Postoperative Opioid Use Prediction Calculator Performance in Patients Undergoing Gynecologic Surgery

This prognostic study was approved by the Duke University Institutional Review Board, which considered this quality improvement study as exempt, granting a waiver of patient consent. The study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis Plus Artificial Intelligence ( TRIPOD+AI ) reporting guideline. 13 The original GO-POP uses 7 variables (preoperative patient-reported anxiety, anticipated opioid need, age, operating time, pregabalin use, smoking status, and education level). In response to feedback from GO-POP end users and potential users, we updated the calculator to facilitate preoperative prescribing, which can simplify the postoperative course for patients by enabling them to pick up prescriptions before surgery. Additionally, pregabalin had been removed from our institution’s enhanced recovery after surgery protocol and, although collected for this study, was expected to be removed from GO-POP. Patients aged 18 years or older undergoing scheduled gynecologic surgery and able to read and speak English were prospectively identified from November 1, 2023, to November 1, 2024, from the following surgical specialties within a single academic health system: general gynecology, gynecologic oncology, minimally invasive gynecologic surgery (MIGS), and urogynecology. Patient identification continued by specialty or subspecialty until survey and opioid use data were obtained for approximately 25% of the total cohort. This process ensured that each specialty and subspecialty was represented, as different subspecialties reflect different patient populations that vary in age, education, surgical indication, and mental health comorbidities, all of which contribute to a person’s pain experience. Patients undergoing only minor procedures, such as pelvic examinations under anesthesia, intrauterine device insertions, hysteroscopies, and loop electrosurgical excision procedures, were excluded. Enhanced recovery after surgery protocols are the standard of care at the authors’ institution and include preoperative teaching about postoperative expectations, though pain management recommendations may have varied. Clinical information was abstracted from the electronic health record and stored in a Research Electronic Data Capture database hosted at Duke University. 14 Patients were identified by reviewing surgeons’ clinical and operative schedules. Variables previously identified as predictors of opioid prescribing were collected. Route of surgery, malignant neoplasm, history of preoperative opioid use, and indication for surgery were abstracted. These variables were evaluated during the original model development and found not to be key predictors of outpatient opioid use; therefore, they were not included. Specifically, operative time was found to be a better predictor than route of surgery. 8 Preoperative patient-reported anxiety was assessed by asking, “On a scale from 0 to 100, how anxious are you regarding your upcoming surgery?” Preoperative anticipated opioid need was measured by asking, “Do you anticipate requiring below average, average, or above average amount of pain medication following surgery?” 15 Additional candidate predictors collected were patient age at the preoperative visit and the surgeon’s anticipated operative time. Patients received postoperative email surveys to assess the number of opioid pills used, the average pain score over the past week, and the pain score at the time of the survey. These surveys occurred weekly until patients reported no longer using opioids for postoperative pain or until 6 weeks after surgery. For patients with preoperative opioid use, surveys continued until they returned to baseline use or until 6 weeks after surgery. Preoperative opioid use was determined by reviewing patients’ medical histories and medications in the electronic health record. If patients did not respond to the email survey, they were called to obtain responses based on clinician and patient availability. Missing predictor values, including educational attainment, preoperative anxiety, and anticipated opioid use, were imputed using the aregImpute function in the R package Hmisc, version 5.2-6. 16 We generated 71 imputed datasets. The number of imputations was determined using the R package howManyImputations, version 0.2.5. Data were analyzed from January 15, 2025, through July 27, 2026. The original model was externally validated in the expansion cohort without refitting. The prespecified updated proportional odds model included anticipated pain medication use, restricted cubic splines for age at the preoperative visit (3 knots) and anticipated operating time (5 knots), educational attainment, smoking status, and anxiety; 71 imputation-specific models were fitted. The full model fitting and prediction workflow was internally validated using 500 participant-level bootstrap repetitions. Optimism was defined as the difference between bootstrap apparent performance and performance on the original sample, and it was subtracted from the original apparent performance. Model fitting was performed using the R package rms, version 8.1-1. 