Introduction
General-purpose search engines often fail to meet the specific needs of users seeking medical information. This failure stems from two primary issues: the "vocabulary gap," where lay users lack the technical terminology to find precise results, and a "structural mismatch," where the traditional flat list of search results does not reflect the inherently structured nature of medical knowledge. Presented at AAAI 2008, iMed is an intelligent medical search engine designed to bridge these gaps by guiding users through a diagnostic process and organizing results into a meaningful hierarchy.
Methodology: Interactive Diagnosis and Query Formation
To address the vocabulary gap, iMed employs diagnostic decision trees based on professional medical knowledge. Instead of requiring a precise query, the system initiates an interactive questionnaire. For instance, if a user enters "chest pain," the system navigates a decision tree by asking clarifying questions to narrow down potential conditions.
Once potential conditions are identified, iMed performs automatic query formation. It generates multiple weighted queries for various diseases and their associated clinical aspects, leveraging the MeSH (Medical Subject Headings) ontology to suggest relevant medical phrases and refine the search.
The Hierarchical Output Interface
The core innovation of iMed is its three-level hierarchical interface, which replaces the standard sequential list of links. This structure mirrors the clinical reasoning of medical professionals:
- Level 1 (Topic): Results are organized by disease names (e.g., dissecting aneurysm).
- Level 2 (Aspect): Within each disease, information is categorized into clinical aspects such as symptoms, diagnosis, treatment, and prevention.
- Level 3 (Web Pages): Specific documents are listed under their respective aspects.
By providing immediate access to these categories, the system reduces the cognitive load on users. The interface also includes "overview" snippets at the Topic and Aspect levels, allowing users to quickly rule out irrelevant conditions without navigating away from the search results.
Evaluation and Results
The system's search index was built from 22GB of data crawled from reputable sources like WebMD, Healthline, and the Merck Manual. Evaluation was conducted using 30 standardized cases from the USMLE Step 2 CS exam, comparing iMed’s hierarchical interface against traditional sequential search results. The results demonstrated significant improvements across several metrics:
- Diagnostic Accuracy: The success rate in identifying the correct diagnosis increased from 29% to 35%.
- Efficiency: The average time required to complete a search task dropped from 39 minutes to 30 minutes.
- User Effort: Users viewed fewer pages (11.3 vs. 14.7) and required fewer search iterations (2.9 vs. 3.7).
- User Satisfaction: Subjective ratings for ease of use and usefulness were significantly higher for the iMed interface.
Conclusion and Future Directions
iMed demonstrates that leveraging the "closed domain" nature of medicine allows for more proactive and structured information delivery. While currently focused on diseases, future work aims to expand the topic scope to include medical procedures and exams. Additionally, the authors plan to refine the aspect sets to be more context-sensitive, ensuring that only relevant clinical categories (e.g., excluding "surgery" where inapplicable) are presented to the user.
Summary updated