Preprints.org logo
Preprint
Review

This version is not peer-reviewed.

Machine Learning and Postoperative Pain: A 2026 Critical Narrative Review

Submitted:

21 September 2026

Posted:

22 September 2026

You are already at the latest version

Abstract
Postoperative pain remains a common and clinically important problem, with inadequate control contributing to delayed recovery, prolonged opioid exposure, and risk of chronic postsurgical pain. Machine learning has been suggested as a strategy to improve postoperative pain management by predicting pain trajectories, estimating opioid requirements, identifying patients at risk for poor outcomes and improving the clinical management. This critical review examines current applications of machine learning in postoperative pain, including preoperative risk stratification, intraoperative and postoperative prediction, objective assessment, wearable monitoring, discharge opioid prescribing, and use in specialized populations.We found that although many models demonstrate moderate to strong predictive performance, their clinical utility remains uncertain. Most studies are retrospective, single-center, or procedure-specific, with limited external validation and heterogeneous outcome definitions. While many models incorporate postoperative variables, this influences the stage of care at which they can be applied, making them more useful for guiding postoperative management than for informing preoperative or intraoperative interventions. As a result, future work should move beyond prediction accuracy toward prospective validation, actionable hypothesis driven multi-modal models, and large-scale trials that determine whether machine learning–guided care improves pain control while reducing unnecessary opioid exposure.
Keywords: 
machine learning
;  
pain
;  
predictive modeling
Subject: 
Medicine and Pharmacology  -   Anesthesiology and Pain Medicine

Introduction

Postoperative pain remains a pervasive challenge in surgical care, with up to 80% of patients reporting significant discomfort after procedures.[1] If inadequately controlled, this pain initiates a cascade of adverse physiological effects, including increased sympathetic tone, impaired pulmonary function, and delayed mobilization, which collectively prolong hospital stays and elevate the risk of chronic postsurgical pain.[2] Historically, the foundation of acute pain management has relied heavily on opioid-based analgesia. [2,3] However, the limitations of this approach are increasingly evident. Opioid administration is associated with dose-dependent side effects such as respiratory depression, ileus, and sedation, and more critically, it serves as a common entry point for long-term opioid dependence and substance use disorders.[4,5] New persistent opioid use is a recognized postoperative complication characterized by continued opioid use after the normal healing period in previously opioid-naïve patients, affecting approximately 5-7% of patients after common surgical procedures.[6] Furthermore, persistent opioid use after surgery may be a risk factor for development of opioid use disorder.[7]
In recent years, there has been growing interest in leveraging clinical data and machine learning (ML) to integrate precision medicine-focused concepts into perioperative pain management. A wide range of predictive models have been developed to forecast postoperative pain trajectories, estimate opioid requirements, and identify patients at risk for inadequate pain control and prolonged opioid use.[8,9,10] These approaches leverage a variety of sources, such as existing databases, facial expressions, wearable data, and procedural variables, to generate patient-specific pain predictions.[8,9,11] However, despite promising performance metrics in existing retrospective and prospective studies, these models have largely remained confined to research settings and not operationalized into real clinical practice. As a result, few, if any ML-based tools, have been successfully adopted in routine perioperative care. Addressing this disconnect is essential to realizing the full potential of these technologies.
This special article examines the current landscape of ML applications in postoperative pain management. We explore key use cases and discuss the opportunities and barriers to translating these models from development to bedside practice.

Search Strategy and Selection Criteria

A structured literature search was conducted in PubMed to identify studies evaluating ML applications in postoperative pain management. The search query combined terms related to machine learning and artificial intelligence with terms related to postoperative pain and pain assessment. Additional studies were identified through manual review of the reference lists of included articles and relevant review papers. Only English-language articles were considered. The search was conducted through April 2026.
Studies were included if they evaluated ML-based approaches for predicting, assessing, or managing postoperative pain or opioid related outcomes in surgical patients. Eligible studies included retrospective and prospective observational studies, cohort studies, clinical validation studies, and proof-of-concept investigations involving adult or pediatric populations. Studies examining preoperative risk stratification, intraoperative or postoperative pain prediction, objective pain assessment, wearable-based monitoring, discharge opioid prescribing, or specialized surgical populations were considered. Review articles, editorials, conference abstracts without sufficient methodological detail, and studies not focused on postoperative pain outcomes were excluded. Because the objective of this article was to provide a critical narrative overview of the field, formal risk-of-bias scoring and quantitative pooling of results were not performed.

