A Knowledge-Injection Framework for Zero-Shot Adaptation of LLMs to Delirium Prediction

2026-04-27 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Medical Devices & Health Technology · Depth: Expert, extended

Summary

A lightweight knowledge-injection framework significantly improves zero-shot ICU delirium prediction using large language models. This framework augments structured electronic health record data with an external clinical knowledge report at inference time, bypassing fine-tuning or retrieval. Evaluating LLaMA 3.1 8B and LLaMA 3.3 70B on 3,160 ICU admissions from the MIMIC IV dataset, the 8B model's AUROC increased by 8.57 percentage points, reducing its performance gap to a GPT-5.2 reference from 15.66 to 7.09 AUROC points. The 70B model saw a 1.99 percentage point AUROC gain, narrowing its gap from 5.30 to 3.31 points. Crucially, gains depend on clinically meaningful content, with a concise, qualitative knowledge report (v1) outperforming a more granular, numerical one (v2). This approach offers a practical, privacy-preserving solution for resource-constrained clinical settings.

Key takeaway

For Machine Learning Engineers deploying LLMs in clinical settings with limited resources or strict privacy needs, consider implementing inference-time knowledge injection. This method allows smaller, open-weight models like LLaMA 8B to significantly improve zero-shot prediction performance, narrowing the gap to larger frontier models without costly fine-tuning or complex RAG pipelines. Prioritize concise, qualitative knowledge reports over overly prescriptive ones, as they prove more effective and can even reduce inference time for larger models.

Key insights

Inference-time knowledge injection significantly enhances smaller LLMs' zero-shot clinical prediction without fine-tuning or retrieval.

Principles

• Smaller LLMs benefit most from external knowledge injection.
• Qualitative, flexible knowledge reports outperform overly granular ones.
• Semantic relevance of injected knowledge is crucial for performance gains.

Method

Generate a task-specific clinical knowledge report, summarize patient EHR data into text, then combine both into a single prompt for an off-the-shelf LLM.

In practice

• Design knowledge reports with flexible, qualitative guidance, not strict numerical thresholds.
• Consider open-weight LLMs with knowledge injection for privacy-sensitive on-premises deployments.
• Use SHAP analysis to confirm LLM engagement with injected knowledge sections.

Topics

• Large Language Models
• Zero-Shot Prediction
• Clinical AI
• Delirium Prediction
• Knowledge Injection
• EHR Data

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.
• In2Lab-TNT at #SMM4H-HeaRD 2026: An Application of QTT’s Terminological Entanglement to Leverage Insomnia Detection in Clinical Notes — Modeling clinical concept dependencies improves LLM-based insomnia detection, boosting recall and stabilizing performance.
• Reading between the Lines: Leveraging Large Language Models for Global Dementia and Depression Assessment from Clinical Interviews — LLMs can effectively assess neuropsychiatric disorder severity from speech, enabling automated screening.
• Vasudev Awatramani at #SMM4H-HeaRD 2026: A Two-Pass LLM Pipeline with Deterministic Rule Derivation for Interpretable Insomnia Detection in Clinical Notes — Combining LLM evidence extraction with deterministic rule application provides interpretable clinical label prediction.
• No More RAG: Parametric Knowledge Injection into LLMs — Injecting knowledge at the final LLM layer via LoRA preserves inference efficiency by avoiding KV cache invalidation.
• T2D-Bench: Evidence-Gated Evaluation of LLM Outputs for Type 2 Diabetes Using a Multi-Layer Clinical-Lifestyle Knowledge Graph — LLMs require evidence-gated evaluation against clinical knowledge graphs to ensure medical accuracy and guideline compliance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.