With more than 290 million surgical procedures performed globally each year management of postoperative pain is an important health care priority. While opioids are an effective and important tool for the treatment of pain in the perioperative period, there continues to be increased scrutiny with concerns around long-term dependence. As a result, there has been a proliferation of studies that evaluate prolonged postoperative opioid use, typically defined based on filled pharmacy claims for opioids after surgery using health administrative databases. Despite these studies, identification of patients at the highest risk of dependence remains a challenge. Opioid dependence occurs due to a complex interplay of cultural, psychologic, and neurobiologic factors which can be difficult to assess using traditional statistical models. In this editorial, we raise 2 important questions. First, how do we identify patients at the highest risk for opioid dependence after surgery? Second, and perhaps most importantly, how do we as perioperative physicians move beyond describing the size of the problem towards improving quality of life and reducing opioid-related harm?

In this issue of Anesthesia & Analgesia, 2 studies address important gaps in the literature. Simpson et al included 2611 patients undergoing spine surgery to evaluate the predictive accuracy of 5 different statistical models. They demonstrate that patients at the highest risk of persistent opioid use, defined as patient-reported opioid use 3 months postoperatively, could be predicted with greater accuracy using machine learning. Machine learning is pattern recognition developed into an algorithm to allow data (information) to be efficiently used by computational methods to improve performance, classify new data, or achieve accurate and precise predictions (output). The machine learns with each piece of additional information to improve on its output, and machine learning forms the backbone of artificial intelligence (AI). Machine learning algorithms are categorized into various types, namely supervised, unsupervised, semisupervised, and reinforcement learning. Supervised learning involves training on labeled datasets, unsupervised learning discovers patterns in unlabeled data, and semisupervised learning combines both. Furthermore, the analysis and interpretation of machine learning algorithms can take either a data-driven approach, relying on empirical patterns, or a theory-driven approach, emphasizing understanding the underlying principles governing the algorithm’s behavior. To be valid, supervised machine learning algorithms, as demonstrated in the study by Simpson et al need to accurately recognize a pattern in a known dataset (internal validity), which are typically tested in a proportion of the larger dataset that was used to build the algorithm. The model is then tested on an external dataset (external validity). Selecting an appropriate statistical model is critical and commonly used model approaches include logistic regression, decision tree, random forest, neural network, and others. Performance of a model is evaluated using metrics (performance measures), which depend on whether the model is a classification or a regression model.

Despite the perceived sophistication of machine learning approaches, traditional statistical methods, such as logistic regression, still offer robust discriminative accuracy and prediction. In a large meta-analysis of 71 studies, clinical prediction models did not demonstrate improved discrimination with machine learning relative to traditional statistical methods. Moreover, these traditional techniques are often less resource intensive, can be implemented in smaller health care settings, and offer greater transparency compared to more complex and opaque machine learning models. However, it is when datasets are expanded to include greater complexity or granularity, such as those derived from electronic medical records (EMRs), that machine learning techniques often excel.

Optimistically, the findings of Simpson et al provide a glimpse into a hopeful future, where machine learning has the potential to “analyze vast datasets” and predict those at the highest risk of harm incorporating the complex characteristics that may “escape clinician observation” in real time. However, moving forward, it is critical to assess whether these models can be incorporated into the physician workflow and evaluate whether prediction of risk can improve patient outcomes in clinical trials. Furthermore, each model is only as effective as the granularity and quality of the input data. Specific methodological flaws can occur due to inappropriate identification of population exposed; inappropriate diagnosis due to poor sensitivity or specificity; and inappropriate outcomes. Impressively, Simpson et al leveraged outpatient clinic notes, EMR data, preoperative pain scores, and prescription data. Some anesthesiology departments have recently partnered with departments of genetics and neuroradiology to obtain genetic and functional magnetic resonance imaging data to include in machine learning models, allowing incorporation of not only the social and clinical factors but the neurobiologic factors that may predict chronic pain and opioid use disorder. Ultimately, identification of patients at the highest risk of harm should enable clinicians to deliver tailored interventions to reduce harmful opioid use.

In this context, Gong et al studied the analgesic prescription and patient characteristics of >260,000 opioid-naïve patients who were dispensed opioids within 7 days of undergoing nontraumatic surgery in New Zealand. They identified that 9.1% of these patients filled another prescription for an opioid between 3 months and 1 year after discharge from surgery. Patients who were dispensed multiple opioids, a higher total morphine equivalent dose, or required a change in the opioid filled early after surgery were more likely to fill a prescription 91 to 365 days after surgery. Importantly, those coprescribed a nonopioid analgesic were less likely to fill a subsequent opioid prescription during this time. While the need for higher doses of opioid or the need to switch analgesics may reflect increased pain postoperatively, these findings also suggest that the choice of the initial analgesic prescribed by perioperative physicians may have lasting effects. For example, the Michigan Opioid Prescribing Engagement Network has developed surgery-specific analgesic prescribing guidelines that tailor opioid prescriptions and nonopioid adjuncts to the pain associated with each procedure. These guidelines were associated with a 63% reduction in prescription size for laparoscopic cholecystectomy and the decreased amount of opioid prescribed did not correlate with increased prescription refills.

