Development and Internal Validation of a Perioperative Model for Predicting Early Postoperative Nausea and Vomiting After Gynecologic Laparoscopic Surgery.

Save

Background

Postoperative nausea and vomiting (PONV) is a frequent complication after gynecologic laparoscopic surgery. The classic Apfel score is widely used for risk assessment, but its individual-level discrimination may be limited in this relatively homogeneous high-risk population.

Methods

We conducted a single-center prospective observational model-development study of 487 patients undergoing gynecologic laparoscopic surgery under general anesthesia. Nineteen candidate predictors were considered. Exploratory univariable screening (P<0.10), followed by backward elimination using the Akaike information criterion, was used to derive the multivariable logistic model.

The blood-loss cut-point was selected by comparing candidate cut-points that retained at least 10% of patients in each group. Internal validation used 1000 bootstrap resamples in which the complete screening, cut-point selection, and backward-elimination procedure was repeated.

Results

Overall, 216 patients (44.4%) developed 0–24 h PONV. The final eight-predictor model had an apparent AUC of 0.706 and Brier score of 0.214. After bootstrap correction of the complete development procedure, the AUC was 0.666, the Brier score was 0.229, the calibration slope was 0.793, and the calibration intercept was −0.001.

The apparent AUC exceeded that of the Apfel score (0.586; paired DeLong difference 0.121, 95% CI 0.070–0.171; P<0.001).

Conclusion

The perioperative model showed moderate apparent discrimination, which decreased after correction for model-development optimism. The model is preliminary and requires prospective external validation, recalibration where necessary, and clinical-impact evaluation before routine use.

Introduction

Postoperative nausea and vomiting (PONV) remains one of the most common complications following anesthesia and surgery. Although generally not life-threatening, PONV can cause considerable patient discomfort, delay oral intake, prolong recovery, increase the need for rescue medication, and reduce satisfaction with perioperative care.1–3 Early PONV, particularly within the first 24 h after surgery, is clinically important as it directly affects recovery in the post-anesthesia care unit and the early postoperative ward period.

Patients undergoing gynecologic laparoscopic surgery are at relatively high risk for PONV. Female sex is a well-established risk factor, and laparoscopic surgery, anesthetic exposure, postoperative analgesic strategy, and opioid use may further increase this risk.1,4–7 Accurate risk assessment is therefore important for identifying patients who may benefit from more active prophylactic or rescue antiemetic treatment.

The Apfel score is among the most widely used clinical tools for PONV risk prediction.8 It comprises four simple factors: female sex, non-smoking status, history of motion sickness or PONV, and postoperative opioid use. Its primary advantage is simplicity, and it can be easily applied before or immediately after surgery. However, the performance of the Apfel score may be reduced in selected surgical populations or in patients receiving opioid-based patient-controlled analgesia.4,8,9 In gynecologic surgery, all patients are female, and most are non-smokers.

Consequently, many patients already possess two Apfel risk factors, leaving limited room for further risk stratification.

Moreover, the Apfel score does not incorporate many perioperative factors that become available during or shortly after surgery and may be clinically relevant for early PONV. For instance, postoperative PCIA reflects postoperative opioid exposure and is directly related to the postoperative opioid-use component of the Apfel score.1,6,7 Other factors, such as blood loss, intraoperative hypothermia, sedative use, anesthetic technique, and analgesic strategy, may also influence early PONV risk.1,10,11 Therefore, a model that integrates baseline susceptibility with perioperative information may improve risk prediction in this setting.

In this study, we evaluated the classic Apfel score as a preoperative benchmark and developed a perioperative model for postoperative risk reassessment immediately after surgery, once intraoperative exposures and the postoperative analgesia plan were available. The target outcome was PONV during the first 24 h after surgery. Because the proposed model uses later information than the Apfel score, their comparison was intended to quantify additional discrimination at a later prediction time point rather than to compare two tools available at the same clinical moment.

Materials and Methods

Study Design and Patient Population

This was a prospective observational study of patients who underwent gynecologic laparoscopic surgery under general anesthesia at our institution. Perioperative data were collected prospectively during routine clinical care. Patients were included if they had complete perioperative records and available documentation of PONV within the first 24 h after surgery.

Patients were excluded if they had missing key outcome data, incomplete anesthesia or medication records, conversion to non-laparoscopic surgery, or clinical conditions that rendered routine PONV assessment unreliable. The analyzed dataset comprised 487 patients, including 216 events, with complete values for all analyzed predictors and the outcome. All available eligible records formed the development sample.