Preoperative Risk Stratification

ML models have explored to use preoperative data to forecast postoperative pain trajectories and opioid needs. These models could guide selection of enhanced recovery after surgery pathways, facilitate shared decision-making regarding surgical approach and recovery expectations, and enable preemptive interventions such as regional anesthesia or multimodal analgesia for patients identified as high risk.
Parthipan et al. demonstrated that preoperative pain, surgery type, and opioid tolerance, along with SSRI-opioid interactions, were strong predictors of postoperative acute pain. Their model achieved area under the receiver operating characteristics curve (AUC) of 0.87 at discharge that declined to 0.69 at later follow-up.[11] This deterioration illustrates an important limitation: models calibrated to the immediate postoperative period may not capture the psychosocial and behavioral factors that increasingly shape pain over time. Moreover, although outcomes measured at discharge can guide postoperative interventions such as intensified analgesic management or referral to transitional pain services, their value for informing earlier perioperative decision making is inherently limited. Sun et al. reported excellent performance (AUC ~0.92-0.96) using a broad set of clinical, cognitive, and intraoperative variables.[12] While this multidimensional approach likely improves discrimination, inclusion of postoperative variables complicates interpretation of the model’s practical value, as prediction partly depends on information unavailable during preoperative decision-making. This reflects a recurring tension in ML pain prediction studies, as models optimized for accuracy often rely on data inaccessible at the point when intervention would be most clinically useful. Nair et al. demonstrated more modest performance (~70-72% accuracy), with procedure type and medical history as dominant predictors.[13] Notably, the model showed poor sensitivity for patients with high opioid requirements, highlighting a common limitation where models perform well on majority low-risk groups but fail in the clinically critical tail of the distribution.[13] This is a clinically significant weakness because these high-need patients are precisely the subgroup in whom accurate prediction would be most actionable. A model that performs well for the majority of lower-risk patients but fails to detect outliers offers limited value for targeted intervention.
Taken together, these studies suggest that strong statistical discrimination does not necessarily translate to clinical utility. Pain is a subjective and context-dependent outcome influenced by psychosocial stressors, prior pain experiences, cultural expectations, and social determinants of health, many of which remain poorly represented in structured datasets. Consequently, even high-performing models may encounter an upper limit in predictive accuracy. Furthermore, dependence on retrospective single-center cohorts, limited external validation, and inconsistent thresholds for clinical action constrain implementation. As a result, current models may be better positioned as tools for population-level risk stratification and hypothesis generation rather than reliable systems for individualized, real-time perioperative pain management. This gap reflects a broader problem in which pharmacogenomic, physiologic, behavioral, and medication adherence data are studied in isolation rather than integrated within a unified analytic framework.
Building on this, risk assessment models have been proposed to support preventive, multidisciplinary, and patient-specific pain management by leveraging existing electronic medical record data without requiring additional data collection.[14] Kagerer et al. achieved an AUC of 0.82 in predicting postoperative pain using 807 features mapped from medical record data, such as demographics, ICD-10 diagnoses, laboratory values, procedure codes.[14] A major strength is its use of existing clinical data without requiring additional measurements. This improves scalability compared to more complex or specialized data-driven models. However, the large number of input features increases model complexity, and depending on the modeling approach, may make it more difficult to identify the clinical drivers of predictions. The absence of external validation also reduces confidence in generalizability across institutions. In addition, like many predictive models in this space, its clinical utility remains limited by a lack of defined implementation pathways and uncertainty regarding how risk estimates should translate into perioperative management decisions.

