Publications

Purpose: ChatGPT and other large language models (LLMs) have gained immense popularity since their commercial release in 2022, with applications in various sectors including health care. We sought to evaluate their deployment in anesthesiology and critical care in a systematic review. Our aim was to describe the integration of LLMs in the field by showcasing and categorizing their current applications, assessing their performance in patient care, and reviewing application-specific ethical and practical challenges in deployment. Methods: Respecting Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines, we systematically searched through PubMed®, Embase, the Cochrane Central Register of Controlled Trials, and Web of Science®, from inception until 1 August 2024. We extracted all papers investigating LLMs in anesthesiology or critical care and reporting results. We segmented the literature into major themes and highlighted key findings and limitations. Results: From 480 retrieved articles, we included 45 papers. The evaluated models (GPT-4, GPT-3.5, Google Bard [now Gemini], LLaMA, and others) showed diverse applications in four segments: intensive care unit, patient education, medical education, and perioperative care. Large language models, especially newer models, are promising in predicting clinical scores, navigating simple clinical scenarios, and managing preoperative anxiety. Their performance remains below the clinician level in predicting outcomes, solving complex clinical scenarios (i.e., airway management), board examinations, and generating patient-directed documents, although newer models performed better than older ones. Conclusion: While LLMs are not yet equipped to fully assist physicians in anesthesiology and critical care, they have significant potential, and their capabilities are rapidly improving. Supervised use for select tasks can streamline patient care. Further trials are warranted as new versions of models become available.

No abstract available

Background: The transversus abdominis plane (TAP) block is a regional anaesthetic technique targeting thoracolumbar sensory nerve afferents to provide analgesia to the anterolateral abdominal wall. With emerging RCTs supporting its efficacy across various surgeries, an updated, comprehensive review is warranted to assess its analgesic effectiveness and opioid sparing potential. Methods: A systematic review and meta-analysis were conducted to assess the analgesic efficacy of the TAP block. We searched MEDLINE, Embase, and Cochrane CENTRAL from inception to May 2024. We included RCTs that compared the TAP block with placebo, local infiltration, epidural analgesia, or intrathecal morphine and reported 6-, 12-, or 24-h postoperative pain scores or cumulative 24-h postoperative morphine consumption. Meta-analyses were performed for all outcomes. Results: A total of 123 RCTs were included in this review. The TAP block group showed significant reductions across all outcomes when compared with placebo. When compared with local infiltration, the TAP block also showed significant reductions in 6-h (standardised mean difference [SMD]=-0.89; 95% confidence interval [95% CI], -1.35 to -0.43; P<0.001), 12-h (SMD=-1.27; 95% CI, -2.08 to -0.46; P=0.002), and 24-h (SMD=-0.66; 95% CI, -1.01 to -0.32; P<0.001) postoperative pain, and in cumulative 24-h postoperative morphine consumption (SMD=-0.99; 95% CI, -1.52 to -0.47; P<0.001). However, it only showed a reduction in 12-h pain scores when compared with epidural analgesia and did not differ significantly with intrathecal morphine for any of the outcomes. Conclusions: The TAP block is an effective anaesthetic modality for pain management and opioid sparing after abdominal surgery.

Background Artificial intelligence (AI) embedded in point-of-care ultrasound (POCUS) could reduce operator dependence in left ventricular ejection fraction (LVEF) assessment and accelerate bedside decision-making by non-cardiologists and non-radiologists. We systematically reviewed prospective comparative studies evaluating real-time AI-assisted POCUS for LVEF across point-of-care settings. Methods PubMed/MEDLINE, Embase, and Cochrane were searched from inception to June 11, 2025. We included prospective analytical observational studies in ICU, ED, perioperative, ward, or home settings that used AI during image acquisition/interpretation to estimate LVEF in real time against a reference standard. Risk of bias was appraised with QUADAS-2. Due to heterogeneity, findings were narratively synthesized. Results Twelve studies met inclusion criteria across ICU, emergency, perioperative, ward, and community settings. For continuous outcomes, agreement between AI-derived and reference LVEF ranged from r = 0.56 - 0.92 and ICC = 0.84 - 0.94. Bland–Altman analyses demonstrated broad limits of agreement, from approximately -31.8% to +20.0%, often indicating a tendency for AI to underestimate LVEF compared with reference measures. For categorical classification at a 50% threshold, sensitivities ranged 70 - 93%, specificities 89–100%, and AUROCs 0.85 - 0.98. Weighted kappa values ranged 0.49 - 0.83. No study achieved low risk of bias across all QUADAS-2 domains. Conclusions AI-assisted POCUS can approximate reference LVEF and reasonably classify reduced LVEF at the bedside, supporting its use as a screening/triage tool where comprehensive echocardiography is not immediately available. Future studies should adopt intention-to-diagnose designs, standardize Simpson’s biplane reference, minimize test delay, report calibration/failure rates, and incorporate image-quality feedback.