As anesthesiologists and surgeons strive to create interventions that alleviate pain and diminish the risks associated with opioid-related harm, various opportunities emerge.

First, developing effective interventions to reduce harmful opioid use hinges on a thorough understanding of the indications prompting a filled opioid prescription. In a recent mixed cohort of surgical patients in Europe, most patients who reported opioid use 12 months postoperatively were taking opioids for reasons unrelated to the index surgery even amongst those who were opioid-naïve preoperatively. Similarly, in a second prospective cohort of patients who received guideline-directed opioid prescribing, long-term opioid prescriptions were most often the result of a new painful medical condition or new surgery. In a third cohort study of >600 postoperative patients, pain and opioid use at 3 months were not statistically associated. If the reason for an opioid prescription filled in the months after surgery is not delineated, then interventions that reduce pain or opioids during the index surgery may not be effective at reducing “persistent” opioid use. Indeed, both interventions to reduce early perioperative pain, such as regional anesthesia and state policies, such as those which limit opioid prescribing have failed to reduce persistent opioid use after surgery in large observational studies. Clinical trials and future studies must work to include an assessment of both opioid use and postoperative pain, and their underlying drivers, as these may respond differently to proposed interventions. Furthermore, these studies suggest a “persistent” filled opioid prescription might not always represent opioid-related harm. At the bedside, physicians and policymakers alike may need to reconcile with the reality that there may be some level of opioid use in the population. While we must remain mindful of the risk of opioid diversion, misuse, and adverse outcomes, a subset of patients may benefit from opioid analgesics to ensure quality of life.

Second, patients presenting for surgery are a highly heterogenous population that may respond differently to interventions that aim to treat pain or reduce the likelihood of opioid-related harm. Simpson et al identified age, preoperative opioid use, preoperative pain, and body mass index as the most important factors in the prediction of future filled opioid prescriptions among patients undergoing spine surgery. The study by Gong et al studied all-comers for surgery, irrespective of age, and identified preoperative hypnotic or nonopioid analgesic use, higher comorbidity burden, and ethnicity to be associated with persistently filled opioid prescriptions. The variation in risk factors identified across studies further underscores the notion that the drivers of pain and opioid requirements may be unique to each patient, surgical procedure, and cultural context. By comprehending the underlying drivers for prolonged opioid usage among different subgroups, such as age, ethnicity, history of chronic pain and other medical comorbidity, and cultural and environmental influences, we can tailor perioperative interventions to address these root causes and increase the likelihood of successful outcomes. Future studies should incorporate these high-risk subgroups in their study design to ensure they can capture potential heterogeneity of treatment effect.

Finally, these studies highlight opportunities to change our approach to clinical care. In Canada and the United States, nearly 80% of patients fill an opioid prescription within the first 7 days of outpatient surgery while in the study by Gong et al approximately 20% of surgical patients in New Zealand fill an opioid prescription within the same time frame. These findings reiterate the significant variation in analgesic prescribing practices globally, and that a proportion of patients may not require a postoperative opioid prescription. While attaining effective opioid-free analgesia postsurgery can be challenging for certain patients and procedures, an increasing body of evidence indicates that many patients undergoing specific operations, such as appendectomy and thyroidectomy, can be suitably discharged with opioid-free analgesic prescriptions. As anesthesiologists are often not responsible for the analgesic prescription at discharge, we must continue to collaborate with the surgical multidisciplinary team. In one study, patients who filled opioid prescriptions by surgeons in the highest quintile of opioid prescribing intensity had an elevated risk of persistent opioid use postoperatively. Therefore, provider education for both anesthesiologists and surgeons may play an important role in reducing potential opioid-related harm. Furthermore, in the studies by both Simpson et al and Gong et al patients with preoperative pain and analgesic use were among those at the highest risk of persistent opioid requirements postoperatively. While further study is required, perioperative optimization of pain and opioid requirements through transitional pain services may benefit these high-risk patients. Additionally, multifaceted approaches to opioid management that involve opioid adjuncts, enhanced recovery after surgery as well as bundled care pathways may also facilitate individualized opioid prescribing after surgery.

In summary, Simpson et al demonstrated the potential to improve our prediction of high-risk patients using novel machine learning techniques, while Gong et al address the important challenge of identifying potential interventions to reduce persistent opioid use after surgery. The synergy of these studies lies in recognizing that accurate prediction is most valuable when integrated with effective strategies to reduce the risk of harm. Given the scarcity of interventions that have durably reduced the incidence of persistent postsurgical opioid use, a concentrated effort is needed to understand the underlying mechanisms driving this usage. Further research must focus on developing interventions that both alleviate pain and mitigate the risks associated with opioid-related harm, recognizing that these objectives may not always align. It is only then that we will be able to fully realize the benefits of improved prediction.