Data Characteristics
Home healthcare clinical trial prescreening data originates from voluntary health questionnaires, device usage records, and physiological metrics from wearable devices. Data update frequencies vary. Questionnaires are typically single-entry or periodic reviews. Device usage records may update daily or per use. Physiological data may transmit in real-time or near real-time. Document structures include structured or semi-structured text for questionnaires, covering fields like age, gender, medical history, and medication. Device usage records may include operation time, mode, and results. Physiological data is typically numerical, such as blood pressure (mmHg), blood glucose (mmol/L), and heart rate (bpm). Data often contains medical abbreviations and specific device model information.
Constraints from Data Characteristics on Multiturn Conversation and Prompts
The diversity and varied update frequency of home healthcare data require robust context management in multiturn conversations. Structured information from questionnaires needs precise extraction to avoid misinterpretation. The discrete nature of device usage records requires the dialogue system to integrate data from different time points for comprehensive assessment. The numerical nature of physiological metrics means prompt design must include numerical range validation and unit conversion logic. For example, if a user mentions "blood pressure is a bit high," the system needs to ask for specific values and evaluate them against normal ranges. Users may be unfamiliar with medical terminology, so prompts must translate professional terms into common language and handle typos or vague descriptions in user input. Sensitive health information in the data also requires prompt design to prioritize privacy protection, avoiding leading questions.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8 turns | Balances user memory and system resources, covers common prescreening flows. |
temperature | 0.3–0.5 | Ensures accuracy and consistency of responses, prevents excessive divergence. |
Chunk size (Segment Length) | 500 characters | Accommodates text length of questionnaires and device records, reduces semantic loss. |
Recall count (Recall Count) | Top 5 entries (Top 5 items) | Balances recall efficiency and relevance, ensures critical information retrieval. |
Similarity threshold (Similarity Threshold) | 0.75 | Improves matching precision, filters for knowledge highly relevant to user intent. |
Rerank result count (Rerank Return Count) | 3 items | Further refines results, provides the most relevant and concise suggestions. |
Common Pitfalls
- "Invalid token" or "insufficient permissions" errors occur mid-conversation. This is due to expired model call credentials or misconfigurations, leading to API authentication failure.
- The system repeatedly asks for medical history information already provided by the user. This indicates improper multiturn conversation context management, with
maxContextset too short or historical dialogue state not correctly passed. - After a user inputs "blood pressure 140/90," the system fails to correctly recognize and provide risk alerts. This is because the prompt lacks logic for parsing blood pressure values or unit conversion rules.
Verification
- Simulate various user profiles and conduct full prescreening process tests. Check if key information (e.g., age, underlying conditions) is correctly captured and used for assessment.
- Intentionally input vague or ambiguous health descriptions. Observe if the system clarifies through follow-up questions and check the accuracy and guidance of these questions.
- Examine system feedback when handling abnormal inputs (e.g., non-numeric physiological indicators, invalid device models). Confirm if friendly error messages or guidance are provided.
Note: The values provided are common starting points. Measure them against specific 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.