17 The outcome was the number of opioid pills used after discharge. Both models used proportional odds ordinal logistic regression. Performance was evaluated separately at clinically relevant binary thresholds of at least 2, 5, and 10 pills. At each threshold, discrimination was quantified using the binary C statistic (area under the receiver operating characteristic curve), and overall prediction error was quantified using the Brier score. The scaled Brier score was calculated separately at each threshold as 1 minus the Brier score divided by the null Brier score, where the null Brier score was prevalence multiplied by 1 minus prevalence. 18 Calibration intercepts and slopes were estimated using logistic regression of each observed binary threshold outcome on the logit of the pooled predicted probability, without binning. Grouped observed proportions were used only for graphical display. For external validation of the original model, 95% CIs were obtained from 2000 participant-level bootstrap resamples, with pooled predictions held fixed. For the updated model, apparent and optimism-corrected calibration curves were derived from the 500-repetition internal validation procedure. Predicted probabilities were averaged across the 71 imputation-specific models; no postfit, threshold-specific recalibration was applied. The 95% CIs for optimism-corrected performance were obtained by a nested bootstrap in which the entire model-fitting and optimism-correction procedure was repeated within 500 outer participant-level resamples, with bias-corrected intervals computed from the resulting distribution. Model performance was stratified by specialty to assess potential biases that may arise from specialty-specific prescribing practices. Model performance was also stratified by patient self-reported race, in accordance with current reporting guidelines, to ensure no overperformance or underperformance by racial group, even though race was not a predictor in the model. 13 For the analysis, race and ethnicity were grouped as African American or Black; Hispanic or non-Hispanic; White; and other (American Indian or Alaska Native, Asian, Native Hawaiian or Pacific Islander, multiracial, preferred not to answer [grouped due to low numbers]). Secondary outcomes included gynecologic oncology patients whose opioid prescriptions were guided by GO-POP, average pain score over the past 7 days, and pain score at the time of the survey. We also included the number of patients with at least 1 refill request, clinician visit, phone call, and urgent care or emergency department visit for postoperative pain. Secondary outcomes were evaluated until patients were no longer using opioids for postoperative pain or until 6 weeks after surgery. Variables were collected for the overall cohort to assess differences between patients with and without opioid use data. The sample size required for development of an updated prediction model was calculated using criteria proposed by Rodriguez et al 8 and Riley et al 19 and information established from previous GO-POP model fittings. Using a mean (SD) outcome of 3 (11) pills, an R 2 statistic from the existing model of 0.19, and a global shrinkage factor of 0.9, 589 participants were required to consider up to 16 parameters in the original model. Therefore, we planned to include at least 775 patients, assuming that approximately 25% would be lost to follow-up at postoperative assessment. The statistical analyses were performed using R, version 4.6.0 (R Foundation for Statistical Computing). No hypothesis tests were performed. Results are reported as point estimates with 2-sided 95% CIs, with null values of 0.5 for the C statistic, 0 for the calibration intercept, and 1 for the calibration slope.

A total of 884 patients scheduled for gynecologic surgery were prospectively identified. Eight duplicate records, 3 patients younger than 18 years, and 3 patients undergoing only minor procedures were excluded, leaving 870 eligible patients. Thirty-five patients had no recorded actual operating time because surgery was canceled or delayed; 835 underwent an operation. Among patients who underwent surgery, 219 lacked the opioid use outcome, and 616 were included in the analysis (median [IQR] age, 48 [39-65] years; 179 identifying as African American or Black [29.1%], 376 as White [61.0%], and 61 as other [9.9%] race; 38 identifying as Hispanic [6.2%] and 553 as non-Hispanic [89.8%] ethnicity) ( Figure 1 ; Table 1 ). Baseline characteristics of patients with and without opioid use follow-up were similar. The analysis cohort included 136 