Augmented Intelligence: Developing and Deploying AI in Anesthesiology & Critical Care

Christian S. Guay, MD

With the release of ChatGPT in November 2022, artificial intelligence (AI) has quickly become one of the most popular topics for mainstream headlines and social media. This transformative technology is quickly spreading through multiple industries, including healthcare, where AI has been studied for decades. In the session, “Artificial Intelligence and Machine Learning: Reality and Hype,” held Friday, April 14 at the IARS 2023 Annual Meeting, experts weighed in on the development and applications of AI in the perioperative and critical care settings.

Hannah Lonsdale, MBChB, an Assistant Professor of Anesthesiology at Vanderbilt University, opened the session with a general overview of the field of AI in anesthesiology. From a bird’s-eye view, the process of developing an AI model can be divided into discrete steps: acquire large data sets, preprocess the data, divide the data into training and validation sets, train and validate the model, tune the model’s hyperparameters and deploy the model. Of course, you can take a deep dive into each of these steps. For example, training can be supervised, semi-supervised, unsupervised or occur via reinforcement learning using artificial neural networks. Once deployed, some models can continue to learn and improve their performance. Alternatively, models may also be “locked” to their original configuration, resulting in static but predictable performance. Once deployed, AI models can then assist clinicians, usually by providing patient-specific predictions based on multidimensional clinical datasets.

A host of studies have shown that AI models built in this way outperform clinicians when making predictions, so why aren’t they more widely used? Most studies to date have been limited to single institutions, and developing your own in-house model can be a very time and resource-intensive process. Demand for AI expertise is also at an all-time high, further driving up costs. Unfortunately, models do not tend to perform as well when deployed in external institutions, which raises the issue of healthcare equity – will these powerful tools only be available at a subset of large medical centers who have the resources to develop them? Another concern is that most studies have focused on model development and validation rather than implementation and ethical considerations. For example, how should liability be shared by a clinician and the AI model they use for decision support? Data privacy and systematic bias are also important considerations that need to be addressed before large-scale deployment. Finally, there is the challenge of developing trust with clinicians. AI models are often viewed as black boxes and occasionally seen as a threat to job security. Focusing on clinician support rather than replacement will be important in building the necessary trust for real-world use.

Sabry Ayad, MD, a Professor of Anesthesiology at Cleveland Clinic, built on Dr. Lonsdale’s presentation with principles of AI systems and reviewed a large body of evidence for AI in perioperative care. He emphasized the importance of collecting high quality data and having robust preprocessing pipelines in place before training, citing the adage that the output of any model can only be as good as its input. The breadth of AI applications in anesthesiology is impressive:

• improving depth of anesthesia monitoring;
• titrating hypnotics, analgesics and neuromuscular blockade using closed loop systems;
• optimizing mechanical ventilation;
• automating analysis of ultrasound images;
• predicting chronic pain and response to therapy; optimizing operating room logistics including inventory management, resource utilization, scheduling and billing;
• generating predictions and risk stratification for hypotension, acute kidney injury and difficult airways;
• improving research trial design, analytics and recruitment; and
• increasing quality of care by identifying best practices and preventing medical errors.

Ronald Pearl, MD, PhD, a Professor in the Department of Anesthesiology, Perioperative and Pain Medicine at Stanford University, extended the discussion to critical care. Similar to the perioperative literature, AI has been studied for diverse uses in the intensive care unit (ICU), with a main focus on clinical prediction tools. Dr. Pearl highlighted that humans struggle with multidimensional data and rely heavily on heuristics to compensate for the fact that we often make decisions on less than six data points. The ICU is a data-rich environment and therefore ideal for AI systems that can handle the overwhelming amount of high-resolution multidimensional data streams. Fatigue is also common among ICU clinicians, so the assistance of an indefatigable decision-support system could help maintain high quality care 24/7. As noted by the previous speakers, external validation is an ongoing issue for large-scale deployment of AI systems in ICUs across multiple institutions. During the Q&A period, moderated by Jonathan Wanderer, MD, MPhil, FASA, FAMIA, Professor of Anesthesiology and Chair of Periprocedural Informatics Governance Committee at Vanderbilt University Medical Center, Dr. Pearl noted that institution-specific models may be the best way forward to account for the significant heterogeneity that limits model generalizability. He also noted the importance of focusing on clinician augmentation rather than replacement.

Matthew Zapf, MD, Assistant Professor of Anesthesiology at Vanderbilt University, concluded the session with a discussion of technical AI concepts and their implementation. Like the other speakers, he noted the critical importance of institutional support and multidisciplinary teams when building an AI framework that can actually be used in the clinical environment. Another recurring point was the need to account for racial algorithmic bias during model development and training. Ultimately, validated models can provide real-time decision-support and ease documentation burden for clinicians, allowing them to dedicate more time to directly caring for patients and their families.