Author: Young May Cha, MD
The Daily Dose
Artificial intelligence and machine learning were expected to transform healthcare by converting large clinical datasets into personalized decision-support tools. Despite the excitement and the growing number of published models, AI has had relatively little effect on everyday clinical decision-making.
At the 2026 IARS and SOCCA Annual Meeting, experts discussed why healthcare may be entering the “trough of disillusionment,” a stage in which early enthusiasm declines because a technology has not yet produced its promised clinical benefits. The panel focused particularly on the challenges of applying data science and AI to pediatric perioperative care.
The messy middle between development and clinical use
Theodora Wingert, MD, explained that developing and publishing a predictive model is only the first step. The difficult work begins afterward, during what she called the “messy middle.”
Before a model can be used clinically, it must be tested with real-time data, validated in different patient populations, incorporated into clinical workflows, and continuously evaluated after implementation.
Pediatric perioperative care creates additional difficulties. Pediatric populations are smaller and more heterogeneous than adult populations. Children also demonstrate major differences in physiology, development, anatomy, and normal clinical measurements depending on age.
These factors make it difficult to develop models that perform reliably across neonates, infants, children, and adolescents.
A model may also be statistically accurate but clinically useless. For example, an algorithm that predicts deterioration only after clinicians already recognize that the patient is critically ill does not improve decision-making.
Useful models must provide accurate information early enough to change treatment. Successful implementation therefore requires collaboration among anesthesiologists, surgeons, nurses, statisticians, informaticists, engineers, administrators, and electronic health-record specialists.
AI should be evaluated like a new medication
Hannah Lonsdale, MBChB, compared the development of AI systems with the development of new pharmaceuticals.
A drug is not adopted simply because it works in an early laboratory experiment. It must pass through multiple stages of testing, including validation, safety assessment, effectiveness studies, impact analysis, and implementation research.
AI models should undergo a similarly rigorous process.
The foundation must be a clinically important question supported by a large, detailed, and reliable dataset. Models developed from incomplete or low-quality information may produce misleading predictions regardless of how sophisticated the algorithm appears.
Many predictive systems also generate large numbers of false-positive alerts. This can create alarm fatigue, unnecessary testing, additional expense, and reduced confidence among clinicians.
Extremely large datasets from multiple institutions may be necessary to improve a model’s positive predictive value and demonstrate that it performs reliably outside the hospital where it was developed.
Traditional statistics may sometimes be better
Artificial intelligence is not automatically superior to conventional biostatistical methods.
Dr. Lonsdale described the NEO-READY model, which predicts whether a newborn is ready for discharge from the neonatal intensive care unit. The model was created using conventional logistic regression rather than a more complicated machine-learning approach.
The appropriate analytical method should be selected according to the clinical question and available data. Machine learning may be useful for extremely large datasets, complex nonlinear relationships, images, and unstructured text. Traditional statistical methods may be more transparent, easier to validate, and equally or more effective for other problems.
More complicated technology does not necessarily produce a more useful clinical tool.
Temporal drift and continuing validation
A model that performs well today may gradually become less accurate as clinical practice, patient populations, documentation systems, medications, and hospital policies change.
This process is known as temporal drift.
Predictive models must therefore be monitored and recalibrated throughout their clinical life. Validation cannot stop when the model is introduced.
AI systems may also create an unusual paradox. Once a model helps clinicians prevent complications, the outcomes used to train the model may occur less frequently. The model can then appear to perform worse because clinical care has improved.
Organizations must distinguish between genuine deterioration in model performance and apparent deterioration caused by successful clinical intervention.
Predicting persistent opioid use in teenagers
Tori Sutherland, MD, MPH, discussed the use of national claims data to study postoperative opioid prescribing and persistent use among teenagers.
Teenagers in the United States were prescribed more than twice the recommended maximum amount of opioids following several common operations. Prescribing rates were also considerably higher than those observed in other countries.
Prescription data do not prove that patients consumed all the medication. Researchers therefore used a new opioid prescription filled three to six months after surgery as a surrogate measure for persistent opioid use.
Factors associated with an increased risk included:
• Filling an opioid prescription before surgery
• Being older than 15 years
• Having chronic pain diagnoses such as headaches or abdominal pain
• Having depression, anxiety, substance-use disorders, or other mental health diagnoses
Unexpectedly, some of the highest-risk procedures were relatively low-pain procedures, including endoscopy, incision and drainage, and certain laparoscopic operations.
These findings prompted institutions to review prescribing practices and provide additional education to trainees about appropriate opioid prescribing.
Clinical significance
AI and machine learning have not yet consistently outperformed traditional methods or transformed pediatric perioperative care.
The greatest challenge is not creating another predictive model. It is developing a clinically meaningful tool, validating it across institutions, introducing it into existing workflows, demonstrating that it improves care, and monitoring it continuously for declining accuracy or unintended consequences.
Healthcare organizations should avoid adopting AI simply because it is new or technologically impressive. Each system should address an important clinical problem and provide information early enough to influence treatment.
AI may eventually help anesthesiologists predict complications, personalize care, improve prescribing, and identify high-risk patients. However, success will depend on high-quality data, multidisciplinary cooperation, rigorous validation, thoughtful implementation, and continued recalibration.
The path from data to better clinical decisions may be difficult and messy, but acknowledging these challenges is necessary for artificial intelligence to move beyond hype and become a dependable part of patient care.
Thank you to The Daily Dose, IARS, SOCCA, and the Society for Technology in Anesthesia for allowing us to summarize this important discussion of artificial intelligence in pediatric perioperative medicine.