Intraoperative and Early Postoperative Pain Prediction

ML models leveraging intraoperative and early postoperative physiologic signals have demonstrated promising ability to predict acute pain in the post-anesthesia care unit, although performance varies depending on the signal type and model complexity.
Models such as Zhang et al.’s DoseFormer have demonstrated strong predictive performance for acute postoperative pain using intraoperative vital signs and patient characteristics.[8] Built using graph-based deep learning architectures with attention mechanisms, the model achieved an AUROC of 0.98 and F1 score of 0.85.[8] A major strength of this approach is its ability to integrate longitudinal physiologic trends with static clinical variables, potentially capturing nonlinear perioperative pain patterns that traditional regression models overlook. While these results are impressive, they should be interpreted cautiously. The model was developed in a retrospective cohort restricted to a single procedure type, which likely reduced heterogeneity and may have inflated performance relative to what would be expected in broader surgical populations. In addition, an AUROC approaching 1.0 raises concern for overfitting, particularly when complex architectures are applied to relatively homogeneous datasets. The graph-based design may capture nonlinear temporal patterns that simpler models miss, but its limited interpretability may hinder clinician trust and make implementation difficult unless predictions can be clearly linked to actionable features.
Objective pain assessment tools are also being studied to address patients unable to self-report, such as those who are sedated or cognitively impaired.[15,16] Ryu et al. developed XGBoost models using photoplethysmography (PPG) waveform features collected throughout the perioperative period, achieving AUCs of 0.819 intraoperatively and 0.927 postoperatively, with performance exceeding that of a commercial surgical pain index.[15] A key strength of this study is its use of standard pulse oximetry data, which are routinely available and do not require additional monitoring equipment. However, several aspects limit its clinical applicability. First, the model was trained on pain assessments obtained at discrete events that are intrinsically associated with nociceptive stimulation, such as intubation and incision. This design may allow the model to distinguish procedural phases rather than detect pain in a more generalizable sense. Additionally, model performance depended on engineered waveform features that may be sensitive to motion artifact, vasopressor use, or hemodynamic instability, potentially limiting reproducibility across institutions and monitoring platforms. Together, these considerations suggest that the study provides an important proof of concept, but additional validation in more heterogeneous surgical settings is needed before PPG-based pain prediction can be considered clinically actionable.
Morrison et al. demonstrated that individual intraoperative nociception metrics had limited predictive ability for moderate to severe PACU pain (AUC 0.65–0.67).[17] However, combining multiple nociception-derived variables using penalized logistic regression improved performance to a cross-validated AUC of 0.753, suggesting that nociception signals alone may be insufficient and that multivariable models integrating additional physiologic and clinical data are likely necessary for reliable postoperative pain prediction.[17] Despite this improvement, several limitations temper the clinical significance of the results. First, the study population was restricted to relatively healthy women undergoing a single type of minimally invasive gynecologic surgery, which limits generalizability to other procedures and patient populations. Additionally, the NOL index itself reflects autonomic responses that can be influenced by factors unrelated to pain, including anesthetic depth, beta-blockade, hemodynamic changes, and surgical stimulation. As a result, the study supports the concept that intraoperative nociception contains predictive information, but it also highlights that these signals alone explain only a modest proportion of postoperative pain variability and are not yet sufficiently robust for routine clinical decision-making.
Subramanian et al. developed a multimodal machine learning framework integrating ECG, EMG, electrodermal activity, and respiratory signals to classify pain levels with over 80% balanced accuracy.[10] In theory, such systems could enable automated and more objective pain monitoring, reduce the burden of repeated nursing assessments, and facilitate earlier recognition of inadequate analgesia. Continuous physiologic monitoring may also provide a more granular view of pain trajectories over time, with potential applications in both individualized care and quality improvement. However, the practical value of this approach remains uncertain. In the study, single modalities, particularly respiratory rate, often performed as well as or better than the full multimodal model, suggesting that the added complexity of combining multiple biosignals may offer limited incremental benefit. Moreover, implementation would require multiple sensors, robust signal preprocessing, and specialized analytic infrastructure, all of which create substantial barriers to integration into routine postoperative workflows.
Other models have expanded predictor sets to include genetic and experimental pain sensitivity variables. Kumar et al. incorporated SNPs (e.g., OPRM1, COMT), cold pain test scores, pupillary response to fentanyl, and clinical characteristics to predict postoperative fentanyl requirements, pain scores, and time to first analgesic. Model performance was moderate (R2 ~0.31-0.53 across outcomes), suggesting that even biologically rich, multimodal inputs only partially explain variability in pain and opioid needs.[18] In addition to only modest predictive gains, this approach requires genetic testing and specialized experimental assessments that increase cost and workflow complexity, making it impractical for routine perioperative use.