With 216 events and 19 candidate predictor parameters, the dataset provided approximately 11 events per candidate parameter as a descriptive measure of sample size; this ratio was not treated as proof against overfitting.

The study protocol was approved by the Ethics Committee of Shantou Central Hospital (approval No.2024–047). Written informed consent was obtained from all participants before enrollment.

Outcome Definition

The primary outcome was 0–24 h PONV, defined as the occurrence of nausea, vomiting, or both within the first 24 h after surgery, based on postoperative nursing records, anesthesia follow-up records, and medication records. Patients were divided into the PONV group and the non-PONV group according to this endpoint.

Data Collection

Baseline clinical characteristics included age, body mass index (BMI), non-smoking status, history of motion sickness or PONV, migraine history, history of hyperemesis gravidarum, and ASA grade.

Anesthesia- and surgery-related variables included anesthesia maintenance method, anesthesia duration, surgical duration, blood loss, intraoperative hypothermia, and systolic and diastolic blood pressure fluctuations. Anesthesia was maintained with propofol-based total intravenous anesthesia in 302 patients (62.0%) and combined intravenous–inhalational anesthesia in 185 (38.0%); sevoflurane, the only volatile anesthetic used at our institution, was recorded in 184 (37.8%). Nitrous oxide was not used; carbon dioxide was used to establish the pneumoperitoneum.

Propofol and cisatracurium were recorded in all patients, and neuromuscular blockade was reversed with neostigmine combined with atropine. The recorded intraoperative antiemetic prophylaxis included tropisetron in 286 patients (58.7%), metoclopramide in 134 (27.5%), dexamethasone in 134 (27.5%), and methylprednisolone in 33 (6.8%). These drugs were administered intraoperatively for prophylaxis, but no uniform standardized antiemetic prophylaxis protocol was used and exact administration times were unavailable.

Recorded dezocine and midazolam doses also represented intraoperative administration and therefore preceded postoperative PONV surveillance; however, exact administration times and the time of first PONV onset were not collected.

The classic Apfel score was reconstructed from female sex, non-smoking status, history of motion sickness or previous PONV, and planned postoperative opioid exposure. All patients received one point for female sex. The score distribution was 1 in 3 patients, 2 in 172, 3 in 238, and 4 in 74; scores 1 and 2 were combined for category-based analyses because only three patients had a score of 1.

Postoperative PCIA supplied the postoperative-opioid component.

Development of the Perioperative Model

The performance of the classic Apfel model was first evaluated. We then developed a perioperative model using logistic regression. Logistic regression was selected because it is transparent, clinically interpretable, and suitable for binary clinical outcomes.12,13

The total Apfel score was not entered into the perioperative model. This decision was made because postoperative PCIA already represents postoperative opioid use, which is one of the four Apfel components. Including both the total Apfel score and postoperative PCIA would duplicate information and reduce model interpretability.

Therefore, the perioperative model was developed using clinically meaningful Apfel-related factors and perioperative variables rather than the total Apfel score.

Nineteen candidate predictors were considered: age, body mass index, non-smoking status, history of motion sickness or previous PONV, migraine history, history of hyperemesis gravidarum, ASA grade, anesthesia maintenance technique, anesthesia duration, surgical duration, blood loss, intraoperative hypothermia, systolic and diastolic blood pressure fluctuations, midazolam dose, dezocine dose, flurbiprofen axetil use, tropisetron use, and postoperative PCIA. Completeness, distribution, clinical interpretability, and redundancy were reviewed before modeling. Univariable logistic regression was used as an exploratory screening step, and variables with P<0.10 entered the initial multivariable model.

Backward elimination was then performed by sequentially removing the variable whose exclusion produced the greatest reduction in AIC until no further reduction occurred. This reproducible process removed tropisetron use, age, systolic blood pressure fluctuation, surgical duration, and anesthesia duration from the initial screened model.

Blood-loss cut-points corresponding to observed values were evaluated when at least 10% of patients remained in each group; the cut-point yielding the lowest AIC after the complete selection procedure was chosen. A threshold of >50 mL was selected in the original dataset. The final model included postoperative PCIA, history of hyperemesis gravidarum, blood loss >50 mL, flurbiprofen axetil use, dezocine dose per 5 mg, ASA grade per one-grade increase, intraoperative hypothermia, and midazolam dose per 1 mg.

The fitted equation was: logit(p) = −2.7398 + 0.6432 (PCIA) + 0.6805 (history of hyperemesis gravidarum) + 0.7924 (blood loss >50 mL) - 0.5241 (flurbiprofen axetil use) + 0.3913 (dezocine dose per 5 mg) + 0.9654 (ASA grade per one-grade increase) + 0.3271 (intraoperative hypothermia) - 0.2131 (midazolam dose in mg).

