Clinicians Using MedSearch AI Conversational Search Answer Questions 79% Faster, 34% Fewer Searches

Read the paper →

MedSearch: Accelerating Evidence-Based Medical Decision-Making with Real-Time AI Search – Preliminary Study Shows 79% Faster Responses and Reduced Search Burden

In the fast-paced world of clinical medicine, timely and reliable access to up-to-date evidence is critical for optimal patient care. Yet healthcare professionals often grapple with inefficient information retrieval methods, relying on traditional literature searches that can be time-consuming and fragmented. Addressing this gap, MedSearch introduces an AI-powered conversational agent designed to perform real-time, evidence-based internet searches tailored to medical queries. Preliminary findings reveal that clinicians using MedSearch answered clinical questions 79% faster and with 34% fewer searches compared to those using traditional methods, highlighting a promising leap forward in clinical information access.

This innovative tool integrates the strengths of large language models with live internet connectivity, allowing it to dynamically source and reference the latest validated medical literature. Such adaptability offers a powerful complement to clinician workflow, potentially transforming how medical knowledge is accessed at the point of care. In this blog post, we explore the study evaluating MedSearch’s performance among healthcare professionals, its methodological rigor, and the implications for clinical practice and future deployment.

Study Partnership & Context

The study was conducted by researchers at Arkangel AI in Bogotá, Colombia, motivated by a recognized need to enhance real-time access to validated medical knowledge within clinical and academic settings. The Colombian healthcare environment, marked by diverse patient populations and resource variability, provides a meaningful backdrop for testing digital tools aimed at improving clinical decision support. This collaboration reflects a growing international interest in leveraging AI to streamline evidence retrieval and integrate clinical knowledge more seamlessly into healthcare workflows.

The study specifically engaged a cohort of medical students, general practitioners, and specialists, reflecting a broad spectrum of clinical experience and information needs. By comparing MedSearch with conventional search techniques, the researchers sought to assess not only the speed and efficiency but also the user acceptability of this AI-assisted approach in real-world clinical reasoning scenarios.

Study Design and Methodology

The study employed a randomized design involving 25 preliminary participants divided into two groups:

• Group A (MedSearch users): 13 participants used the MedSearch conversational agent to answer clinical case questions.
• Group B (Traditional methods): 12 participants relied on conventional search strategies excluding any AI-based platforms.

Participants tackled four clinical case scenarios, each accompanied by four carefully designed questions spanning general diagnosis, differential diagnosis, state-of-the-art research, and general medical knowledge. Key metrics recorded included the average time to answer each question and the number of searches performed per question. Additionally, Group A participants completed a brief survey evaluating the perceived utility, confidence in responses, likelihood of daily use, and willingness to recommend MedSearch to peers, using a simple 1-to-3 scoring scale.

MedSearch operates by combining advanced large language model techniques (LLMs) with live internet connectivity, enabling dynamic retrieval of up-to-date, evidence-backed information with direct access to source references. This contrasts with typical LLMs deployed offline or trained on static datasets, addressing a crucial limitation in medical AI tools.

Key Results

• Response Time: Median response time per question was 46 seconds for MedSearch users, versus 1.75 minutes for traditional searchers — a staggering 79.8% reduction in time.
• Number of Searches: MedSearch users performed on average 3.81 searches per question compared to 5.38 by the traditional group, representing a 34.1% decrease.
• User Acceptability Scores (1–3 scale):
  • Perceived Utility: 3.0 (helpful in answering questions)
  • Confidence in Accuracy: 2.7 (generally confident)
  • Likelihood of Daily Use: 2.8 (probable use)
  • Recommendation to Peers: 3.0 (would recommend)

These preliminary results underscore the efficiency gained by integrating an AI conversational search tool into clinical problem-solving without compromising user confidence or satisfaction.

Interpretation & Implications

The dramatic reduction in search time and query volume suggests that MedSearch can streamline clinical workflows by rapidly surfacing relevant, evidence-based answers. This capability may reduce cognitive load on providers, speed up decision-making, and ultimately support better patient outcomes through timely incorporation of current best evidence. The high acceptability and confidence ratings further indicate that clinicians find the tool trustworthy and practical, important factors for adoption and sustained use.

By contrast, traditional search methods—often requiring multiple platforms, textbooks, or databases—present logistical barriers to efficient information retrieval. MedSearch’s conversational interface offers a more intuitive and interactive way to navigate complex medical knowledge bases, lowering the barrier for evidence-based practice in busy clinical environments.

Nonetheless, given the preliminary nature of these findings, further validation with larger and more diverse cohorts is essential. Additionally, the study did not yet evaluate diagnostic accuracy or patient outcomes linked directly to MedSearch use, areas ripe for future investigation. Potential biases in question difficulty or participant familiarity with digital tools similarly merit careful examination.

Deployment & Scalability

MedSearch’s architecture enables seamless integration into clinical settings where rapid decision support is necessary—such as emergency departments, outpatient clinics, and academic hospitals. Its real-time internet connection differentiates it from closed LLM systems and allows for continuous updating of evidence with minimal manual intervention.

Challenges for widespread deployment include ensuring data privacy, securing reliable internet access, and providing user training to optimize effective use. Addressing these hurdles will be critical in settings with limited IT infrastructure. However, the flexibility of MedSearch’s design allows adaptability to multiple languages, specialties, and healthcare systems globally, suggesting broad potential beyond the initial Colombian context.

Moreover, the underlying technology could be extended to support allied health professions, patient education, and public health information dissemination, accelerating knowledge transfer across the medical ecosystem.

Conclusion & Next Steps

The MedSearch pilot study presents compelling early evidence that AI-enabled, real-time evidence retrieval can markedly enhance clinical information access. By significantly reducing search time and effort while maintaining user trust and perceived utility, MedSearch offers a promising avenue to embed up-to-date medical knowledge into routine practice.

Moving forward, larger-scale trials assessing clinical accuracy, decision impact, and patient outcomes will be crucial. Parallel efforts should focus on integrating MedSearch with electronic health records and refining its interface for diverse user groups. With continued development and validation, tools like MedSearch could become indispensable assets in the pursuit of evidence-based, efficient, and patient-centered care.

References

• Castaño-Villegas N, Villa MC, Llano I, Zea J. Validación preliminar de MedSearch: un agente conversacional para responder preguntas médicas en tiempo real basadas en evidencia. Arkangel AI, Bogotá, Colombia, 2024.
• Liu N, Chen L, Tian X, Zou W, Chen K, Cui M. From LLM to Conversational Agent: A Memory Enhanced Architecture with Fine-Tuning of Large Language Models. arXiv 2024. http://arxiv.org/abs/2401.02777
• Gilson A, Safranek CW, Huang T, Socrates V, Chi L, Taylor RA, Chartash D. How Does ChatGPT Perform on the United States Medical Licensing Examination (USMLE)? JMIR Med Educ. 2023;9:e45312.
• Thirunavukarasu AJ, Ting DSJ, Elangovan K, et al. Large language models in medicine. Nat Med. 2023;29:1930–1940.