Objective Pain Assessment After Surgery

Facial expressions have also been used to predict Numeric Rating Scale (NRS) scores.[19,20] Fontaine et al. developed a model that identified severe pain (≥7/10) with 77.5% sensitivity, vastly outperforming bedside nurses who identified only 17.0% of cases via visual observation.[19] This suggests that deep-learning systems using computer vision can provide a standardized metric to assist physicians in detecting acute distress. However, the model had a mean error of 2.4 points, which may prevent it from meeting the rigorous threshold for clinical use, as a 2-point shift on the NRS often represents the minimum clinically important difference that triggers a change in analgesic strategy.[19] An error of this magnitude could lead to inappropriate opioid titration or the failure to treat significant distress. Similarly, Park et al. found that machine learning models based on facial expression analysis predicted severe postoperative pain with high accuracy (AUC 0.93), outperforming models derived from physiologic measures such as vital signs and analgesia nociception index data.[16] While facial recognition offers a noninvasive and intuitive method for automated pain assessment, implementation remains limited by privacy concerns, dependence on camera positioning and lighting conditions, and reduced reliability in masked, sedated, or partially obscured patients. Collectively, these studies demonstrate strong predictive performance in controlled settings but face barriers to widespread adoption due to limited external validation, workflow integration challenges, interpretability concerns, and uncertainty regarding real-world generalizability.
Similar approaches have been explored in pediatric populations, where pain assessment is particularly challenging. Aydin and Ozyazicioglu found that machine learning–based facial expression analysis more closely approximated self-reported pain scores than caregiver or nurse assessments (ICC 0.414), supporting its potential role in patients with limited ability to communicate.[20] However, performance was also only moderate, and implementation would require reliable video capture and validation across more diverse pediatric populations. Together, these limitations help explain why facial recognition systems remain investigational rather than part of routine pediatric pain assessment.
Wearable devices represent another promising avenue for extending postoperative monitoring beyond the hospital. Morimoto et al. developed ML models using physiologic data from smart rings to predict poor postoperative pain alleviation in a cohort of 37 patients with 70% accuracy and an AUROC of 0.762.[21] While this study demonstrates the feasibility of collecting wearable data during recovery, the small sample size limits confidence in model stability and generalizability. In addition, performance was only comparable to existing models based on routine preoperative variables, raising questions about whether wearable devices provide enough incremental benefit to justify the added cost, data management burden, and reliance on patient adherence.
Soley et al. achieved stronger predictive performance in a cohort of 347 patients by integrating wearable-derived features with preoperative electronic health record (EHR) data.[22] Their stacking ensemble model reached an AUROC of 0.88 for severe acute postoperative pain and 0.74–0.90 across models for predicting chronic opioid use.[22] The superior performance of Soley et al.’s approach likely reflects both the added value of combining longitudinal physiologic signals with structured clinical data and the benefit of a larger, more diverse training cohort. This suggests that wearable data alone may be insufficient and that integration with EHR variables is key to maximizing predictive utility. However, several predicted outcomes were aggregated over approximately one-day intervals or longer. While this may be valuable for retrospective trend detection, predictions at such coarse time scales are less useful for real-time clinical intervention, where decisions about analgesic adjustments often need to occur over hours rather than days.
Despite these promising findings, models remain early in development. Most studies involve small cohorts and single-center datasets, which limit generalizability. In addition, physiologic signals captured by wearables may be influenced by numerous factors unrelated to pain, including sleep disruption, medications, mobility limitations, and comorbid conditions. Practical barriers also remain, including patient adherence to device use, integration of wearable data into clinical information systems, and uncertainty about how clinicians should act on algorithm-generated alerts. Consequently, although wearable-based machine learning systems demonstrate potential for continuous postoperative monitoring, further large-scale validation and workflow integration studies are needed before they can be routinely implemented in perioperative care.

