Does the future of evidence-based healthcare lie in autonomous research?
Introduction
The healthcare landscape is undergoing a profound transformation with the integration of artificial intelligence (AI) meaning that health technology assessment (HTA) stands at a critical crossroads. While a previous discussion examined how generative AI is reshaping traditional HTA processes, a more sophisticated and promising development is emerging: agentic deep research. This approach represents a fundamental reimagining of how evidence is generated, synthesised, and applied to inform healthcare decision-making.
What is Agentic Deep Research?
Agentic deep research involves using AI "agents" that autonomously conduct in-depth research tasks, iteratively searching and analysing information to produce comprehensive insights. Unlike a simple query-response chatbot or a manual literature search, an agentic research system can chain together multiple steps – querying databases, reading and summarising documents, extracting data, and even performing basic reasoning or calculations – with minimal human guidance. The goal is to achieve a depth and breadth of analysis comparable to that of a skilled human researcher, but at a much greater speed and scale.
Notably, leading technology organisations have demonstrated that such AI agents can synthesise hundreds of hours of online sources and distil them into reports at the level of a research analyst. In practical terms, this means a HTA team could delegate labour-intensive research tasks (such as combing through thousands of journal articles or scanning real-world data repositories) to an AI agent and receive an organised, digestible summary of findings.
How Agentic AI Differs from Traditional Generative AI
Agentic AI represents a significant evolution beyond standard generative AI tools. Whilst both use large language models (LLMs) as their foundation, agentic systems possess several distinctive characteristics:
Autonomy: Agentic AI systems can independently analyse data, assess situations, update research plans, execute actions, and make rational decisions that optimise performance toward their goals, reducing the need for constant human oversight. This capability allows them to function effectively in complex and open-ended environments.
Goal-Oriented Behaviour: Agents are designed to pursue specific objectives. They accomplish this by breaking down complex tasks into manageable sub-tasks, planning workflows, prioritising actions, and dynamically adjusting processes based on the latest information to achieve the desired outcome efficiently.
Memory and Learning: Agentic systems possess memory, allowing them to retain information from past interactions and data. Importantly, they learn from this experience, often using techniques such as reinforcement learning or processing feedback from human operators, which enables them to refine their strategies and improve performance over time.
Environmental Adaptation: These systems are not static; they can perceive changes in their environment, whether from new data, user interactions, or external factors, and adjust their strategies in real-time to maintain effectiveness. This adaptability is crucial in dynamic fields such as healthcare research.
Tool Use: Agentic AI does not operate in isolation. It is designed to leverage available resources and tools. This can include accessing databases, searching websites, running software programmes, interacting with other AI models and utilising specialised algorithms to perform specific parts of its task. LLMs are particularly important, as they provide capabilities in understanding natural language, reasoning through problems, processing text, and retrieving domain-specific knowledge.
Leading Deep Research Platforms
Several leading AI providers have established strong research capabilities, each possessing unique strengths pertinent to HTA applications:
OpenAI's o3 Deep Research: OpenAI's o3 reasoning model features advanced dedicated deep research capabilities. This system can autonomously search through multiple sources, evaluate the quality of information, and synthesise findings into coherent analyses. It demonstrates a particular strength in understanding complex medical and scientific literature, making it potentially valuable for evidence synthesis.
Perplexity Pro Deep Research: Perplexity AI's Pro offering features enhanced deep research functionality specifically designed for information retrieval and synthesis. The platform integrates directly with academic databases and provides detailed citations, addressing a critical need in HTA for transparent, verifiable evidence trails.
Google's Gemini Pro 2.5 Deep Research: Google's Gemini platform leverages the company's extensive search and knowledge graph capabilities to provide comprehensive research features. The system excels at understanding the relationships between various pieces of information and can identify patterns across diverse data sources. Its integration with Google's academic search tools offers specific advantages for automating literature reviews.
These platforms represent significant advancements in autonomous research capabilities; however, each has different strengths and limitations for specific HTA applications. The best choice of platform depends on the particular HTA task and organisational needs.
Transformative Use Cases for Agentic Deep Research in HTA
The collaboration between autonomy and learning is particularly relevant for the deep research required in HTA. It suggests that an agentic system could go beyond executing predefined search protocols. It might autonomously identify promising new data sources, refine its understanding of the complex HTA question based on initial findings, and dynamically adjust its search or analysis strategy mid-task. This mirrors the iterative and exploratory nature of research conducted by human experts, potentially uncovering insights that rigid, non-adaptive automated processes might miss.