Model Evaluation

Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC). Because both models were evaluated in the same patients, their apparent AUCs were compared using the paired DeLong method.14 The AUC difference is reported with a paired 95% confidence interval. Sensitivity, specificity, positive predictive value, negative predictive value, and accuracy at the Youden-index cut-point were treated as descriptive, sample-dependent operating characteristics rather than a clinically optimal treatment threshold.

Calibration was assessed using calibration plots, calibration intercept, calibration slope, and the Brier score.15 Model fit was summarized using AIC and BIC. Decision curve analysis compared the perioperative model with the Apfel model, treat-all, and treat-none strategies across threshold probabilities from 0.05 to 0.70.16 Threshold analyses were exploratory and were not interpreted as evidence that model use improves patient outcomes.

Internal validation used 1000 bootstrap resamples.12 Within every resample, all 19 univariable screening tests, blood-loss cut-point evaluation, and backward AIC elimination steps were repeated. The resulting model was evaluated in the bootstrap sample and the original dataset; mean optimism was subtracted from apparent AUC, Brier score, calibration slope, and calibration intercept. Predictor-selection and blood-loss cut-point frequencies were also examined to assess model stability.

Statistical Analysis

Continuous variables were presented as median and interquartile range and compared using the Mann–Whitney U-test. Categorical variables were presented as number and percentage and compared using the chi-square test or Fisher exact test, as appropriate. Logistic regression coefficients, standard errors, odds ratios, and 95% confidence intervals are reported.

No predictor or outcome values were missing in the 487-patient analysis dataset. A two-sided P value <0.05 was used for inferential analyses; P<0.10 was used only for exploratory candidate screening. Analyses were performed using R-compatible logistic-regression, ROC, and bootstrap procedures.

Results

Patient Characteristics

A total of 487 patients were included in the final analysis. Among them, 216 patients developed PONV within the first 24 h after surgery, with an overall incidence of 44.4%. The median age of the entire cohort was 40.0 years, and the median BMI was 22.9 kg/m2 (Table 1).Table 1Baseline and Perioperative Characteristics According to 0–24 h PONV StatusVariableOverall (n=487)Non-PONV (n=271)PONV (n=216)P valueDemographics and clinical historyAge, years40.0 (33.0–48.0)39.0 (32.0–46.0)42.0 (34.0–49.0)0.011BMI, kg/m222.9 (21.2–25.0)23.1 (21.2–25.1)22.8 (21.2–24.7)0.464Non-smoking status480 (98.6)267 (98.5)213 (98.6)1.000Motion sickness/PONV history211 (43.3)112 (41.3)99 (45.8)0.366Migraine history129 (26.5)64 (23.6)65 (30.1)0.132History of hyperemesis gravidarum259 (53.2)123 (45.4)136 (63.0)<0.001ASA grade0.103 I14 (2.9)11 (4.1)3 (1.4) II460 (94.5)255 (94.1)205 (94.9) III13 (2.7)5 (1.8)8 (3.7)Anesthesia and surgical factorsAnesthesia maintenance0.646 TIVA302 (62.0)171 (63.1)131 (60.6) Combined intravenous-inhalational185 (38.0)100 (36.9)85 (39.4)Anesthesia duration, min200.0 (150.0–260.0)190.0 (140.0–250.0)212.5 (170.0–270.0)0.004Surgical duration, min155.0 (110.0–210.0)150.0 (100.0–200.0)165.0 (130.0–220.0)0.002Blood loss, mL50.0 (10.0–100.0)30.0 (10.0–50.0)50.0 (20.0–100.0)<0.001Intraoperative hypothermia119 (24.4)57 (21.0)62 (28.7)0.064SBP fluctuation, mmHg39.0 (29.0–50.0)37.0 (26.5–49.0)41.0 (31.0–52.0)0.005DBP fluctuation, mmHg29.0 (22.0–37.0)28.0 (21.0–36.0)31.0 (23.0–38.0)0.035Key medications and postoperative factorsMidazolam, mg0.0 (0.0–0.0)0.0 (0.0–0.0)0.0 (0.0–0.0)0.080Dezocine, mg0.0 (0.0–0.0)0.0 (0.0–0.0)0.0 (0.0–5.0)0.012Flurbiprofen axetil, mg50.0 (0.0–50.0)50.0 (50.0–50.0)50.0 (0.0–50.0)0.005Tropisetron, mg5.0 (0.0–5.0)5.0 (0.0–5.0)5.0 (0.0–5.0)0.010Postoperative PCIA179 (36.8)77 (28.4)102 (47.2)<0.001Classic Apfel score3.0 (2.0–3.0)3.0 (2.0–3.0)3.0 (2.0–3.0)<0.001Classic Apfel score category<0.001 Score <=2175 (35.9)111 (41.0)64 (29.6) Score 3238 (48.9)133 (49.1)105 (48.6) Score 474 (15.2)27 (10.0)47 (21.8)Notes: Data are presented as median (IQR) or n (%).