Discharge Opioid Prescription Planning

ML models are being explored to optimize opioid prescribing at discharge, an area in which it is critical to balance adequate pain control with the risk of overprescribing.[9,23,24] These tools could support more personalized discharge prescriptions, identify patients who may need closer follow-up for pain management, and reduce excess prescribing that contributes to opioid diversion and misuse.
Gabriel et al. developed models to predict opioid refills after ambulatory surgery, identifying regional nerve block use, PACU pain scores, smoking history, and perioperative opioid consumption as key predictors, with moderate performance (AUC = 0.75).9 Similarly, Salehinejad et al. reported accuracies around 0.78–0.79 for predicting opioid refills, with procedure type, in-hospital pain scores, and discharge opioid dose as dominant predictors, but these features again reflect postoperative care processes rather than baseline risk, suggesting partial circularity and limiting early intervention potential.[23] Although these studies identified patients at increased risk of requiring additional opioids, much of the predictive signal came from postoperative variables such as PACU pain scores, in-hospital pain intensity, and discharge opioid dose. Because these data are only available after surgery, they offer limited value for earlier risk stratification or preoperative planning. In addition, moderate discrimination may not be sufficient to support individualized prescribing decisions.
In contrast, Simpson et al. developed models to predict persistent opioid use following major spine surgery, with a balanced random forest achieving excellent discrimination (AUC=0.88).[25] The most influential predictors included age, preoperative opioid use, preoperative pain scores, and body mass index, demonstrating the value of incorporating preoperative patient characteristics for risk prediction. However, their models also targeted a procedure-specific population with a high baseline prevalence of persistent opioid use, which may limit generalizability to broader surgical populations.
Klemt et al. reported excellent performance (AUC > 0.83) for predicting prolonged opioid use, defined as opioid prescriptions extending beyond 90 days, after primary total knee arthroplasty.[24] The model’s strongest predictors were preoperative opioid duration, substance use, and depression, which are well-established clinical risk factors that are readily identifiable without advanced analytics.[24] This suggests that much of the model’s predictive power stemmed from a small number of highly informative variables rather than from the complexity of the machine learning approach itself. In addition, the study was limited to a single procedure type, which may reduce generalizability to other surgical populations. Although the authors demonstrated favorable calibration and decision curve analysis, the study did not assess whether incorporating the model into clinical practice changes prescribing behavior or improves patient outcomes. As a result, the incremental benefit of this approach over simpler risk stratification methods remains uncertain.
Collectively, these studies suggest that while ML models can achieve moderate to high discrimination, much of their predictive power derives from a small set of well-known clinical risk factors and, in many cases, from postoperative variables that reduce preoperative applicability. This highlights a key limitation in the literature: models often optimize prediction rather than clinical usefulness, with performance inflated by data that are not actionable at the time decisions must be made. As a result, despite consistent identification of high-risk phenotypes, including patients with prior opioid exposure, high pain scores, and psychosocial comorbidities, the added value of complex ML approaches over simpler risk stratification tools remains uncertain, and integration into clinical workflows is hindered by limited generalizability, timing of predictors, and lack of clear intervention pathways.

