Artificial Intelligence in Anaesthesia: Current Uses, Recent Advances, and Future Directions

1. Background: Why AI in Anaesthesia?

1. Depth of anaesthesia (DoA) monitoring
2. Image-guided regional techniques
3. Perioperative risk prediction
4. Automated drug administration

2. Depth of Anaesthesia Monitoring

• The Bispectral Index (BIS) has long processed EEG signals using a proprietary algorithm. Newer AI-based monitors go further: deep neural networks can predict BIS values in real time and flag transitions between anaesthetic states with greater accuracy than legacy rule-based systems.
• A deep learning model (Hwang et al., 2023, PMID 37673590) achieved interpretable BIS prediction using convolutional neural networks applied to raw EEG, providing clinicians visibility into why the model made its prediction - addressing the "black box" concern.

3. Closed-Loop Drug Delivery ("Autopilot Anaesthesia")

• Reviewed 26 studies (2010-2025) on ML-based closed-loop anaesthesia
• Architectures used: reinforcement learning (RL), convolutional neural networks (CNNs), and ensemble tree methods (e.g. random forests)
• Key finding: ML-guided controllers outperformed conventional PID algorithms in maintaining target BIS range, reducing total propofol dose, and improving haemodynamic stability
• Specifically, a meta-analysis within the closed-loop literature showed: +15.3% time in target BIS range, -0.8 mg/kg/h propofol consumption, and -3.2 min time to eye opening after surgery
• No system was associated with increased intraoperative awareness despite lower drug doses

4. Perioperative Risk Prediction

4a. Hypotension Prediction Index (HPI)

• HYPE-2 RCT (Schuurmans et al., 2025, PMID 39576150, Crit Care Med): Machine learning-derived early warning with treatment protocol significantly reduced hypotension burden during cardiac surgery and ICU stay - one of the first RCTs demonstrating clinical outcome benefit from a perioperative AI tool.
• A 2024 narrative review (PMID 38325215) confirmed HPI reduces intraoperative hypotension episodes across multiple surgical settings.

4b. Other Perioperative Risk Models

• Postoperative delirium prediction: ML models integrating age, cognitive baseline, intraoperative data, and inflammatory markers to identify high-risk patients for targeted preventive strategies.
• PONV risk: Classical risk scores (Apfel) are being augmented by ML models using EHR data that outperform them in discrimination.
• Long-term pain after surgery: Machine learning applied to preoperative patient characteristics to predict risk of chronic postsurgical pain (presented at ASRA 2025, Hospital for Special Surgery, using knee replacement data).
• Sepsis in ICU/AICU: RL models (Roggeveen et al.) trained on ICU data optimise fluid and vasopressor therapy, though specificity to surgical patients remains limited.

5. Image-Guided Regional Anaesthesia

• Brachial plexus blocks
• Neuraxial procedures (epidurals, spinals) - ultrasound-AI fusion for landmark identification in obese or anatomically difficult patients

6. Operating Room Management

7. Critical Care Applications

• Real-time early warning systems for deterioration
• Ventilator weaning optimisation
• Sepsis recognition and treatment guidance
• Fluid responsiveness prediction

• "Black box" algorithms reduce clinician trust
• Databases used for training often don't reflect target populations (fairness/bias)
• Real-time data integration from multiple streams is technically difficult
• Legal liability for AI-assisted errors remains unresolved

8. Natural Language Processing and Large Language Models

• Preoperative documentation: Automated generation of anaesthetic plans from EHR data
• Patient communication: AI-generated, personalised consent and education materials in multiple languages and reading levels (ASRA 2025 data)
• Clinical documentation: Intraoperative record generation, reducing clerical burden
• Differential diagnosis support: Medical generalist LLMs assist with cross-specialty decision-making during complex perioperative presentations

9. Obstetric Anaesthesia

• Predictive models for spinal hypotension (a common problem in obstetric regional anaesthesia)
• Personalised vasopressor dosing algorithms
• Risk stratification for conversion from regional to general anaesthesia

10. Education and Training

• AI-driven simulation with adaptive feedback (adjusting difficulty based on trainee performance)
• VR/AR platforms for regional anaesthesia procedural training
• Automated assessment of competency in ultrasound scanning and airway management
• NLP analysis of patient questions to guide what trainees are taught about consent

11. Future Directions

12. Barriers and Ethical Considerations

• Algorithmic bias: Most training datasets over-represent certain demographics; models may perform poorly in underrepresented groups.
• Black-box interpretability: Clinicians and regulators need to understand why AI makes a recommendation, especially for high-stakes decisions.
• Data privacy and security: Medical data used for AI training requires robust governance; cybersecurity of connected drug delivery systems is a non-trivial risk.
• Liability: If an AI-assisted decision leads to patient harm, responsibility attribution between clinician, institution, and vendor remains legally unsettled.
• Workflow integration: Even well-validated tools fail at scale if they don't fit clinical workflow or generate alert fatigue.
• Patient acceptance: Variable; patient-facing AI (consent tools, communication aids) requires explicit explanation and consent.

Summary

• Lopes et al., 2024 - Systematic review of AI in anaesthesia (PMID 37864754)
• Zarei & Torgerson, 2025 - ML for precision anaesthesia/closed-loop (PMID 41458798)
• Bellini et al., 2024 - AI in operating room management (PMID 38353755)
• Schuurmans et al., 2025 - HYPE-2 RCT: ML hypotension prediction (PMID 39576150)
• Pinsky et al., 2024 - AI in critical care (PMID 38589940)
• Shu et al., 2025 - Applications of AI in anaesthesiology (Springer)