patients (22.1%) from general gynecology, 154 (25.0%) from gynecologic oncology, 163 (26.5%) from MIGS, and 163 (26.5%) from urogynecology ( Table 2 ). Among the 616 patients with opioid use outcomes, 7692 pills were prescribed, and 2991 were used. Among 579 patients with an opioid type recorded, 554 (95.7%) were prescribed oxycodone at the preoperative visit, discharge, or both. a Missing values were addressed using multiple imputation (71 imputed datasets); no participants were excluded. Abbreviations: BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); EIN, endometrial intraepithelial neoplasia; NOS, not otherwise specified. Other race included American Indian or Alaska Native, Asian, Native Hawaiian or Pacific Islander, multiracial, and preferred not to answer, which were grouped because no individual category included sufficient numbers of patients to estimate threshold-specific discrimination. Included trade school, unknown, or not answered. Included risk-reducing surgery for hereditary cancer predisposition or family history ( BRCA1 , BRCA2 , Lynch syndrome, PALB2 , CHEK2 , or RAD51C ; vulvar, vaginal, or cervical dysplasia or carcinoma; oophorectomy or surgical menopause for breast cancer; adenomyosis, endometrial hyperplasia, or endometrial mass; gender-affirming surgery; infertility, fertility-preserving, or obstetric indications; peritoneal malignancy or borderline ovarian tumor; pelvic floor or urologic conditions; and other benign conditions. Abbreviation: MIGS, minimally invasive gynecologic surgery. Anticipated surgical approaches were check box responses and were not mutually exclusive. In the outcome cohort, 575 participants selected 1 approach, 39 selected 2 approaches, and 2 selected 3 approaches; in the overall cohort, 1 selected 0 approaches, 806 selected 1 approach, 61 selected 2 approaches, and 2 selected 3 approaches. Most often mini-laparotomy myomectomy, with the remainder including hysteroscopic procedures, diagnostic laparoscopy, laparoscopic myomectomy or ovarian cystectomy, and excision of abdominal wall endometriosis. Missing data represent patients whose surgery was canceled or delayed beyond the study period. The GO-POP guided prescribing for 101 gynecologic oncology patients. The median total number of perioperative pills prescribed for outpatient use after discharge was 10 (IQR, 8-15 pills) ( Table 3 ). The median number of pills used after discharge was 2 (IQR, 0-7 pills), and 226 patients (36.7%) used no pills. Median use was 3 pills (IQR, 0-10 pills) in general gynecology, 1 pill (IQR, 0-6 pills) in gynecologic oncology, 5 pills (IQR, 0-10 pills) in MIGS, and 1 pill (IQR, 0-4 pills) in urogynecology. Median use was 5 pills (IQR, 0-10 pills) among patients with an anticipated open approach and 2 pills (IQR, 0-7 pills) among those without an anticipated open approach. Abbreviations: ED, emergency department; MIGS, minimally invasive gynecologic surgery; NA, not applicable. Represents total pills prescribed preoperatively and postoperatively for outpatient use after discharge. Missing data represent patients who were scheduled for surgery, but surgery was canceled or delayed beyond the study period. For the original GO-POP model, threshold-specific C statistics were 0.70 (95% CI, 0.66-0.74), 0.72 (95% CI, 0.68-0.76), and 0.74 (95% CI, 0.70-0.78) for use of at least 2, 5, and 10 pills, respectively. Calibration intercepts were −0.11 (95% CI, −0.30 to 0.08), −0.35 (95% CI, −0.54 to −0.17), and −0.51 (95% CI, −0.71 to −0.31), respectively. Thus, the original model overpredicted probabilities at thresholds of 5 or more and 10 or more pills, whereas calibration at the threshold of 2 or more pills was compatible with ideal calibration (eFigure and eTable 1 in Supplement 1 ). Threshold-specific discrimination varied across surgical specialties from 2 or more to 10 or more pills, respectively (general gynecology: C statistic range, 0.63 [95% CI, 0.53-0.73] to 0.70 [95% CI, 0.60-0.79]; gynecologic oncology: C statistic range, 0.69 [95% CI, 0.60-0.77] to 0.70 [95% CI, 0.59-0.80]; MIGS: C statistic range, 0.62 [95% CI, 0.52-0.71] to 0.63 [95% CI, 0.54-0.71]; urogynecology: C statistic range, 0.71 [95% CI, 0.62-0.78] to 0.80 [95% CI, 0.67-0.91]). Threshold-specific discrimination also varied across racial groups (African American or Black: C statistic range, 0.65 [95% CI, 0.56-0.73] for ≥2 pills to 0.65 [95% CI, 0.57-0.73] for ≥10 pills; White: C statistic range, 0.71 [95% CI, 0.66-0.76] for ≥2 pills to 0.77 [95% CI, 0.70-0.83] for ≥10 pills; other: C statistic range, 0.59 [95% CI, 0.45-0.73] for ≥2 pills to 0.68 [95% CI, 0.54-0.82] for ≥5 pills) (eTable 2 in Supplement 1 ). For the updated preoperative model, optimism-corrected C statistics were 0.71 (95% CI, 0.66-0.74), 0.73 (95% CI, 0.68-0.76), and 0.73 (95% CI, 0.67-0.77) for use of at least 2, 5, and 10 pills, respectively. Corresponding Brier scores were 0.21 (95% CI, 0.21-0.23), 0.20 (95% CI, 0.19-0.22), and 0.15 (95% CI, 0.14-0.17), respectively, and scaled Brier scores were 13.3% (95% CI, 6.1%-16.7%), 14.1% (95% CI, 6.7%-17.1%), and 10.4% (95% CI, 3.0%-13.7%), respectively. Figure 2 shows apparent and optimism-corrected logistic calibration of the updated model. At the thresholds of at least 2, 5, and 10 pills, optimism-corrected calibration intercepts were −0.05 (95% CI, −0.09 to −0.01), −0.04 (95% CI, −0.08 to 0.002), and −0.03 (95% CI, −0.07 to 0.01), and slopes were 0.93 (95% CI, 0.80-1.02), 1.06 (95% CI, 0.89-1.19), and 1.15 (95% CI, 0.85-1.37), respectively. No postfit, threshold-specific recalibration was applied. Predicted probabilities were calculated within each of 71 imputation-specific proportional odds models and pooled by taking their arithmetic mean. Points and bars indicate observed proportions and exact binomial 95% CIs in 10 approximately equal-sized groups; rug marks show the distribution of predicted probabilities. Optimism-corrected calibration intercepts and slopes were estimated from 500 participant-level bootstrap repetitions. No postfit, threshold-specific recalibration was applied. The median pain score over the preceding 7 days at week 1 was 4 (IQR, 3-6), and the median pain score at contact was 2 (IQR, 1-4). Pain-related health care use was descriptively comparable in the opioid use outcome and overall cohorts. The respective counts were 70 patients (11.4%) and 110 patients (12.6%) for at least 1 telephone call, 23 (3.7%) and 43 (4.9%) for at least 1 opioid refill, 18 (2.9%) and 33 (3.8%) for at least 1 postoperative visit, and 2 (0.3%) and 9 (1.0%) for at least 1 urgent care or emergency department visit. No patients had a pain-related readmission ( Table 3 ).

This prognostic study found that the original opioid use prediction model showed moderate threshold-specific discrimination in a broader gynecologic surgical population but overpredicted use at the higher thresholds of pill use. The prespecified updated model relied only on preoperative characteristics and showed moderate threshold-specific discrimination after internal validation, with calibration intercepts near 0 and slopes from 0.93 to 1.15. The updated calculator will be made publicly available in the Duke Health Risk Calculator Library. 20 We recommend external institutional validation before broad implementation. Multiple efforts have been made to predict opioid use among gynecology patients. Wong et al 10 developed the Postoperative Opioid Calculator for Hysterectomy, which relies on 9 preoperative questions to predict high (>15 tablets of oxycodone) and low (<5 tablets of oxycodone) opioid use. 10 Straubhar and colleagues developed a simple algorithm based on 7 factors to predict postoperative opioid use in gynecologic oncology patients, and it showed high prescriber compliance and a lower number of excess pills prescribed. 11 , 12 Our study expands the previously developed, validated, and implemented GO-POP for use in a gynecologic oncology division at an academic health center to apply to gynecology specialty and subspecialty patients alike. 8 , 9 The original GO-POP used rigorous methods to predict the number of outpatient opioid pills a patient would use. When implemented, the original calculator was associated with fewer opioid pills prescribed without worse patient-reported pain control or more health care encounters for pain. In this study, the original model showed moderate discrimination across a broad range of gynecologic surgeries, with miscalibration at higher thresholds of pill use. The updated model uses 6 preoperative predictors and showed moderate threshold-specific discrimination after internal validation. Among the 616 patients with opioid use outcomes, 7692 pills were prescribed and 2991 were used, leaving a mean (SD) of 7.6 (6.6) unused pills per patient. In 2003, approximately 600 000 hysterectomies were performed in the US. 21 At that annual volume, overprescribing 5 opioid pills per patient would leave approximately 3 million unused pills in the community. This estimate does not include other gynecologic procedures, such as myomectomies and laparoscopies. Model estimates may support, but should not dictate, prescribing; clinicians should continue to use clinical judgment and shared decision-making. Education and strategies for opioid disposal should also be explored. Improving both prescribing and disposal practices may support pain management and opioid stewardship. This study had several strengths. First, it built on an existing, validated, and implemented postoperative prediction calculator. Previous evaluations showed a reduction in postoperative opioid prescriptions without an increase in health care encounters for pain. 8 , 9 This study prospectively extends generalizability to both specialty and subspecialty patients undergoing major surgery. Second, the modified model incorporated feedback to support preoperative prescribing. Third, the calculators relied on easily obtainable variables to predict the number of opioids a patient would use after surgery. This study also had several limitations. First, it carried inherent limitations associated with surveys. We achieved postoperative contact with approximately 70% of patients, which is consistent with response rates in the literature 22 but this could have influenced the actual number of opioid pills used and our ability to assess pain scores after surgery. Additionally, as our data relied on self-report, some patients may not have honestly reported opioid use. Challenges in assessing medication adherence are rife in the literature. 