Recent advances in large language models (LLMs) have enabled promising progress in diagnosis prediction from electronic health records (EHRs). However, existing LLM-based approaches tend to overfit to historically observed diagnoses, often overlooking novel yet clinically important conditions that are critical for early intervention. To address this, we propose EviCare, an in-context reasoning framework that integrates deep model guidance into LLM-based diagnosis prediction. Rather than prompting LLMs directly with raw EHR inputs, EviCare performs (1) deep model inference for candidate selection, (2) evidential prioritization for set-based EHRs, and (3) relational evidence construction for novel diagnosis prediction. These signals are then composed into an adaptive in-context prompt to guide LLM reasoning in an accurate and interpretable manner. Extensive experiments on two real-world EHR benchmarks (MIMIC-III and MIMIC-IV) demonstrate that EviCare achieves significant performance gains, which consistently outperforms both LLM-only and deep model-only baselines by an average of 20.65% across precision and accuracy metrics. The improvements are particularly notable in challenging novel diagnosis prediction, yielding average improvements of 30.97%.

Graphs with abundant attributes are essential in modeling interconnected entities and enhancing predictions across various real-world applications. Traditional Graph Neural Networks (GNNs) often require re-training for different graph tasks and datasets. Although the emergence of Large Language Models (LLMs) has introduced new paradigms in natural language processing, their potential for generic graph mining-training a single model to simultaneously handle diverse tasks and datasets-remains under-explored. To this end, our novel framework ${\sf MuseGraph}$MuseGraph, seamlessly integrates the strengths of GNNs and LLMs into one foundation model for graph mining across tasks and datasets. This framework first features a compact graph description to encapsulate key graph information within language token limitations. Then, we propose a diverse instruction generation mechanism with Chain-of-Thought (CoT)-based instruction packages to distill the reasoning capabilities from advanced LLMs like GPT-4. Finally, we design a graph-aware instruction tuning strategy to facilitate mutual enhancement across multiple tasks and datasets while preventing catastrophic forgetting of LLMs' generative abilities. Our experimental results demonstrate significant improvements in five graph tasks and ten datasets, showcasing the potential of our ${\sf MuseGraph}$MuseGraph in enhancing the accuracy of graph-oriented downstream tasks while improving the generation abilities of LLMs.