MedSEBA: AI-Driven Medical Evidence Synthesis

• MedSEBA is an AI-powered platform that synthesizes trustworthy medical answers by retrieving and summarizing peer-reviewed PubMed research.
• It employs biomedical named entity recognition and cosine similarity ranking to extract top relevant studies, ensuring precise answer synthesis with traceable citations.
• The system features interactive visualizations that map research stances and temporal trends, enabling both lay users and experts to assess evolving scientific consensus.

MedSEBA is an interactive, AI-powered system designed to generate synthesized, evidence-based answers to medical questions by grounding responses in peer-reviewed research dynamically retrieved from PubMed. It combines LLMs with biomedical literature retrieval and answer composition strategies to overcome the challenges posed by information overload, unreliable online sources, and the rapidly evolving nature of medical research. By tracing its generated summaries directly to supporting or refuting studies, MedSEBA addresses the dual needs of lay users seeking trustworthy medical advice and researchers needing high-level insights into shifting scientific consensus.

1. Objectives and Functional Scope

MedSEBA is conceptualized to serve both non-experts and domain professionals by (a) synthesizing reliable medical answers from vetted literature, and (b) visualizing the stances and temporal evolution of scientific consensus on any queried medical claim. The system operates by transforming user questions into targeted PubMed searches, retrieving relevant studies, and generating answers that systematically reference those studies. Each generated answer is accompanied by traceable citations and visual evidence maps to continually ground its claims in the corpus of evolving medical literature.

2. Evidence Retrieval and Ranking Methodology

MedSEBA initiates its answer generation by transforming user queries into optimized Boolean search expressions via SciSpacy's biomedical named entity recognition. For example, a question about predictors of poor surgical outcomes in elderly cardiac surgery patients becomes:

predictors AND poor AND surgical outcomes AND elderly cardiac surgery AND (patients OR outpatients OR visitors to patients) AND ("journal article"[Publication Type] OR "review"[Publication Type])

This query is submitted to PubMed to fetch 50 candidate articles. These documents are embedded using a sentence transformer (BMRetriever), and ranked using cosine similarity, defined mathematically as:

$\cos(\theta) = \frac{A \cdot B}{\|A\| \|B\|}$

where $A$ and $B$ are vector representations of the query and document respectively. The top 20 most relevant articles are selected for answer synthesis.

3. Structured Answer Synthesis and Citation Integration

The system prompts a variant of GPT-4 (GPT-4o) with a list of abstracts and a detailed task-specific prompt to elicit a structured answer. The output is organized into clearly demarcated thematic sections, each summarizing a group of findings. Crucially, every key point is supplemented by inline, clickable citations that map directly to the source studies. For further transparency, MedSEBA identifies the single most germane sentence from each study in relation to the query, allowing users to rapidly validate the linkage between summary claims and primary evidence.

4. Research Stance Labeling and Consensus Visualization

A distinguishing feature of MedSEBA is its explicit mapping of research stances. For every cited study, GPT-4o assigns a label—"supported," "refuted," or "neutral"—with respect to the medical claim posed in the query. The system aggregates these stances into visualizations such as:

• Stacked bar charts reflecting the distribution of support/refutation/neutrality among retrieved studies.
• Time series plots showing how research stances evolve over publication years, and correlating these trends with citation counts as a proxy for impact.

This mechanism facilitates rapid, high-fidelity appraisal of the direction and volatility of scholarly consensus, helping users to recognize shifts, divergences, or consolidation in scientific viewpoints over time.

5. Usability Evaluation and Feedback

MedSEBA underwent empirical validation via a user study involving medical experts and graduate students. The system achieved an average System Usability Scale (SUS) score of 81.7, surpassing usability benchmarks. Study participants consistently described it as easy to navigate, informative, and trustworthy, highlighting the value of direct linkage to supporting studies. However, users identified challenges in exhaustively representing all relevant evidence due to output length constraints, and recognized a need for further precision in extracting the single most salient sentence per publication.

6. Applications and Prospective Enhancements

MedSEBA's framework is applicable to both everyday health advice, where lay users benefit from transparent, evidence-grounded answers, and to advanced research workflows, where survey or hypothesis forming demands precise synthesis of up-to-date literature and temporal consensus analysis. A plausible implication is that support for locally deployed LLMs could extend MedSEBA’s reach to privacy-sensitive, clinical applications where data confidentiality and rapid, contextual synthesis from non-centralized databases become paramount.

7. Significance and Limitations

MedSEBA presents an integrated pipeline uniting advanced neural retrieval, LLM-based synthesis, and interactive research visualization. The system directly addresses the contemporary challenges of medical advice reliability and literature overload in the digital age. Nevertheless, output completeness remains bounded by system constraints on response length and citation density, and stance assignment—though robust—reflects the inferential capacity and any inherent limitations of the underlying LLM (GPT-4o). Future developments may target broader document coverage, refined sentence extraction, and enhanced local deployability.

---

MedSEBA embodies a data-driven approach to trustworthy medical question answering, leveraging automated evidence retrieval, NLP-based synthesis, and visual consensus mapping for both public health and academic research contexts (Vladika et al., 30 Aug 2025).