P values were calculated using Mann–Whitney U-test, chi-square test, or Fisher exact test as appropriate. Classic Apfel score components were female sex, non-smoking status, history of motion sickness or PONV, and postoperative opioid use. Because all patients underwent gynecologic surgery, female sex contributed one point for all patients.

Apfel score 1 was merged with score 2 because of the small sample size. Bold text indicates column headings, section headings, and overall P values for multi-category variables.

Compared with patients without PONV, those with PONV were slightly older (median age 42.0 vs 39.0 years; P=0.011). BMI was similar between the two groups. Nearly all patients were non-smokers, and the proportion of non-smoking status did not differ between groups.

A history of motion sickness or previous PONV was also comparable between groups. In contrast, a history of hyperemesis gravidarum was more frequent in patients with PONV than in those without (63.0% vs 45.4%, P<0.001; Table 1).

Surgical and anesthesia-related factors also differed between groups. Patients with PONV had longer anesthesia duration (median 212.5 vs 190.0 min; P=0.004) and longer surgical duration (median 165.0 vs 150.0 min; P=0.002). Blood loss was greater in the PONV group (P<0.001).

Systolic and diastolic blood pressure fluctuations were also more pronounced in the PONV group (Table 1).

Medication and postoperative analgesic strategy were associated with early PONV. Postoperative PCIA was used more frequently in patients with PONV than in those without (47.2% vs 28.4%, P<0.001). Dezocine dose was higher in the PONV group (P=0.012), whereas flurbiprofen axetil use was less frequent in the PONV group (P=0.005; Table 1).

Predictive Performance of the Classic Apfel Score

The classic Apfel score was significantly associated with 0–24 h PONV. The distribution of Apfel score categories differed between patients with and without PONV (P<0.001; Table 1). The observed incidence of PONV increased across Apfel score categories, with the highest incidence in patients with a score of 4 and the lowest in those with scores ≤2.

The trend was statistically significant (P for trend <0.001; Figure 1A).Figure 1Performance of the classic Apfel score for predicting 0–24 h postoperative nausea and vomiting (PONV). (A) Observed incidence of 0–24 h PONV according to classic Apfel score category. Apfel score 1 was merged with score 2 because of the small number of patients in the lowest-score category.

The incidence of PONV increased significantly with higher Apfel score category (P for trend < 0.001). (B) Receiver operating characteristic (ROC) curve of the classic Apfel model for predicting 0–24 h PONV. The area under the curve (AUC) was 0.586.

The blue dot represents the descriptive Youden-index cut-point (sensitivity 21.8%, specificity 90.0%).Bar chart & ROC curve of Apfel score vs post-op nausea/vomiting.Image A displays a bar chart with a significant trend (p less than 0.001). The x-axis is labeled ′Classic Apfel score category′ with categories less than or equal to 2, 3 and 4. The y-axis shows ′Observed 0 to 24 h PONV incidence (percent)′ from 0 to 80.

Bar values are 36.6 percent (64/175) for less than or equal to 2, 44.1 percent (105/238) for 3 and 63.5 percent (47/74) for 4. Image B features a receiver operating characteristic graph. The x-axis is ′1 - Specificity′ (0.0 to 1.0) and the y-axis is ′Sensitivity′ (0.0 to 1.0).

A diagonal reference line runs from (0.0, 0.0) to (1.0, 1.0). The Apfel model curve (AUC equal 0.586) starts at (0.0, 0.0), reaches a point near (0.1, 0.218) with Sensitivity equal 21.8 percent and Specificity equal 90.0 percent, then continues to (1.0, 1.0). The legend includes ′Apfel model (AUC equal 0.586)′ and ′Reference.′.

The apparent AUC of the correctly reconstructed Apfel score was 0.586 (95% CI, 0.541–0.630; Figure 1B). At the descriptive Youden cut-point of 4, sensitivity was 21.8% and specificity was 90.0%.

Development of the Perioperative Model

To improve prediction, we developed a perioperative logistic regression model. The model did not include the total Apfel score, as postoperative PCIA represents postoperative opioid use — one of the Apfel components. Instead, the model incorporated Apfel-related clinical information and perioperative variables.

