Multiturn Conversation and Prompts for Telemedicine Clinical

Data Characteristics

Telemedicine clinical trial pre-screening data originates from Electronic Health Records (EHR), Patient-Reported Outcomes (PROs), wearable devices, and remote consultation records. This data typically exists as unstructured text, semi-structured tables, and structured numerical values. Data updates frequently. Some physiological indicators update in real-time. Consultation records and PROs data update on an event-triggered basis. EHR data updates in batches according to the archiving cycles of medical institutions. Document structures vary, including free-text physician diagnostic reports, structured lab and examination results, and remote consultation texts containing patient chief complaints and medication history. Fields and units are highly specialized. For example, "serum creatinine (SCr)" units are typically mg/dL or umol/L. "New York Heart Association Functional Classification (NYHA Class)" uses Roman numerals I-IV. Identifying disease codes (e.g., ICD-10) and drug names (e.g., ATC classification) in text is crucial.

Constraints on Multiturn Conversation and Prompts

The diversity of data sources in telemedicine requires multiturn dialogue systems to effectively integrate information from different formats. High update frequency challenges the real-time nature of the knowledge base, necessitating more frequent index update strategies. The combination of unstructured text and specialized fields means prompt design must balance natural language understanding with professional terminology parsing. For example, patients may describe symptoms colloquially in a multiturn conversation, but the system must map these to precise medical concepts or ICD-10 codes. The presence of specialized units requires unit normalization or conversion when extracting numerical values to avoid misjudgments due to inconsistent units. Multiturn conversation context management must be rigorous to ensure that in complex medical Q&A, the system can track patient condition progression and accurately reference critical medical information from historical conversations, such as previously mentioned allergy history or past medical history. This constrains the system's maxContext setting and the accuracy of entity recognition.

Configuration Guidelines

Configuration Item	Suggested Value	Rationale
`Chunk Size`	`500–800 characters`	Balances the completeness of professional text with RAG recall efficiency, avoiding truncation of critical medical information.
`Recall Count`	`Top 8–12`	Ensures coverage of potentially relevant information from multiple data sources, balancing recall precision and computational overhead.
`Similarity Threshold`	`0.75–0.85`	Reduces the risk of false recall of irrelevant medical documents, improving the accuracy of pre-screening results.
`Rerank Count`	`Top 3–5`	Further refines RAG results, focusing on core information most relevant to clinical trial pre-screening.
`maxContext`	`3000–4000 tokens`	Maintains contextual coherence in multiturn medical conversations, tracking patient history and symptom changes.
`QUERY_REWRITE_ENABLE`	`true`	Optimizes colloquial or vague patient queries, converting them into more precise medical queries.

Common Pitfalls

Symptom: The model fails to correctly identify drug dosage units mentioned by the patient in the conversation, leading to incorrect pre-screening condition judgments. Reason: The prompt does not explicitly instruct the model to focus on and parse units after numerical values, or the knowledge base has not standardized units.
Symptom: The model "forgets" critical medical history mentioned by the patient early in a multiturn conversation, leading to inappropriate clinical trial recommendations. Reason: The context window maxContext is set too small, or the context management strategy fails to effectively retain important entity information.
Symptom: When extracting patient symptoms, the system incorrectly maps non-medical descriptions to ICD-10 codes, causing deviations in pre-screening results. Reason: The prompt's instructions for entity recognition in unstructured text are not precise enough, or there is a lack of domain dictionary assistance for recognition.

Verification Steps

Conduct multiple simulated patient consultations to observe if the model can correctly identify and reference critical medical information (e.g., medical history, allergies) from the conversation history.
Check the pre-screening result details to confirm that the knowledge base documents cited by the model are highly relevant to the patient's query and do not contain content inconsistent with medical common sense.
Test via API calls to verify the accurate extraction of specific medical terms (e.g., NYHA Class III) or units (e.g., mg/dL).
Compare pre-screening results from different Similarity Threshold and Rerank Count configurations to determine if the thresholds effectively filter noise while retaining highly relevant results.

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.