PromptDx: Differentiable Prompt Tuning for Multimodal In-Context Alzheimer's Diagnosis
Researchers have introduced PromptDx, a novel framework for diagnosing Alzheimer's disease that mimics clinical analogical reasoning by referencing past cases. Unlike traditional deep learning models that rely on fixed parametric memory, PromptDx utilizes In-Context Learning (ICL) to diagnose new patients based on similar historical records. The system addresses limitations in existing ICL frameworks, such as TabPFN, which struggle with heterogeneous multimodal data due to non-differentiable preprocessing. PromptDx employs a Differentiable Prompt Tuning (DPT) mechanism, using a lightweight adapter to align masked multimodal modeling with the pre-trained ICL engine. Validated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset using 3D MRI and tabular biomarkers, the method outperforms traditional baselines. Notably, it achieves superior diagnostic performance using only 1% of context samples, compared to 30% required by standard ICL methods, demonstrating significant data efficiency and manifold condensation capabilities. The framework also shows generalizability across various tabular datasets, offering a more clinically aligned and efficient paradigm for medical diagnosis.
Wire timeline
PromptDx: Differentiable Prompt Tuning for Multimodal In-Context Alzheimer's Diagnosis
Researchers have introduced PromptDx, a novel framework for diagnosing Alzheimer's disease that mimics clinical analogical reasoning by referencing past cases. Unlike traditional deep learning models that rely on fixed parametric memory, PromptDx utilizes In-Context Learning (ICL) to diagnose new patients based on similar historical records. The system addresses limitations in existing ICL frameworks, such as TabPFN, which struggle with heterogeneous multimodal data due to non-differentiable preprocessing. PromptDx employs a Differentiable Prompt Tuning (DPT) mechanism, using a lightweight adapter to align masked multimodal modeling with the pre-trained ICL engine. Validated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset using 3D MRI and tabular biomarkers, the method outperforms traditional baselines. Notably, it achieves superior diagnostic performance using only 1% of context samples, compared to 30% required by standard ICL methods, demonstrating significant data efficiency and manifold condensation capabilities. The framework also shows generalizability across various tabular datasets, offering a more clinically aligned and efficient paradigm for medical diagnosis.
cs.AI updates on arXiv.org