Background: Advances in artificial intelligence (AI) have enabled large language models (LLMs) to generate complex and contextually relevant medical responses. However, their potential in clinical decision support for anesthesiology remains underexplored. This study evaluated the accuracy and clinical relevance of high-performing LLMs in response to anesthesia-related questions and compared their performance with traditional online search methods. Clinician perceptions of AI were also assessed. We hypothesized that top-performing large language models would outperform lower-tier models and traditional internet search tools by generating responses rated as more accurate, complete, and clinically relevant to anesthesiology-focused questions, as measured by higher mean evaluator scores on a 10-point Likert scale. Methods: Ten LLMs: GPT-4o, Claude-Sonnet 3.5, DeepSeek R1, Llama 3.1 Instruct 70B, Gemini 2.0, GPT o1-preview, GPT o1, GPT o3-mini, NOVA Pro, and Mistral. All models were tested using ten common general anesthesia questions developed by TMH and validated by 6 physicians. Two Google search conditions served as baselines: a default search conducted in a cleared browser (unpersonalized), and a personalized Google Snippet Search performed in a browser regularly used by a clinician. Four board-certified anesthesiologists independently rated each response on a 10-point Likert scale. An ad hoc Physician Perception Questionnaire captured clinicians' use of AI, trust in its output, and reliance on traditional information sources. Results: LLM performance varied significantly (F = 5.89, P <.0001). DeepSeek R1 achieved the highest overall score (7.7), whereas Gemini 2.0 Flash recorded the lowest among LLMs (5.2). The Google Snippet Search scored 5.3, the lowest overall. Pairwise Welch's t tests showed that DeepSeek R1 significantly outperformed Llama, o3-mini, and Mistral (P <.001). Survey results indicated limited AI use in clinical practice; clinicians prioritized source credibility and continued to favor traditional resources. Conclusions: Although LLM-generated responses differed in quality, DeepSeek R1 and Claude-Sonnet 3.5 produced answers most consistent with expert clinical judgment. The poor performance of several models, coupled with clinician skepticism, underscores the need for further validation before integrating AI into routine anesthesiology decision support.

Background: Residual neuromuscular blockade (rNMB) after surgery can lead to complications such as respiratory distress, increased length of hospital stay, and higher healthcare costs. Sugammadex, a selective relaxant-binding agent, has shown efficacy in reversing rocuronium-induced neuromuscular blockade more rapidly than neostigmine. This study aimed to evaluate the effectiveness of a low dose of sugammadex (0.5 mg.kg-1) compared to standard neostigmine with glycopyrrolate for the reversal of moderate rocuronium-induced neuromuscular blockade. Methods: This randomized, double-blind, controlled trial included adult patients undergoing surgery with rocuronium-induced moderate neuromuscular blockade. Participants were randomized to receive either low-dose sugammadex or standard-dose neostigmine with glycopyrrolate at the end of surgery. The primary outcome was the time to achieve a Train-of-Four (TOF) ratio of ≥ 0.9 after administration of the reversal agent. Secondary outcomes included extubation time, incidence of sugammadex rescue therapy for incomplete reversal, and postoperative complications. Statistical analysis used t-tests and Mann-Whitney-Wilcoxon tests for continuous variables and chi-square tests for categorical variables, with significance set at p < 0.05. Results: The median time to reach TOF ratio of ≥ 0.9 was significantly shorter in the sugammadex group (4.3 min, IQR: 3.2-6) compared to the neostigmine group (20.6 min, IQR: 10.1-21.3, p < 0.001). Extubation times were also reduced with sugammadex, with a median of 11.6 (IQR: 8.6-15.1) minutes versus 25.9 (IQR: 17.7-29.7) minutes in the neostigmine group (p < 0.001). Only 4.8% of patients in the sugammadex group required rescue therapy to reverse the neuromuscular block with TOF ratio of ≥ 0.9, compared to 60.6% in the neostigmine group (p < 0.001). No significant differences were observed in postoperative respiratory complications or PACU length of stay. Conclusions: Low-dose sugammadex provides a faster and more reliable reversal of moderate neuromuscular blockade than standard-dose neostigmine, with implications for improved operating room efficiency and patient safety. Continuous neuromuscular monitoring remains essential, as a small proportion of patients may still require additional intervention.