The reproducible selection procedure retained eight variables: postoperative PCIA, history of hyperemesis gravidarum, blood loss >50 mL, flurbiprofen axetil use, dezocine dose, ASA grade, intraoperative hypothermia, and midazolam dose (Table 2 and Figure S1). The selected >50 mL blood-loss threshold had the lowest AIC among eligible observed cut-points in the original dataset.Table 2Univariable and Multivariable Logistic Regression for Predictors Included in the Perioperative ModelPredictorUnivariable OR (95% CI)P valueBetaSEMultivariable OR (95% CI)P valueIntercept——−2.73980.93970.065 (0.010–0.407)0.004Postoperative PCIA2.254 (1.548–3.282)<0.0010.64320.20521.903 (1.273–2.844)0.002History of hyperemesis gravidarum2.046 (1.420–2.948)<0.0010.68050.19951.975 (1.336–2.920)<0.001Blood loss >50 mL2.389 (1.601–3.565)<0.0010.79240.21942.209 (1.437–3.396)<0.001Flurbiprofen axetil use0.523 (0.349–0.783)0.002−0.52410.22330.592 (0.382–0.917)0.019Dezocine, per 5 mg1.683 (1.206–2.350)0.0020.39130.18701.479 (1.025–2.134)0.036ASA grade, per grade2.380 (1.037–5.461)0.0410.96540.45022.626 (1.087–6.346)0.032Intraoperative hypothermia1.512 (0.998–2.289)0.0510.32710.22911.387 (0.885–2.173)0.153Midazolam, per 1 mg0.812 (0.650–1.016)0.068−0.21310.11560.808 (0.644–1.013)0.065Notes: The intercept OR is shown for completeness but has no direct clinical interpretation. Continuous variables were scaled as indicated.Abbreviations: OR, odds ratio; CI, confidence interval; SE, standard error; PCIA, patient-controlled intravenous analgesia.

In multivariable analysis, postoperative PCIA was independently associated with a higher risk of 0–24 h PONV (OR 1.90, 95% CI 1.27–2.84; P=0.002). History of hyperemesis gravidarum was also associated with increased PONV risk (OR 1.97, 95% CI 1.34–2.92; P<0.001). Blood loss >50 mL was among the strongest predictors in the model (OR 2.21, 95% CI 1.44–3.40; P<0.001).

Higher dezocine dose was associated with increased PONV risk (per 5 mg increment: OR 1.48, 95% CI 1.03–2.13; P=0.036). Higher ASA grade was also associated with increased risk (OR 2.63, 95% CI 1.09–6.35; P=0.032; Table 2 and Figure S1).

Flurbiprofen axetil use was associated with lower predicted PONV risk, but this finding does not establish a direct antiemetic effect. Treatment was not randomized and may reflect analgesic strategy, procedure type, bleeding risk, renal function, pain severity, or clinician preference. The association should therefore be regarded as predictive and potentially confounded, not as evidence supporting a treatment recommendation.

A nomogram was generated from the apparent eight-variable logistic model (Figure 2). Because the bootstrap results indicated optimism and selection instability, the nomogram is presented as a research representation of the apparent model and should not be used for routine clinical decisions before external validation and recalibration.Figure 2Nomogram based on the perioperative model for predicting 0–24 h PONV. The nomogram was constructed using postoperative patient-controlled intravenous analgesia (PCIA), history of hyperemesis gravidarum, blood loss >50 mL, flurbiprofen axetil use, dezocine dose, American Society of Anesthesiologists (ASA) grade, intraoperative hypothermia, and midazolam dose.

Each predictor corresponds to a point value, and the total points are converted into the predicted probability of 0–24 h PONV.A nomogram predicting 0 to 24 h PONV risk using perioperative factors.The nomogram predicts 0 to 24 hour postoperative nausea and vomiting risk based on several perioperative factors. At the top, a scale labeled ′Points′ ranges from 0 to 100. Below, various predictors are listed with corresponding scales: ′Postoperative PCIA′ with ′No′ and ′Yes′, ′Hyperemesis gravidarum history′ with ′No′ and ′Yes′, ′Blood loss greater than 50 mL′ with ′No′ and ′Yes′, ′Flurbiprofen axetil use′ with ′Yes′ and ′No′, ′Dezocine dose′ in milligrams ranging from 0 to 15, ′ASA grade′ from 1 to 3, ′Intraoperative hypothermia′ with ′No′ and ′Yes′ and ′Midazolam dose′ in milligrams from 7.5 to 0.

