Multi-turn Conversation and Prompts for Ophthalmic Clinical Trial Pre-screening

Ophthalmic clinical trial pre-screening data originates from hospital electronic medical record systems, reports from ophthalmic examination equipment

Data Characteristics

Ophthalmic clinical trial pre-screening data originates from hospital electronic medical record systems, reports from ophthalmic examination equipment (e.g., OCT, visual field meters, fundus cameras), and patient self-reported questionnaires. This data updates frequently; outpatient examination reports can generate new data weekly or even daily. Document structures vary, including unstructured progress notes, structured examination reports (e.g., visual acuity, intraocular pressure, corneal topography), and semi-structured follow-up records. Fields contain numerous specialized terms such as "macular edema," "optic nerve atrophy," and "intraocular pressure fluctuation range." Units include millimeters of mercury (mmHg), LogMAR for visual acuity, and diopters (D). The same field can appear in different forms across various reports.

Constraints on Multi-turn Conversations and Prompts

The diversity and complexity of ophthalmic data demand high accuracy in multi-turn conversations and prompt design. Specialized terms and abbreviations in unstructured medical records require prompts with strong semantic understanding to extract key information correctly. Numerical fields in structured examination reports, with their varying units and normal ranges, necessitate precise conversion and comparison by the dialogue system during questioning or explanation. High data update frequency means the knowledge base requires frequent synchronization to ensure real-time and effective pre-screening. Patient self-reported questionnaires, with their subjective descriptions, require prompts to guide users in providing objective, quantifiable information to reduce ambiguity. These factors require the ability to flexibly handle multiple data types in multi-turn conversations and provide intelligent guidance based on context.

Configuration Settings

Configuration ItemRecommended ValueRationale
maxContext800–1200 charactersOphthalmic medical records and reports typically contain extensive information, requiring a sufficiently long context window to maintain dialogue coherence and understanding.
chunkLength300 charactersEnsures each knowledge chunk contains enough information while avoiding excessive length that could lead to information redundancy or semantic drift.
recallCounttop 5Clinical pre-screening requires comprehensive consideration of patient information. Appropriately increasing the recall count helps cover more relevant knowledge.
similarityThreshold0.75The clinical domain demands high accuracy. A higher similarity threshold reduces interference from irrelevant information.
rerankCount3After reranking, focus on the few most relevant pieces of information to improve dialogue efficiency and accuracy.
queryRewriteenabledAddresses the differences between ophthalmic professional terms and patient colloquial descriptions, improving retrieval accuracy through rewriting.

Common Pitfalls

  • The dialogue contains excessive irrelevant information, leading to model "hallucinations." This occurs when the knowledge base recallCount is set too high or the similarityThreshold is too low, failing to effectively filter noise.
  • The model fails to understand patient symptom descriptions, for example, not associating "blurry vision" with "decreased visual acuity." This happens when prompt design does not sufficiently cover various expressions for common ophthalmic symptoms.
  • In multi-turn conversations, the model repeatedly asks for already provided information. This indicates an insufficient maxContext setting, leading to context loss and an inability to maintain dialogue coherence.

Verification Steps

  • Conduct multi-turn simulated dialogues to verify the model's ability to correctly understand ophthalmic professional terms and patient colloquial descriptions, and to provide relevant suggestions.
  • Review the knowledge base recall content in the dialogue details to confirm that the recalled document segments are highly relevant to the current dialogue context and contain no obvious incorrect information.
  • Test various complex ophthalmic cases to ensure the model maintains contextual coherence in long dialogues and does not repeatedly ask for already provided information.

The values provided are common starting points and should be measured against your own samples.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.