Purpose: Accurate hemodynamic monitoring is essential for personalized care in surgical and intensive care unit (ICU) settings, and radial artery catheters are the gold standard. Nevertheless, central-to-peripheral arterial blood pressure gradients (CPAPGs) have been reported in up to 77% of patients, potentially leading to misinformed clinical decisions. Identifying patients at risk of CPAPGs is crucial to adapt monitoring strategies. We aimed to identify the risk factors contributing to CPAPGs and assess their level of consensus in a systematic review of the literature. Methods: For this systematic review, we searched MEDLINE/PubMed®, Embase, and Cochrane databases (CENTRAL and Database of Systematic Reviews) from inception to 15 July 2024. Studies on adult patients with simultaneous radial and central arterial pressure monitoring were included. We extracted and analyzed risk factors and assessed the consistency of the literature for each. Results: From 741 retrieved articles, we included 55 studies (N = 5,598) evaluating the presence of CPAPGs during cardiopulmonary bypass, hepatic transplant, ICU stay, and other settings. Overall, 90 unique potential risk factors were evaluated, with 36 showing a statistically significant association with CPAPGs. Complex procedures and smaller radial artery diameter were the most consistent predictors, with both showing significant association with CPAPGs in the 3 studies evaluating them. Vasopressor usage showed statistical significance in half the studies evaluating it, and demographic factors, potentially associated with smaller radial artery diameter, such as older age, shorter stature, and female sex, were also statistically significant risk factors in some studies. Conclusions: The present systematic review identified numerous risk factors for CPAPGs. In patients with multiple contributing factors, clinicians should consider using more central arterial cannulation to minimize the risk of inaccurate pressure measurements.

This study investigated whether artificial intelligence (AI)-based visual assistance in video laryngoscopy (VL) could be a solution to reduce the technique's learning curve. Twenty volunteers with no prior intubation experience were randomly assigned to 2 equal groups: standard VL or AI-enhanced VL (AI-VL). Participants performed 10 consecutive intubations. The AI-VL group showed a trend toward steeper learning curves for intubation time (P > .05). The AI system functioned as a real-time virtual instructor, which provided visual feedback that helped maintain Cormack-Lehane (C-L) views.

Background: In the UK, total intravenous anaesthesia (TIVA) is used in 25% of general anaesthetics and is gaining traction because of its lower environmental impact and effectiveness in reducing postoperative nausea and vomiting (PONV). Although meta-analyses have compared TIVA and inhalational anaesthesia (IA), the optimal delivery method-manual infusion or target-controlled infusion (TCI)-remains underexplored. This review addresses this gap, leveraging the rapidly growing body of evidence to guide optimal anaesthetic practice. Methods: We searched PubMed, Embase, Cochrane CENTRAL and Web of Science from inception to October 10, 2024. Studies comparing TIVA and IA across several patient-related and efficiency outcomes were included. Meta-analyses were performed for all outcomes. Subgroup analyses were performed to assess the contribution of different factors including a comparison of TCI with manual infusion in TIVA. Results: In total, 385 RCTs were included. No significant difference in ClassIntra grade 3-4 adverse events was observed between TIVA and IA (risk ratio [RR]: 1.00, 95% confidence interval [CI]: 0.88-1.12; P=0.97). Subgroup analysis also showed no significant difference for TCI (RR: 0.89, 95% CI: 0.66-1.21; P=0.46) or manual infusion (RR: 1.03, 95% CI: 0.90-1.17; P=0.70) groups. IA was favoured in recovery times and costs, whereas the incidence of PONV and agitation on emergence favoured TIVA. No statistical difference was observed in our other outcomes. Conclusions: TIVA and IA are comparably safe, with TIVA reducing PONV and agitation, whereas IA offers faster recovery and lower costs. The use of TCI in TIVA might decrease postoperative cognitive dysfunction and increase recovery time, highlighting the need for a systematic review directly comparing TCI and manual infusion.

