Disease-Informed Prompting

• Disease-informed prompting is a structured paradigm that encodes targeted medical terminology, clinical guidelines, and diagnostic frameworks into AI prompts.
• It enhances clinical reasoning, supports diagnostic report generation, and improves recognition of rare diseases through explicit ontology use and reasoning scaffolds.
• This approach boosts performance and interpretability in safety-critical applications by calibrating probabilities and enforcing structured, stepwise clinical reasoning.

Disease-informed prompting is a structured paradigm for guiding LLMs, vision-LLMs (VLMs), or hybrid AI systems toward clinically meaningful, reliable outputs by explicitly encoding disease-specific information, reasoning processes, or ontologies within the prompt. Unlike generic prompt engineering—which leverages broad templates or domain-agnostic instructions—disease-informed prompting integrates targeted terminology, expert knowledge, pathophysiological mechanisms, diagnostic frameworks, and context-specific disambiguation rules. This approach underpins a range of applications, including clinical reasoning, diagnostic report generation, rare disease entity recognition, EHR-based prediction, imaging-based classification, and robust chatbot behavior in safety-critical domains.

1. Core Principles and Taxonomy

Disease-informed prompting is characterized by three core tenets: (a) explicit integration of disease-centric terminology or structures, (b) alignment with expert clinical reasoning or specialty guidelines, and (c) systematic adaptation of prompt content and structure based on the risk, rarity, or complexity of the disease under study (Chen et al., 25 Aug 2025). This paradigm encompasses multiple methodological axes:

• Disease-specific lexical enrichment: Incorporation of pathognomonic features, biomarker references, or classification schemas directly within the prompt (Xi et al., 12 Aug 2025, Guo et al., 2024).
• Reasoning process scaffolding: Inclusion of stepwise clinical reasoning—such as differential diagnosis enumeration, analytic pathophysiology, or Bayesian inference—mirroring clinician decision strategies (Savage et al., 2023).
• Structural and ontological encoding: Leveraging domain ontologies (e.g., UMLS, SNOMED, ICD), semantic causal graphs, or curated disease concept banks to constrain and inform the model's latent state (Chen et al., 25 Aug 2025, Zhao et al., 24 Oct 2025, Mehta et al., 4 Mar 2025).
• Task-specific response guidance: Prompting for structured outputs (e.g., forced-decision JSON, disease label tokens, calibrated probabilities), often with hard constraints to suppress ambiguity (Sweidan et al., 24 Sep 2025).

Table 1: Exemplars of Disease-Informed Prompting Components

Paper / Domain	Disease-Specific Elements	Prompting Modality
(Savage et al., 2023) Diagnostic	Structured differential diagnosis,	Reasoning-type instruction
Reasoning in Medicine	Bayesian inference	Few-shot with rationale
(Xi et al., 12 Aug 2025) Rare Disease NER	Ontology definitions, disambiguation rules	Zero-/few-shot + task schema
(Guo et al., 2024) LVLM VQA	Pathology definitions, imaging signs	Explanation + agent referral
(Zhao et al., 24 Oct 2025) EGO-Prompt	Semantic causal graph (SCG)	Multi-step prompt evolution

2. Prompt Frameworks and Engineering Strategies

Disease-informed prompting frameworks commonly involve modular, composable templates that encode both general clinical context and nuanced disease knowledge. Strategies include:

a) Structured Prompt Templates

A prevalent schema concatenates:

1. Role/Instruction block, e.g. “You are a clinical evaluator for [disease].”
2. Task description and feature specification, sometimes listing normal/abnormal ranges (e.g. in EHR prediction (Zhu et al., 2024)).
3. Domain-aware context, such as detailed entity definitions, prevalence boundaries, or diagnostic criteria (Xi et al., 12 Aug 2025).
4. In-context examples, carefully chosen by semantic similarity to the test case or balanced on rare/majority classes (Xi et al., 12 Aug 2025, Sweidan et al., 24 Sep 2025).
5. Constraining output schema (e.g., forced-decision JSON with confidence scores, explicit labeling, closed-class selection) (Sweidan et al., 24 Sep 2025, Chen et al., 2024).

b) Multi-Dimensional Evaluation and Optimization

Frameworks such as EMPOWER and EGO-Prompt evaluate prompts along disease-weighted axes: clarity, domain specificity, guideline alignment, and predicted factual accuracy. These are optimally combined (with disease- or task-dependent weights) in evolutionary or reinforcement-learning loops that iteratively refine candidate prompts by textual gradients or domain expert feedback (Chen et al., 25 Aug 2025, Zhao et al., 24 Oct 2025).

c) Hybrid Model Integration

Some systems leverage population-level encoders or cross-modal retrieval modules as external “advisors” or context enhancers, with their representations injected into the prompt as soft tokens, summarized features, or structured results (Moghaddam et al., 20 Apr 2026, Jin et al., 2023).

3. Methodologies for Reasoning, Disambiguation, and Knowledge Injection

a) Emulating Clinical Reasoning

In diagnostic tasks, prompts can instruct LLMs to perform stepwise differential diagnosis, analytic pathophysiology, or explicit Bayesian updating over disease priors and evidence (e.g., via recursive rationales or probability chains). Representative templates enumerate key features (“petechiae,” “ST elevations”) and force the model to map case findings to pathophysiological rules or likelihood updates (Savage et al., 2023).

b) Handling Rare, Ambiguous, or Underrepresented Diseases

For rare disease NER or minority-class VQA, prompts introduce ontology-backed definitions, tight output format controls, and domain-specific disambiguation heuristics (e.g., excluding common diseases, enforcing dictionary constraints, and adding “none if absent” rules) (Xi et al., 12 Aug 2025, Guo et al., 2024). In few-shot regimes, semantically-guided example selection (cluster-KNN, inquiry-KNN) efficiently spans the relevant entity landscape (Xi et al., 12 Aug 2025).

c) Calibrated Probabilities and Interpretability

Advanced frameworks incorporate external probabilistic anchors (e.g., mapping MMSE scores to calibrated AD probabilities), interpretability-by-design via concept-based bottlenecks (as in concept-guided VLMs for retinal disease (Mehta et al., 4 Mar 2025)), or chain-of-thought rationales for auditability (Sweidan et al., 24 Sep 2025, Savage et al., 2023).