Special Populations

Models have also been tailored for specialized populations that face unique challenges in postoperative pain management. In pediatrics, for example, developmental differences complicate postoperative pain assessment. A computer vision and machine learning approach developed by Sikka et al. achieved AUCs of 0.84 to 0.94 for detecting postoperative pain based on facial expressions in children following appendectomy. [26] This study highlights the potential for automated, objective pain assessment to reduce variability and mitigate the well-documented tendency of clinicians to underestimate pediatric pain. However, the cohort was small (n=50), restricted to neurotypical children undergoing a single procedure, and evaluated under controlled postoperative conditions. These limitations raise important questions regarding generalizability to more diverse pediatric populations, including younger children, patients with atypical facial expressions, and those undergoing different types of surgery.
Cancer patients represent another population in whom postoperative pain is complicated by chronic opioid exposure, neuropathic pain mechanisms, and polypharmacy. In a retrospective study of 42 hospitalized patients with colorectal cancer, Random Forest models predicted treatment response to analgesic therapy with an AUC of 0.92, while UMAP clustering identified two phenotypic groups with differing analgesic needs.[27] Although these findings are promising, the very small sample size and narrow disease focus create a substantial risk of overfitting and limit confidence in the stability of the identified predictors. In addition, the model predicted response over a 2 to 3 week follow-up period, which is useful for understanding longitudinal pain control but less applicable to immediate perioperative decision-making.

Cross-Cutting Limitations of the Current Literature

Taken together, studies in this review highlight several recurring limitations that continue to restrict the clinical translation of machine learning models for postoperative pain.
Most models have been developed using retrospective, single-center, or procedure-specific datasets, which limits confidence in their generalizability across institutions, surgical populations, anesthetic techniques, and postoperative care pathways. External validation remains uncommon, and models that are evaluated only within the same health system or database in which they were developed may overestimate real-world performance. In addition, studies use heterogeneous outcome definitions, including pain scores at different postoperative time points, opioid consumption, opioid refill requests, and prolonged opioid use. This variability makes it difficult to compare models across studies or determine which outcomes are most clinically meaningful.
Another important limitation is the quality of the data on which these models are trained. Machine learning approaches depend on the assumption that the input variables and outcome labels accurately reflect the clinical phenomenon being predicted. In postoperative pain, this assumption is often problematic. Pain scores are subjective and may vary depending on patient communication, nursing workflow, timing of assessment, and institutional documentation practices. Similarly, opioid administration, refill requests, and discharge prescriptions may reflect clinician prescribing behavior or local practice patterns rather than pain severity alone. As a result, models trained on these data may learn patterns of documentation and treatment rather than true postoperative pain trajectories.
This concern also applies to physiologic and wearable-derived data. Signals such as heart rate, electrodermal activity, sleep disruption, movement, and photoplethysmography may be influenced by anxiety, medications, autonomic tone, mobility restrictions, comorbid disease, and postoperative recovery itself. These variables may therefore act as indirect and nonspecific markers rather than precise measures of pain. Consequently, even technically sophisticated models remain vulnerable to the basic limitation that inaccurate, incomplete, or biased inputs will produce unreliable predictions. This issue is especially important in postoperative pain because the outcome is subjective, multidimensional, and inconsistently measured across studies
Lastly, many models remain better suited for statistical prediction than for clinical decision support. Although advances in interpretability techniques, including SHAP (SHapley Additive exPlanations) and feature ablation analyses, have made it easier to identify the variables driving model predictions, these insights do not necessarily translate into clear, evidence-based clinical actions.[28] For subjective and multifactorial outcomes such as postoperative pain, knowing which factors contribute most to a patient’s predicted risk does not establish how analgesic management should be modified to improve outcomes. Moreover, several models rely on postoperative variables, including PACU pain scores, inpatient opioid use, or discharge prescribing patterns, which are only available after key perioperative decisions have already occurred. As a result, even models with strong performance metrics have seen limited clinical adoption because they often fail to provide timely, interpretable, and actionable predictions that can be linked to specific interventions and improved patient outcomes.