Background: Pharmacological sedation and analgesia are used to alleviate discomfort during awake medical procedures but can cause adverse effects like apnea and hypoxemia, increasing the need for airway management and prolonging recovery. Virtual reality (VR) has emerged as a non-pharmacological intervention to reduce the need for procedural sedatives and analgesics. Methods: A systematic review and meta-analysis were conducted, assessing the impact of VR immersion on intraprocedural sedation and analgesia usage in adults (≥ 18 years). We searched MEDLINE (PubMed), Embase, Cochrane CENTRAL, and Web of Science from inception to August 1st, 2024. We included analytical studies utilizing VR immersion in the intervention arm, and reporting tailored dosages of intraprocedural sedatives (propofol, midazolam) and/or opioids. Statistical analyses used standardized mean differences (SMD), and heterogeneity was assessed with I2. Results: Of 2714 identified papers, 11 (560 patients) were included. VR significantly reduced propofol usage (SMD = -1.70; 95% CI -3.10 to -0.31; P = 0.02; I2 = 92%) and midazolam usage (SMD = -0.29; 95% CI -0.57 to 0.00; P = 0.05; I2 = 0%). However, our analysis showed no reduction in opioid usage (SMD = -0.21; 95% CI -0.60 to 0.19; P = 0.31; I2 = 74%) in the VR group. Conclusions: VR immersion effectively reduces the required dose of intraprocedural sedatives, but its impact on opioid consumption remains unclear, especially in the absence of regional or neuraxial anesthesia. Further research is needed to clarify these effects and optimize VR use in clinical practice.

Study objective: Evaluate the impact of deep neuromuscular blockade on intraoperative nociception Deep neuromuscular blockade has been shown to improve surgical conditions and postoperative outcomes compared to moderate neuromuscular blockade in laparoscopic surgery. Still, its impact on intraoperative nociception and opioid requirement has never been assessed. Design: Monocentric randomised controlled trial. Setting: Operating room. Patients: We included 100 ASA I to III patients who underwent colorectal laparoscopic surgery with desflurane-remifentanil anesthesia. Interventions: Patients were randomised into two groups to achieve either moderate (1-3 train of four response) or deep (1-2 post-tetanic count) neuromuscular block (NMB) with repeated boluses of rocuronium. The Nociception Level (NOL) index guided intraoperative remifentanil administration in both groups. Measurements: The primary endpoint was total intraoperative remifentanil administration per hour of surgery. Secondary endpoints included, Leiden Surgical Rating Scale (L-SRS), intra-abdominal pressure, postoperative pain scores and opioids' consumption. Main results: Ninety-three patients were analysed. Forty-five in the deep group and 48 patients in moderate group. Intraoperative administration of remifentanil was 348 (228-472) μg.h-1 in the deep NMB group compared to 494 (392-618) μg.h-1 in the moderate NMB group (P < 0.001). Lowest L-SRS was 5 (4-5) in the deep NMB group versus 3 (2-5) (P < 0.001) in the moderate NMB group. Mean intra-abdominal pressure was 11.9 (1.3) in the deep NMB group versus 13 (1.3) (P < 0.001) in the moderate NMB group. Secondary postoperative outcomes including pain scores and analgesics administration were not significantly different. Conclusions: This study shows that deep neuromuscular blockade reduces intraoperative NOL-guided administration of remifentanil in colorectal laparoscopic surgeries. It also improves surgical conditions.