23 Second, although inclusion criteria were broad, this study was limited to patients who read and speak English within a single quaternary academic health system. It was not feasible to obtain an interpreter for all points of contact required for this study. Although pain physiology should not differ by race and ethnicity, predictive models should be validated in populations that do not speak English. Moreover, while the catchment area involves multiple surrounding states in the southeastern US, there may be regional differences surrounding opioid use and pain management. Third, while inclusive of all subspecialties, this study was not powered to conclusively determine distinct patterns across surgical subspecialties, and further research is needed to determine whether these differences are reproducible. Fourth, most patients were prescribed oxycodone; therefore, this calculator is most predictive of this type of opioid. Fifth, most patients underwent minimally invasive surgery for benign indications, which may limit generalizability to patients who undergo laparotomy for cancer. Route-specific model performance and associations between operative route and opioid use were not evaluated in this study. Finally, while this study provides temporal validation data for the original GO-POP model in a single institution, the updated preoperative model has only been validated within a single health system and requires external institutional validation.

This prognostic study found that the original opioid use prediction calculator showed moderate threshold-specific discrimination but overpredicted use at the higher thresholds of pill use in the expansion cohort. The updated model showed moderate discrimination after internal validation, with calibration intercepts near 0 and slopes from 0.93 to 1.15. External validation is required before broad implementation. Individualized model estimates may support clinical judgment, shared decision-making, and opioid stewardship.

In 2022, an estimated 4.6 million adult women in the US needed treatment for opioid use disorder. 1 Prescription opioid exposure can lead to persistent use, and surgery is linked to an increased risk of chronic opioid use. 2 , 3 , 4 , 5 Although overprescribing contributes to the opioid epidemic, surgeons may prescribe more tablets than patients use after surgery. 3 , 6 Contributing factors include a desire to ensure adequate pain control, efforts to avoid refill requests, and a lack of procedure-specific guidance. 7 The postoperative period is therefore an important opportunity to limit initial opioid exposure. The Gynecologic Oncology Postoperative Opioid Use Predictive Calculator (GO-POP) was developed and internally validated in a gynecologic oncology patient population at an academic health center. 8 This predictive model estimates the number of opioid pills a patient would use after discharge following surgery. 8 Implementation of GO-POP decreased the number of opioid pills prescribed by 33% without worsening patient-reported pain control or increasing health care encounters for pain management. 9 The GO-POP is now the standard of care within the division of gynecologic oncology at Duke University Health System. Others have also developed postoperative opioid use prediction calculators following gynecologic surgery. Wong et al 10 developed the Postoperative Opioid Calculator for Hysterectomy for patients undergoing minimally invasive hysterectomy, and Straubhar and colleagues developed an algorithm in a gynecologic oncology patient population. 11 , 12 These algorithms are limited by their scope of use as neither has been validated outside the specific patient population or surgical route for which they were developed. The primary aim of our study was to expand and validate GO-POP in a broader gynecologic surgical population, including patients undergoing surgery with gynecologists, gynecologic oncologists, minimally invasive gynecologic surgeons, and urogynecologists. Furthermore, because many surgeons prescribe pain medications preoperatively, the secondary aim was to assess whether GO-POP could maintain accuracy when using only preoperative characteristics.