Future Directions

We suggest that future studies should move beyond retrospective model development and focus on prospective validation in real clinical workflows. Although many models demonstrate promising discrimination in single-center datasets, few have been tested across institutions, surgical populations, anesthetic techniques, or postoperative care pathways. Prospective studies are needed to determine whether model performance is maintained in real time and whether model-guided care improves outcomes, reduces poorly controlled pain, limits unnecessary opioid exposure, or improves postoperative monitoring.
Another important direction is the integration of multimodal data. Postoperative pain is shaped by biologic, psychological, behavioral, procedural, and social factors, and models based on a single data stream are unlikely to capture this complexity. Future systems should integrate EHR-derived clinical variables with patient-reported outcomes, medication adherence, wearable physiologic signals, psychosocial measures, and, where feasible, genomic or pharmacogenomic data. However, additional data sources should be included only if they provide incremental clinical value, remain feasible to collect, and generate predictions that are interpretable and actionable.
Ultimately, the field must move from models that predict postoperative pain to systems that support pain management within clinical decision support. This will require explainable models, clinically meaningful outcome definitions, and prospective trials that evaluate whether algorithm-guided interventions improve care compared with standard practice. Without this shift from prediction to intervention, machine learning models may continue to show strong performance metrics while providing limited benefit at the bedside.

Funding

No funding declared.

Conflicts of Interest Disclosures

Dr. Zeev N. Kain is the President of the American College of Perioperative Medicine and is funded by the National Institutes of Health.