Anesthesia objectives have evolved into combining hypnosis, amnesia, analgesia, paralysis, and suppression of the sympathetic autonomic nervous system. Technological improvements have led to new monitoring strategies, aimed at translating a qualitative physiological state into quantitative metrics, but the optimal strategies for depth of anesthesia (DoA) and analgesia monitoring continue to stimulate debate. Historically, DoA monitoring used patient’s movement as a surrogate of awareness. Pharmacokinetic models and metrics, including minimum alveolar concentration for inhaled anesthetics and target-controlled infusion models for intravenous anesthesia, provided further insights to clinicians, but electroencephalography and its derivatives (processed EEG; pEEG) offer the potential for personalization of anesthesia care. Current studies appear to affirm that pEEG monitoring decreases the quantity of anesthetics administered, diminishes postanesthesia care unit duration, and may reduce the occurrence of postoperative delirium (notwithstanding the difficulties of defining this condition). Major trials are underway to further elucidate the impact on postoperative cognitive dysfunction. In this manuscript, we discuss the Bispectral (BIS) index, Narcotrend monitor, Patient State Index, entropy-based monitoring, and Neurosense monitor, as well as middle latency evoked auditory potential, before exploring how these technologies could evolve in the upcoming years. In contrast to developments in pEEG monitors, nociception monitors remain by comparison underdeveloped and underutilized. Just as with anesthetic agents, excessive analgesia can lead to harmful side effects, whereas inadequate analgesia is associated with increased stress response, poorer hemodynamic conditions and coagulation, metabolic, and immune system dysregulation. Broadly, 3 distinct monitoring strategies have emerged: motor reflex, central nervous system, and autonomic nervous system monitoring. Generally, nociceptive monitors outperform basic clinical vital sign monitoring in reducing perioperative opioid use. This manuscript describes pupillometry, surgical pleth index, analgesia nociception index, and nociception level index, and suggest how future developments could impact their use. The final section of this review explores the profound implications of future monitoring technologies on anesthesiology practice and envisages 3 transformative scenarios: helping in creation of an optimal analgesic drug, the advent of bidirectional neuron-microelectronic interfaces, and the synergistic combination of hypnosis and virtual reality.

BACKGROUND: Classification of perioperative risk is important for patient care, resource allocation, and guiding shared decision-making. Using discriminative features from the electronic health record (EHR), machine-learning algorithms can create digital phenotypes among heterogenous populations, representing distinct patient subpopulations grouped by shared characteristics, from which we can personalize care, anticipate clinical care trajectories, and explore therapies. We hypothesized that digital phenotypes in preoperative settings are associated with postoperative adverse events including in-hospital and 30-day mortality, 30-day surgical redo, intensive care unit (ICU) admission, and hospital length of stay (LOS). METHODS: We identified all laminectomies, colectomies, and thoracic surgeries performed over a 9-year period from a large hospital system. Seventy-seven readily extractable preoperative features were first selected from clinical consensus, including demographics, medical history, and lab results. Three surgery-specific datasets were built and split into derivation and validation cohorts using chronological occurrence. Consensus k-means clustering was performed independently on each derivation cohort, from which phenotypes’ characteristics were explored. Cluster assignments were used to train a random forest model to assign patient phenotypes in validation cohorts. We reconducted descriptive analyses on validation cohorts to confirm the similarity of patient characteristics with derivation cohorts, and quantified the association of each phenotype with postoperative adverse events by using the area under receiver operating characteristic curve (AUROC). We compared our approach to American Society of Anesthesiologists (ASA) alone and investigated a combination of our phenotypes with the ASA score. RESULTS: A total of 7251 patients met inclusion criteria, of which 2770 were held out in a validation dataset based on chronological occurrence. Using segmentation metrics and clinical consensus, 3 distinct phenotypes were created for each surgery. The main features used for segmentation included urgency of the procedure, preoperative LOS, age, and comorbidities. The most relevant characteristics varied for each of the 3 surgeries. Low-risk phenotype alpha was the most common (2039 of 2770, 74%), while high-risk phenotype gamma was the rarest (302 of 2770, 11%). Adverse outcomes progressively increased from phenotypes alpha to gamma, including 30-day mortality (0.3%, 2.1%, and 6.0%, respectively), in-hospital mortality (0.2%, 2.3%, and 7.3%), and prolonged hospital LOS (3.4%, 22.1%, and 25.8%). When combined with the ASA score, digital phenotypes achieved higher AUROC than the ASA score alone (hospital mortality: 0.91 vs 0.84; prolonged hospitalization: 0.80 vs 0.71). CONCLUSIONS: For 3 frequently performed surgeries, we identified 3 digital phenotypes. The typical profiles of each phenotype were described and could be used to anticipate adverse postoperative events.