Artificial intelligence-facilitated evidence-based medicine research and practice: A position statement
Year: 2026
Type: article
Abstract: The world healthcare system has deeply adopted evidence-based medicine (EBM) for clinical practice and policy making by systematically integrating all available knowledge, including research studies and adjusting their biases. The approaches for EBM have been evolving rapidly in the past decades, including novel statistical methods, certainty rating approaches, and sustainable governance of the evidence ecosystem.[1] However, it still faces challenges such as exponential accumulation of novel evidence. The introduction of artificial intelligence (AI) and informative techniques is now facilitating the maintenance of the living evidence ecosystem with quickly iterated approaches from machine learning (ML) to large language model (LLM). These constructive technologies from different areas are revolutionizing the process of EBM research as well as its association with daily practice.[2] The rapid evolution of AI techniques, however, introduces uncertainty and confusion for clinicians and researchers.[3] Strategic consideration from stakeholders of evidence governance calls for long-term foreseeing of technical directions and pipelines in the future, which remains absent. Living in their cocoons, researchers in different disciplines can hardly find the full picture of the evolution of the evidence ecosystem. To address these gaps and provide a clear road map for the trajectory of AI-facilitated EBM in the past, present, and future, a multidisciplinary team in China proposed a framework to assess the degree of AI involvement in the evidence ecosystem. This framework evaluates the current extent of AI adoption in EBM research and clinical practice and offers insights to guide the development of AI technologies in the strategic layer, including generative artificial intelligence (GAI), and industry standards, ultimately enhancing the process of EBM and improving patient outcomes. Development of the Framework. To illustrate a long-term development of the AI-facilitated EBM research and practice in the past, present, and future, the AI4Evidence Working Group of the Chinese Medical Doctor Association initiated the framework. Through six rounds of the virtual and in-person meeting, the core team raised a draft of the taxonomy. A multidisciplinary panel from a full working group, including clinicians, EBM methodologists, epidemiologists, and computer scientists, reviewed, revised, and approved the taxonomy and the manuscript [Figure 1].Figure 1: Structure of tiers assessing the degree of AI involvement in the evidence ecosystem. AI: Artificial intelligence; EBM: Evidence-based medicine; NA: Not available.TIER 0: EBM without AI. This tier represents the era of EBM research and practice before the involvement of any AI technologies. Human forces are the only investigators to operate and maintain the evidence ecosystem. TIER 1: EBM with complementary AI. In Tier 1, human forces take charge of the evidence synthesis, interpretation, and dissemination. AI participates in the evidence ecosystem merely as complementary assistance and support. The limited AI application merely offers light aids to human forces in several phases of data preparation and/or paper writing, such as paper search and screening, data extraction, and cleaning in meta-analysis, rather than core work, including efficient evidence dissemination, individual treatment, and unified data collaboration. AI is unable to be responsible for the quality of work and all details of AI-facilitated work are reviewed by experienced human reviewers. TIER 2: EBM with supportive AI. In Tier 2, AI technology initiates to take limited single tasks independently to further support the human forces in maintaining the evidence ecosystem. For example, AI may filter particular study types (such as randomized controlled trials), conduct core data analysis, and present results in an appropriate format during the evidence synthesis to reduce human workload. AI may also offer suggestions for appropriate statistical methods, evidence dissemination strategies, and generate data for evidence assessment across local regions. Nevertheless, all these procedures are executed by AI tools separately, and human forces remain in charge for all vital decisions. TIER 3: EBM with human supervised AI. In Tier 3, early-stage AI-facilitated automation starts to work. Compared with previous tiers, AI manages major tactical decisions in EBM research, including evidence collection, assessment, and possibly interpretation, and makes some adjustments supervised by human researchers from study design, research conduction to reporting. In practice, both health providers and patients may receive support with AI-generated information for individualized decision-making. Nevertheless, human researchers and health providers still take the full responsibility of the validity and trustworthiness of AI-informed material due to expected unqualified results produced by AI forces. TIER 4: semi-automated EBM. In Tier 4, AI intelligence synthesizes individualized evidence and supports timely for clinical decisions, which means general guidelines might fade out and link EBM research and practice seamlessly with an advanced automation process. Due to global data collaboration, AI-provided evidence, in most cases, is trustworthy and applicable in real-world practice, and people without qualified EBM training can engage without barriers. The gap between human researchers and health providers is eliminated and their role is to comment and approve the work provided by the AI using their clinical intuition and gestalt consideration for real life. TIER 5: automated EBM. In Tier 5, AI-facilitated EBM reaches