References

  1. Gan TJ, Habib AS, Miller TE, White W, Apfelbaum JL. Incidence, patient satisfaction, and perceptions of post-surgical pain: results from a US national survey. Curr Med Res Opin. 2014;30(1):149-160. [CrossRef]
  2. Lovich-Sapola J, Smith CE, Brandt CP. Postoperative Pain Control. Surg Clin North Am. 2015;95(2):301-318. [CrossRef]
  3. Glare P, Aubrey KR, Myles PS. Transition from acute to chronic pain after surgery. Lancet. 2019;393(10180):1537-1546. [CrossRef]
  4. Brummett CM, Waljee JF, Goesling J, et al. New Persistent Opioid Use After Minor and Major Surgical Procedures in US Adults. JAMA Surg. 2017;152(6):e170504. [CrossRef]
  5. Clarke H, Soneji N, Ko DT, Yun L, Wijeysundera DN. Rates and risk factors for prolonged opioid use after major surgery: population based cohort study. BMJ. 2014;348:g1251. [CrossRef]
  6. Brummett CM, Waljee JF, Goesling J, et al. New Persistent Opioid Use After Minor and Major Surgical Procedures in US Adults. JAMA Surg. 2017;152(6):e170504. [CrossRef]
  7. Bicket MC, Lin LA, Waljee J. New Persistent Opioid Use After Surgery: A Risk Factor for Opioid Use Disorder? Ann Surg. 2022;275(2):e288-e289. [CrossRef]
  8. Zhang C, Zhao X, Zhou Z, Liang X, Wang S. DoseFormer: Dynamic Graph Transformer for Postoperative Pain Prediction. Electronics. 2023;12(16):3507. [CrossRef]
  9. Gabriel RA, Simpson S, Zhong W, Burton BN, Mehdipour S, Said ET. A Neural Network Model Using Pain Score Patterns to Predict the Need for Outpatient Opioid Refills Following Ambulatory Surgery: Algorithm Development and Validation. JMIR Perioper Med. 2023;6(1):e40455. [CrossRef]
  10. Subramanian A, Cao R, Naeini EK, et al. Multimodal Pain Recognition in Postoperative Patients: Machine Learning Approach. JMIR Form Res. 2025;9:e67969. [CrossRef]
  11. Parthipan A, Banerjee I, Humphreys K, et al. Predicting inadequate postoperative pain management in depressed patients: A machine learning approach. PLOS ONE. 2019;14(2):e0210575. [CrossRef]
  12. Sun Y, Yu K, Du L, et al. Application of XGBoost in the prediction of acute postoperative pain after major noncardiac surgery in older patients. Mol Pain. 2025;21:17448069251376199. [CrossRef]
  13. Nair AA, Velagapudi MA, Lang JA, et al. Machine learning approach to predict postoperative opioid requirements in ambulatory surgery patients. PLOS ONE. 2020;15(7):e0236833. [CrossRef]
  14. Kagerer C, Jauk S, Kramer D, et al. A Machine Learning-Based Risk Assessment Model for Poor Postoperative Pain Outcome. In: dHealth 2025. IOS Press; 2025:98-104. [CrossRef]
  15. Ryu G, Choi JM, Seok HS, et al. Machine learning based quantitative pain assessment for the perioperative period. Npj Digit Med. 2025;8(1):53. [CrossRef]
  16. Park I, Park JH, Yoon J, et al. Machine learning model of facial expression outperforms models using analgesia nociception index and vital signs to predict postoperative pain intensity: a pilot study. Korean J Anesthesiol. 2024;77(2):195-204. [CrossRef]
  17. Morisson L, Nadeau-Vallée M, Espitalier F, et al. Prediction of acute postoperative pain based on intraoperative nociception level (NOL) index values: the impact of machine learning-based analysis. J Clin Monit Comput. 2023;37(1):337-344. [CrossRef]
  18. Kumar S, Kesavan R, Sistla SC, et al. Predictive models for fentanyl dose requirement and postoperative pain using clinical and genetic factors in patients undergoing major breast surgery. Pain. 2023;164(6):1332-1339. [CrossRef]
  19. Fontaine D, Vielzeuf V, Genestier P, et al. Artificial intelligence to evaluate postoperative pain based on facial expression recognition. Eur J Pain. 2022;26(6):1282-1291. [CrossRef]
  20. Aydın Aİ, Özyazıcıoğlu N. Assessment of postoperative pain in children with computer assisted facial expression analysis. J Pediatr Nurs. 2023;71:60-65. [CrossRef]
  21. Morimoto M, Nawari A, Savic R, Marmor M. Exploring the Potential of a Smart Ring to Predict Postoperative Pain Outcomes in Orthopedic Surgery Patients. Sensors. 2024;24(15):5024. [CrossRef]
  22. Soley N, Speed TJ, Xie A, Taylor CO. Predicting Postoperative Pain and Opioid Use with Machine Learning Applied to Longitudinal Electronic Health Record and Wearable Data. Appl Clin Inform. 2024;15(03):569-582. [CrossRef]
  23. Salehinejad H, Muaddi H, Ubl DS, Sharma V, Thiels CA. Deep learning predicts postoperative opioids refills in a multi-institutional cohort of surgical patients. Surgery. 2024;176(2):246-251. [CrossRef]
  24. Klemt C, Harvey MJ, Robinson MG, Esposito JG, Yeo I, Kwon YM. Machine learning algorithms predict extended postoperative opioid use in primary total knee arthroplasty. Knee Surg Sports Traumatol Arthrosc. 2022;30(8):2573-2581. [CrossRef]
  25. Simpson S, Zhong W, Mehdipour S, et al. Classifying High-Risk Patients for Persistent Opioid Use After Major Spine Surgery: A Machine-Learning Approach. Anesth Analg. 2024;139(4):690-699. [CrossRef]
  26. Sikka K, Ahmed AA, Diaz D, et al. Automated Assessment of Children’s Postoperative Pain Using Computer Vision. Pediatrics. 2015;136(1):e124-e131. [CrossRef]
  27. Froicu (Armeanu) EM, Onicescu (Oniciuc) OM, Creangă-Murariu I, et al. Modeling Pain Dynamics and Opioid Response in Oncology Inpatients: A Retrospective Study with Application to AI-Guided Analgesic Strategies in Colorectal Cancer. Medicina (Mex). 2025;61(10):1741. [CrossRef]
  28. Lundberg S, Lee SI. A Unified Approach to Interpreting Model Predictions. arXiv. Preprint posted online November 25, 2017:arXiv:1705.07874. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.