ARTIFICIAL INTELLIGENCE IN PAIN MEDICINE: OPPORTUNITIES, LIMITATIONS, AND LIABILITY A SYSTEMATIC REVIEW OF THE LITERATURE

G. Del Balzo, A. Malva, A. Martini | University of Verona, Italy

Background and Objective. Artificial intelligence (AI) is increasingly being integrated across the entire pain care pathway; however, its implementation raises important concerns regarding clinical reliability and professional liability. The aim of this systematic review was to provide a structured overview of current AI applications in pain medicine, highlighting both their potential and limitations, and to examine these findings within the legal framework governing physicians' civil liability when using AI-assisted systems.
Methods. A systematic literature search was conducted in PubMed, covering publications from January 2021 to June 2026. Multiple Boolean search strategies were used, including: "artificial intelligence" AND "pain medicine" (5,078 records); "machine learning" AND "chronic pain" AND prediction (464); "deep learning" AND opioid AND prescribing AND risk (10); "artificial intelligence" AND "pain assessment" AND "facial expression" (36); and "artificial intelligence" AND "pain management" AND review AND challenges AND ethics (62). Eligibility criteria included English-language reviews and primary studies involving human subjects, with preference given to original studies published between 2023 and 2025. Following title/abstract screening, duplicate removal, and exclusion of irrelevant records (non-pain dentistry and general nursing), 14 studies were included, comprising narrative reviews, systematic reviews/meta-analyses, and studies based on electronic health records (EHRs).
Results. AI applications demonstrated significant potential in several areas, including automated pain assessment through facial expressions, biosignals, and natural language processing in non-communicative patients; prediction of postoperative pain, pain chronification, and opioid use disorder (AUROC up to 0.94); AI-assisted ultrasound guidance for regional anesthesia; patient selection and programming for neuromodulation; and personalized treatment strategies for cancer pain and migraine. However, the available evidence remains limited by concerns regarding data quality and representativeness, algorithmic opacity ("black-box" models), bias, and insufficient external validation. Several reviews reported a high risk of bias in predictive models (e.g., migraine meta-analysis showing an AUC of 0.86 but a high overall risk of bias according to PROBAST), as well as risks related to clinician over-reliance and deskilling. From a legal perspective, in the Italian healthcare system AI-assisted pain management remains governed by the Gelli-Bianco Law (Law No. 24/2017) and general civil liability principles, integrated with the European AI Act (Regulation (EU) 2024/1689) and Italian Law No. 132/2025, which explicitly states that AI serves exclusively as a decision-support tool, while clinical decision-making remains the physician's responsibility. Under Article 26 of the AI Act, the physician acting as the AI deployer is liable under negligence-based principles rather than strict liability. Outputs generated by knowledge-based systems may be considered analogous to clinical practice guidelines under Article 5 of the Gelli-Bianco Law, whereas extending such equivalence to machine learning systems would require specific legislative provisions.
Conclusions. Current evidence suggests that AI holds considerable promise in pain medicine, although existing studies remain methodologically heterogeneous and frequently affected by a high risk of bias. Clinically safe and legally sustainable implementation of AI requires rigorous validation, adequate physician training in AI literacy, explainability of AI systems, and a clearly defined framework for professional responsibility centered on the physician's critical supervision.