' problem—where tabular-specific inductive biases and non-differentiable preprocessing pipelines fail when applied to heterogeneous data like 3D MRI scans combined with numerical biomarkers. The lightweight adapter architecture acts as a differentiable surrogate, enabling end-to-end optimization across the entire system.

The practical implications are substantial for clinical deployment. Achieving comparable diagnostic accuracy with 1% rather than 30% of reference samples dramatically reduces the computational burden and data requirements for real-world implementation. This efficiency gain translates to faster diagnosis times and lower infrastructure costs for healthcare institutions. The framework's demonstrated generalizability across six diverse tabular datasets suggests applicability beyond Alzheimer's to other heterogeneous medical diagnosis tasks.

The research sets a precedent for how pre-trained models designed for specific data modalities can be adapted to multimodal contexts through intelligent intermediary layers. Future work likely involves expanding to other neurodegenerative diseases and clinical domains where diagnosis-by-reference paradigms align naturally with physician workflows.

• →PromptDx enables in-context learning for multimodal medical data by solving the gradient flow problem between pre-trained models and heterogeneous inputs.
• →The method achieves diagnostic performance using only 1% of context samples compared to 30% in standard in-context learning approaches.
• →Differentiable prompt tuning acts as a bridge between non-differentiable preprocessing pipelines and end-to-end optimization in medical AI systems.
• →Framework generalization across six tabular datasets indicates applicability beyond Alzheimer's to broader medical diagnosis tasks.
• →The diagnosis-by-reference paradigm more closely aligns AI systems with actual clinical practice compared to traditional parametric memory approaches.