4. Applications Across Modalities and Clinical Domains

a) Electronic Health Records (EHR) and Structured Clinical Prediction

Zero-shot prompting of LLMs with unit- and range-augmented EHR tables achieves superior predictive performance to domain-adapted ML/DL across hospital mortality, readmission, and length-of-stay tasks. This is most effective when explicit feature context and role instructions are included (Zhu et al., 2024). Recurrent prompt tuning, as in RePrompT, integrates both longitudinal state and population-level disease embeddings via soft prompt tokens, facilitating superior performance on multi-visit EHR sequences (Moghaddam et al., 20 Apr 2026).

b) Medical Report Generation and Image-based Diagnostic Reasoning

Diagnosis-driven prompts for radiology report generation (PromptMRG) bridge disease classification output and token-level prompt vectors in the decoder, sharply improving clinical efficacy metrics over classic NLG-only architectures. Cross-modal retrieval and logit-adjusted loss further aid underrepresented conditions (Jin et al., 2023). Pathology-detailed VQA prompts or weak-learner-augmented agent judgments suppress hallucinations and enhance recall on imbalanced imaging datasets (Guo et al., 2024).

c) Multimodal and Multi-Agent Chatbots

In high-risk domains such as schizophrenia education, disease-informed prompts enforce strict scope, knowledge sourcing, and disclaimer procedures, with multi-agent critical analysis layers (CAF) dynamically filtering and revising outputs to maintain compliant, non-hazardous conversations (Waaler et al., 2024).

d) Neurological Disease and Graph-based Analysis

Multi-level prompt enrichment, as in BrainPrompt, injects anatomical, demographic, and disease-stage knowledge at ROI-, subject-, and label-level into GNNs, providing both improved generalization in low-data fMRI cohorts and direct interpretability via alignment with clinical progression markers (Xu et al., 12 Apr 2025).

5. Quantitative Performance, Error Analysis, and Empirical Trends

Disease-informed prompting demonstrably increases both performance and robustness across diverse tasks and modalities. Representative findings include:

• Rare disease NER: F1 scores rise from 0.702 (zero-shot) to 0.776 (few-shot, k=8), peaking at 0.837 with fine-tuning; post-filtering and ablation indicate that explicit definitions and demonstration choice are decisive (Xi et al., 12 Aug 2025).
• Clinical prediction from EHR: Zero-shot GPT-4 with knowledge-infused prompts achieves AUROC 93.55 (TJH mortality) vs 85.01 for XGBoost (10-shots), with sharply reduced missing rates when unit/range context is present (Zhu et al., 2024).
• Medical VQA: Pathology-explanation prompting lifts F1 by up to 0.27, increases recall on minority classes by 0.07, and curbs hallucinations (as assessed by POPE metrics) (Guo et al., 2024).
• PromptMRG report generation: Clinical efficacy F1 climbs from 0.447 (prior SOTA) to 0.476, with the largest per-disease F1 gains for rare findings (+12%) (Jin et al., 2023).
• BrainPrompt: Multi-level prompt integration yields +7.54% accuracy improvement in ADNI 5-way disease staging over GCN, with node-level IG spotlighting domain-valid ROIs (Xu et al., 12 Apr 2025).
• Adversarial chatbot safety: Critical Analysis Filter boosts compliance (C≥2) from 8.7% to 67%; high-compliance replies (C≥3) rise drastically in adversarial roles (Waaler et al., 2024).

Typical error taxonomies in disease-informed NER include boundary drift, type confusion, and spurious/missed entity errors—guiding prompt refinement and post-processing heuristics (Xi et al., 12 Aug 2025).

6. Best Practices and Theoretical Implications

• Prompt modularity and explicit constraints (e.g., slot-based templates, label enumeration) focus model attention, reduce off-label errors, and unify system output for downstream processing (Chen et al., 2024, Xi et al., 12 Aug 2025).
• Balanced examples and semantic proximity in few-shot settings enable high performance with minimal labeling cost—performance for rare diseases saturates quickly at k≈4–8 demonstrations (Xi et al., 12 Aug 2025, Sweidan et al., 24 Sep 2025).
• Guideline/ontology integration and real-time feedback (EMPOWER, EGO-Prompt) allow continuous prompt optimization, domain adaptation, and causal reasoning refinement, tailoring system outputs to real-world clinical priorities (e.g., favoring factual accuracy for acute high-risk diseases) (Chen et al., 25 Aug 2025, Zhao et al., 24 Oct 2025).
• Concept bottlenecking and interpretability-by-design: Including tokenized, interpretable features (from LLM-generated knowledge banks or clinical lexicons) enables explanatory rationales, error tracing, and improved generalization to unseen conditions (Mehta et al., 4 Mar 2025, Xu et al., 12 Apr 2025).

Disease-informed prompting, as a general engineering and theoretical discipline, leverages explicit domain knowledge, multi-modal reasoning, and adaptively optimized prompt design to operationalize clinically rigorous, interpretable, and high-performing AI decision systems across contemporary medical applications. This paradigm is being actively extended to encompass causal reasoning, continuous clinical progression modeling, and safety-critical conversational agents. Continued work emphasizes integration of evolving guidelines, domain ontologies, and specialist feedback into every layer of the prompt optimization stack.