The ′Total points′ scale ranges from 0 to 350. At the bottom, ′Predicted 0 to 24 h PONV risk′ is shown with a scale from 0.1 to 0.9. The nomogram is based on an optimized perioperative logistic regression model.

Comparison Between the Perioperative Model and the Classic Apfel Model

The perioperative model had an apparent AUC of 0.706, compared with 0.586 for the Apfel score (Figure 3A and Table 3). The paired DeLong AUC difference was 0.121 (95% CI, 0.070–0.171; P<0.001). This comparison reflects the additional information available at the postoperative reassessment time point and is not an equal-time comparison with the preoperative Apfel score.Table 3Performance Comparison and Internal Validation of the Perioperative ModelDomainMetricClassic Apfel ModelPerioperative ModelDiscriminationApparent AUC0.5860.706DiscriminationApparent AUC 95% CI0.541–0.6300.658–0.751Internal validationOptimism-corrected AUCNot applicable0.666Internal validationCorrected calibration slopeNot applicable0.793Internal validationCorrected calibration interceptNot applicable−0.001DiscriminationSensitivity at descriptive Youden cut-point21.8%58.3%DiscriminationSpecificity at descriptive Youden cut-point90.0%73.4%DiscriminationPPV63.5%63.6%DiscriminationNPV59.1%68.9%Overall performanceAccuracy59.8%66.7%Overall performanceApparent Brier score0.23990.2136Internal validationOptimism-corrected Brier scoreNot applicable0.2288Model fitAIC659.2617.8Model fitBIC667.6655.5Paired comparisonApparent AUC differenceReference0.121Paired comparison95% CI for AUC differenceReference0.070–0.171Paired comparisonPaired DeLong P valueReference<0.001Notes: The apparent AUCs were compared using the paired DeLong method.

Internal validation used 1000 bootstrap resamples that repeated univariable screening, blood-loss cut-point selection, and backward AIC elimination. Youden-index operating characteristics are descriptive and sample-dependent.Abbreviations: AUC, area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value; AIC, Akaike information criterion; BIC, Bayesian information criterion.Figure 3Comparison between the classic Apfel model and the perioperative model. (A) ROC curves comparing apparent model discrimination.

The perioperative model had an apparent AUC of 0.706 versus 0.586 for the Apfel score. The paired DeLong AUC difference was 0.121 (95% CI, 0.070–0.171; P<0.001). Complete-process bootstrap validation produced an optimism-corrected AUC of 0.666 for the perioperative model.

(B) Apparent calibration curves comparing predicted and observed probabilities. Complete-process bootstrap validation yielded a corrected Brier score of 0.2288, calibration slope of 0.793, and calibration intercept of −0.001 for the perioperative model. (C) Exploratory decision curve analysis comparing the perioperative model, the Apfel model, treat-all, and treat-none strategies across threshold probabilities from 0.05 to 0.70.

Net benefit does not establish improved clinical outcomes.Three plots comparing classic Apfel and perioperative models using ROC, calibration and decision curves.Image A: ROC curve with axes for 1-Specificity and Sensitivity (0.0-1.0). A diagonal reference line runs from (0.0, 0.0) to (1.0, 1.0). Legend: Classic Apfel model (AUC 0.586, 95 percent CI 0.541-0.630), Optimized perioperative model (AUC 0.707, 95 percent CI 0.658-0.751).

AUC difference: 0.121; P less than 0.001. Image B: Calibration curve with axes for Predicted and Observed probability (0.0-1.0). A diagonal line indicates perfect calibration.

Series: Classic Apfel model and Optimized perioperative model. Brier scores: Apfel 0.2399, Optimized 0.2136. Image C: Decision curve analysis with axes for Threshold probability (0.0-0.7) and Net benefit (-0.2 to 0.4).

Legend: Treat none, Treat all, Classic Apfel model, Optimized perioperative model. ′Treat none′ is horizontal at 0.0 net benefit. ′Treat all′ slopes down, crossing 0.0 near 0.45 threshold probability, reaching -0.2 near 0.7. Classic Apfel model declines from 0.4 near 0.0 to 0.0 near 0.6. Optimized model declines from 0.4 near 0.0 to 0.05 near 0.7.

At the descriptive Youden-index cut-point of 0.473, the perioperative model yielded sensitivity of 58.3% and specificity of 73.4%. The positive predictive value was 63.6%, the negative predictive value was 68.9%, and accuracy was 66.7% (Table 3). These sample-dependent values were not used to define a clinical treatment threshold.

