Over the past several years, there has been a veritable explosion in the development, research, and utilization of digital technologies termed either artificial intelligence (AI) or machine learning (ML) across a wide breadth of fields, with health care being no exception. AI refers to any digital system that can perform tasks typically associated human intelligence, while ML is a subfield of AI and is the dominant form of AI utilized in health care research, relying on enormous volumes of data to accomplish specific tasks.¹ Prediction has been the primary focus of AI/ML research in health care, focusing on diagnosis, treatment selection, disease phenotyping, and prognostication of treatment outcomes.²

A brief description of ML is warranted as it is relevant to understanding how these methods may be applied in medicine. ML specifically involves the application of statistical algorithms to big datasets. While there are many types of ML, the 3 major forms include supervised learning, unsupervised learning, and reinforcement learning. The type of learning used is dependent on the problem an algorithm is asked to solve.

Supervised learning is the most common form currently utilized in health care and is likely to be most familiar to health care practitioners as it relies on the use of dependent and independent variables and linear model constructs. Regression, either logistic or linear, is the most common form of supervised learning models but may also include nonlinear algorithms such as decision trees. Supervised learning algorithms are given both an input (independent variables) and an output (dependent variables) and are tasked with developing models that predict the output. Unsupervised learning algorithms are only given an input and are asked to create clusters or structures of data, and may be used for phenotyping or drug discovery. Lastly, reinforcement learning tasks software with taking action in an environment so as to maximize a reward, and may be utilized in dynamic settings such as sepsis alert systems in intensive care units.

The increased attention to AI/ML technologies has also impacted the field of urogynecology and pelvic reconstructive surgery (URPS). Perioperative application of AI/ML may be theorized to fall into the following categories: prediction of treatment outcomes, prediction of complications, and postoperative care. To date, several models have been developed within each category. Werneburg et al applied neural networks to data from the ROSETTA trial to develop a prediction model for treatment selection for third-line therapy of overactive bladder with high classification ability,³ while Whitney et al utilized data from the ABC trial and applied linear regression to develop prediction models for postoperative success, time to recurrence, and risk of catheterization.⁴

Several models have been developed for prediction of both stress and urge urinary incontinence after pelvic organ prolapse surgery, as well as prolapse recurrence; all models relied on regression methods and utilized single institution data.^5,6 Models focusing on complications and postoperative care have primarily focused on prediction of infection and opioid utilization. Models predicting postoperative infection have primarily focused on urinary tract and surgical site infection, and were based on either single-institution or administrative databases and had moderate prediction ability.^7,8 With regard to available models that predict opioid dose after pelvic organ prolapse surgery, there are currently 2 available, both of which are based on single-institution data.^9,10

While the current models in URPS cover a relatively broad swath of application within the field, they all suffer from the same important limitation, as do the majority of prediction models. None of these models have been validated in real-world settings, or even in prospective or randomized trials, and thus their ability to provide meaningful clinical impact is unknown. Similarly, it is unknown whether the models pose a threat of harm to patients, as any intervention undertaken based on a prediction might lead to unforeseen consequences. Thus, prior to model implementation in a clinical setting, there needs to be adequate testing through a combination or real-world, prospective, and randomized controlled trials.

Several prominent examples of premature applications of AI include IBM Watson, which has been used with varying success as a physician clinical decision-aid tool across multiple disciplines and diagnostic tool, and Google Health, which developed an AI-based screening method for retinopathy. While both platforms performed brilliantly in the laboratory setting with high predictive performance, neither had significant clinical impact in the real world, and they have largely been sidelined by their creators. While the promise for AI in URPS and medicine in general is great, these examples highlight the potential pitfalls of early application of this technology without fully understanding its potential benefits, harms, and limits.