Let us explore specific use cases where agentic deep research could transform HTA:
1) Autonomous & Adaptive Evidence Synthesis
Systematic Literature Reviews (SLRs) are fundamental to HTA, providing a comprehensive synthesis of evidence on clinical efficacy, safety, health-related quality of life and other domains. Current AI and machine learning tools already offer significant assistance, primarily in the time-consuming task of screening titles and abstracts.
An agentic AI system could significantly elevate the review process. Based on the HTA research question, framed using criteria such as Population, Intervention/Comparison, Outcome, and Study type (PICOS), the agent can autonomously design the initial search strategy and execute comprehensive searches across multiple relevant databases.
The agent could perform initial abstract screening using LLMs or other classifiers. Crucially, leveraging its learning capabilities, it could dynamically refine the PICOS criteria based on real-time feedback collected from human reviewers regarding inclusion/exclusion disagreements and their rationales. This adaptive screening process promises greater efficiency and potentially higher relevance capture than static protocols.
2) Continuous Real-World Evidence Monitoring & Analysis
Real-world evidence (RWE) is becoming increasingly crucial in HTA, offering insights into long-term effectiveness, safety in routine practice, and outcomes in diverse patient populations. AI methods are already being explored to process large, unstructured RWE datasets and identify relevant patient cohorts or data points.
An agentic AI system could offer continuous, proactive RWE intelligence:
Autonomous Monitoring: An agent could be tasked with continuously monitoring a range of predefined RWE sources (e.g., electronic health records, claims databases, disease registries, and potentially even anonymised patient forums or wearable data streams, subject to strict governance and privacy controls).
Signal Detection & Trend Analysis: The agent could autonomously identify emerging safety signals, track effectiveness trends in specific subpopulations post-launch, and detect variations in treatment patterns.
Data Integration and Harmonisation: Leveraging its tool-use capabilities, the agent could integrate data from disparate RWE sources, potentially identifying ways to harmonise different data structures or account for variations in data quality.
3) Intelligent Health Economic Model Development
Health economic models are crucial HTA tools for assessing cost-effectiveness and value for money.
An agentic system could act as an intelligent assistant for model development:
Model Structure Proposal: Based on the synthesised clinical evidence (potentially provided by an SLR agent), the agent could analyse the disease pathway and treatment comparators to propose suitable economic model structures (e.g., Markov cohort model, discrete event simulation).
Parameter Identification and Retrieval: The agent can systematically scan the evidence base (SLRs, RWE analyses, other published literature) to identify and retrieve necessary model parameters, such as transition probabilities, resource utilisation, and costs associated with health states or treatments, as well as utility values.
Gap Analysis and Consistency Checks: The agent can identify missing parameters required for the chosen model structure, highlight inconsistencies in reported values across multiple studies, or pinpoint parameters that require expert elicitation.
The Path Forward: Balancing Innovation with Responsibility
That said, embracing agentic deep research in HTA must be approached thoughtfully. It is crucial to maintain transparency, ensure the reproducibility of AI-generated findings, and mitigate risks such as algorithmic bias or errors. Stakeholders will understandably ask: How did the AI derive this insight? Can we trust its recommendation?
In conclusion, agentic deep research has the potential to support more timely, comprehensive, and inclusive HTA processes worldwide, helping health systems make better-informed decisions about which technologies to adopt. It is a conceptual toolset that any HTA or payer organisation can adapt – not tied to a specific vendor or software, but rather a method of working smarter with AI as a partner. By remaining grounded in rigorous methodology and ethical oversight, HTA practitioners can harness these AI agents to navigate the ever-expanding sea of evidence. The outcome could be mutually beneficial: faster access to practical innovations for patients, more efficient use of resources for payers, and greater clarity and consistency in decision-making across all geographies.
Engage with Decisive Consulting
Are you exploring how AI can optimise your HTA strategy? Reach out to Lance Richard at lance.richard@decisiveconsulting.co.uk for expert insights on harnessing agentic AI while ensuring regulatory compliance and ethical best practices.
Written by Lance Richard
Decisive Edge 9th June 2025