The apparent Brier score was 0.2136 and the apparent AIC and BIC were 617.8 and 655.5, respectively. After 1000 bootstrap resamples that repeated the complete development procedure, the optimism-corrected AUC was 0.666, the corrected Brier score was 0.2288, the corrected calibration slope was 0.793, and the corrected calibration intercept was −0.001 (Table 3). Blood loss >50 mL was selected in 53.0% of bootstrap samples, while 80 mL was selected in 28.4%, indicating uncertainty in the dichotomization.

The apparent calibration plot suggested reasonable agreement between predicted and observed risks, but the optimism-corrected calibration slope of 0.793 indicated overfitting (Figure 3B and Table 3). Decision curve analysis suggested exploratory net benefit over portions of the evaluated 0.05–0.70 threshold range (Figure 3C); however, no clinical threshold was prespecified, and the analysis does not establish improved patient outcomes.

Discussion

In this single-center prospective observational study of 487 patients, the apparent perioperative-model AUC was 0.706, but it decreased to 0.666 after the complete predictor- and cut-point-selection process was repeated during bootstrap validation. The corrected calibration slope of 0.793 also indicated overfitting. These results support cautious further investigation rather than immediate clinical implementation.

The limited performance of the Apfel model in our cohort should not be interpreted as a failure of the Apfel score itself. The Apfel score was designed as a simple and practical screening tool and remains the most commonly used risk score for routine PONV assessment.1,8 Current guidelines and reviews continue to recommend baseline risk assessment to guide prophylaxis, while also emphasizing that prophylaxis should be adjusted according to anesthetic technique, opioid exposure, and patient-level risk.1,17,18 In our study, the observed PONV incidence increased across Apfel score categories, confirming that the score still captured meaningful baseline risk. However, the AUC of 0.586 indicates difficulty in distinguishing individual patients who would or would not develop early PONV.

This is clinically understandable: in gynecologic laparoscopic surgery, all patients are female and most are non-smokers. Thus, two major Apfel components were present in nearly all patients, leaving limited variation for risk separation. Similar concerns have been raised when traditional PONV scores are applied to relatively homogeneous high-risk surgical populations or to patients receiving opioid-based postoperative analgesia.4,7,9

This finding also highlights an important distinction between risk stratification and individualized prediction. A score may demonstrate a significant trend across risk groups while still exhibiting modest individual-level discrimination. In our cohort, the Apfel model identified some high-risk patients with high specificity, but its sensitivity was only 21.8%.

Clinically, this means that many patients who subsequently developed 0–24 h PONV would not have been classified as high risk by the classic score alone. For PONV prevention, this limitation is relevant because the primary clinical challenge is often the under-recognition of patients who may require more aggressive prophylaxis or early rescue therapy.

The perioperative model was therefore developed as a perioperative extension of the Apfel concept rather than as a replacement for it. We deliberately did not include the total Apfel score in the perioperative model because postoperative PCIA corresponds to postoperative opioid use, one of the four Apfel components. Including both the total Apfel score and PCIA would have counted the same clinical information twice and reduced interpretability.

Instead, we separated the risk information into clinically understandable domains: baseline emetic susceptibility, opioid and analgesic exposure, surgical burden, and intraoperative physiological stress. This structure enhances model interpretability and aligns more closely with real perioperative decision-making.

Among baseline factors, history of hyperemesis gravidarum was independently associated with 0–24 h PONV. Although not included in the classic Apfel score, this variable may reflect an individual tendency toward nausea and vomiting in response to hormonal, autonomic, or vestibular stimuli. Previous studies have also suggested that severe nausea and vomiting during pregnancy may be associated with higher postoperative emetic risk in female surgical patients.19 In a gynecologic population, this information is readily obtainable and may provide additional risk discrimination beyond a general history of motion sickness or previous PONV.

However, this finding should be validated externally, as pregnancy-related nausea history may be recorded differently across institutions.

Postoperative PCIA and higher dezocine dose were associated with higher predicted PONV risk. Because the study was observational and these exposures reflect clinical treatment decisions, the coefficients should not be interpreted as causal effects. PCIA and dezocine may also act as proxies for postoperative pain, procedural extent, or clinician selection.

Altering these treatments on the basis of the reported odds ratios cannot be assumed to reduce PONV.

In contrast, flurbiprofen axetil use was associated with lower PONV risk. This finding is plausible, as non-opioid analgesics can reduce opioid requirements, and multimodal analgesia is a core strategy for improving postoperative recovery.1,20 However, this association should be interpreted with caution. Because flurbiprofen axetil use was not randomly assigned in this observational study, its administration may have been influenced by surgical type, bleeding risk, renal function, physician preference, or anticipated postoperative pain.