the highest level, providing healthcare service using EBM principles directly to patients with seamless and real-time integration of research and practice. The role of AI may involve in the patient assessment, decision making, subsequent monitoring, and original research, filling the gap of the evidence. A full evidence ecosystem runs purely with AI, with humans informing their individual-level values and a relevant regulatory framework ensuring verifiably open accountability mechanisms. With the AI of this tier, the medical research and practice may evolve to a very different era that is beyond people’s imagination. This position statement proposes a framework to illustrate the past, present, and future of AI integration in EBM research and clinical practice. From Tier 0 to 5, the role of AI shifts from assisting particular work to operating an evidence ecosystem automatically. Although the AI involvement in EBM progressed gradually without a hard cut-off from Tier 0 to Tier 1, the launch of LLM such as ChatGPT in 2023 represented a landscape of revolution. In the past two years, the rapid development of AI-facilitated tools for EBM research and practice indicates the shift from Tier 1 to Tier 2 has been happening in recent years.[4] In the future, we are expecting more powerful AI tools in accelerating the knowledge transition from research to practice, and finally, at Tier 5, the sustainable and self-evolving evidence ecosystem will merge the research and practice with only voluntary input from the human force. In the end game of AI-automated clinical practice, a person may receive the healthcare services by automated detection of potential impairment, automatic synthesis of best evidence, and timely implementation. The recipient may not feel the ‘treatment’ but can check it as well as the full rationales of the ‘prescription’ or the care they received. Meanwhile, accountability frameworks will be established based on the characteristics of AI-based EBM, ensuring patient safety through operationalization standards and regulatory guardrails. In a long-term perspective, the implementation of AI technology in EBM research and practice is irreversible. At the current exercise of AI in EBM research and practice, the academic community has to balance clinical reasoning and AI-facilitated information. Gestalt consideration in the clinical perspective remains the demonstrating determinant for the certainty rating and clinical decision making compared with statistical parameters and outputs from “black box” algorithms.[4] For AI accelerated EBM materials, the full process should be transparent and checkable by human forces. It is important to minimize the AI-related errors and attribute their responsibilities. For material generated by AI-associated algorithms, the transparency at the reporting level and validation at the research level should be reported for health providers and downstream researchers to provide full information to make the judgment in being confident or not confident with the results.[5] For LLM-generated materials, lack of the trace of the information source as well as the certainty rating system is the main gap preventing their implementation in clinical practice. Representing a Chinese nationwide working group, this position statement proposed a high-level framework to assess the development of AI-facilitated EBM research and practice in six tiers at a historical view. The current AI tools and systems represent that we are at the transition from Tier 1 to Tier 2. Following the framework, the EBM research practice is expecting its accelerated development in the future decades, in parallel with the development of AI technology and its specialized EBM GAI. Expert Committee Jiahui Ma (West China Hospital of Sichuan University); Yaolong Chen (Lanzhou University); Zhenggang Bai (Evidence based Research Center of Nanjing University of TCM); Yifei Chen (Xinjiang Medical University); Da Feng (Huazhong University of Science and Technology); Le Gao (Xi’an Jiaotong University); Long Ge (Lanzhou University); Nana Guo (Hebei Center for Disease Control and Prevention); Yinghui Jin (Zhongnan Hospital of Wuhan University); Lun Li (Central South University Xiangya Second Hospital); Hailun Liang (Renmin University of China); Xing Liao (China Academy of Chinese Medical Sciences); Jun Lyu (The First Affiliated Hospital of Jinan University); Yanan Ma (China Medical University); Xiaolu Nie (Beijing Children’s Hospital); Xiaochun Qiu (Shanghai Jiao Tong University); Yang Song (The Chinese University of Hong Kong); Ying Sun (Fudan University); Buzhou Tang (Harbin Institute of Technology (Shenzhen); Weiwei Wang (Beijing Anding Hospital); Xinyin Wu (Central South University); Shanshan Wu (Capital Medical University); Jun Xia (University of Nottingham Ningbo China); Yizhong Yan (Shihezi University); Fengwen Yang (Tianjin University of Traditional Chinese Medicine); Xiaorong Yang (Shandong University); Zhirong Yang (Chinese Academy of sciences); Houyu Zhao (Peking University Third Hospital); Hao Zhou (Children’s Hospital of Fudan University); Boheng Zhang (Zhongshan Hospital Affiliated to Fudan University); Simin Zhu (The Chinese University of Hong Kong); Feng Sun (Department of Epidemiology and Biostatistics, School of Public Health, Peking University); Sheyu Li (Department of Endocrinology and Metabolism, MAGIC China Centre, Chinese Evidence-based Medicine Centre, West China Hospital of Sichuan University). Funding This work was supported by grants from the 1·3·5 project for disciplines of excellence, West China Hospital, Sichuan University (No. ZYYC24001). Conflicts of interest None.
Source: Chinese Medical Journal
Institution: Peking University
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,000
OpenAlex ID: W7126021718