Therefore, our results should not be interpreted as evidence that flurbiprofen axetil itself has a direct antiemetic effect. A more conservative interpretation is that opioid-sparing analgesic strategies may be associated with a lower burden of early PONV.

Blood loss >50 mL was associated with higher predicted risk and may represent surgical complexity, tissue manipulation, hemodynamic variation, or postoperative pain burden rather than a direct causal mechanism. The threshold was selected from the development data. Its bootstrap instability—50 mL was selected in 53.0% of samples and 80 mL in 28.4%—reinforces the need to evaluate a continuous or flexible blood-loss function in future external data.

ASA grade, intraoperative hypothermia, and midazolam dose also warrant careful interpretation. Higher ASA grade may reflect reduced physiological reserve and greater vulnerability to perioperative stress. Intraoperative hypothermia showed a trend toward higher PONV risk but did not reach statistical significance; it was retained because temperature instability may indicate longer surgery, greater exposure, or less stable perioperative physiology.

Midazolam showed a trend toward lower risk, but the confidence interval crossed unity. These variables should therefore be regarded as supportive predictors that contributed to overall model performance rather than as confirmed independent causal factors.

From a prediction-model perspective, discrimination was moderate in the development data and more modest after optimism correction. The corrected AUC of 0.666, Brier score of 0.2288, and calibration slope of 0.793 show that the magnitude of performance improvement should not be overstated. These results are compatible with a preliminary postoperative risk-reassessment model that requires external validation and possible recalibration.

Decision curve analysis was exploratory. Although the perioperative model showed higher net benefit than comparison strategies over portions of the evaluated threshold range, no treatment threshold was prespecified or linked to a validated benefit-harm trade-off. The analysis therefore suggests potential utility for future study but does not show that using the model would improve clinical outcomes or justify intensified treatment.

The intended prediction time is immediately after surgery, once intraoperative exposures and the planned postoperative analgesic strategy are known, and before prospective assessment of PONV during the subsequent 0–24 h period. The model is therefore a postoperative risk-reassessment approach, not a preoperative replacement for the Apfel score. Its predictors and coefficients may change in settings with different anesthesia, analgesia, or antiemetic practices.

Several limitations should be acknowledged. First, this was a single-center development study without external validation. Second, predictor screening, backward elimination, and blood-loss cut-point selection introduced optimism and instability despite bootstrap correction.

Third, although the antiemetics, dezocine, and midazolam included in the dataset were administered intraoperatively before postoperative PONV surveillance, their exact administration times and the time of first PONV onset were not collected. No uniform standardized antiemetic prophylaxis protocol was used. Fourth, postoperative pain intensity, rescue-antiemetic timing, and opioid-equivalent doses were not consistently captured.

Fifth, PCIA and medication variables reflect treatment decisions and may be confounded by pain severity and procedural complexity. Sixth, the model addresses only 0–24 h PONV in gynecologic laparoscopic surgery and should not be extrapolated to delayed or post-discharge PONV or other surgical populations. Prospective external validation, assessment of calibration and clinical impact, and evaluation of continuous predictor functions are required before implementation.

Conclusions

In conclusion, the perioperative model showed moderate apparent discrimination for 0–24 h PONV in this single-center gynecologic-laparoscopic cohort, but performance decreased after correction for the complete model-development process. The findings support further investigation of postoperative risk reassessment. Prospective external validation, recalibration where necessary, and clinical-impact assessment are required before the model can be recommended for routine practice.

The model should not be extrapolated to delayed or post-discharge PONV or other surgical populations without validation.

Related articles

PubMed

Incidence of Nausea and Vomiting in Patients With Cervical Cancer Undergoing Chemoradiotherapy With Steroid-Sparing.

Background/Aim Although dexamethasone (DEX) is essential for preventing chemotherapy-induced nausea and vomiting during concurrent chemoradiotherapy (CCRT) for cervical cancer, prolonged administration carries risks of adverse events. Therefore, we evaluated the effect of a shortened DEX duration (steroid-sparing therapy) on the total control (TC) of nausea and vomiting in patients with cervical cancer undergoing CCRT comprising cisplatin 40 mg/m2 to avoid the risk of adverse effects associated...

PubMed

Collagenous Gastritis Presenting as Chronic Nausea and Vomiting.

ABSTRACT Collagenous gastritis (CoG) is a rare disorder characterized by subepithelial collagen band deposition in the gastric mucosa. In this report, we describe a case of a 33-year-old woman presenting with chronic, episodic, postprandial nausea and vomiting later found to have biopsy proven CoG. She symptomatically improved following proton pump inhibitor and budesonide initiation and cessation of nonsteroidal anti-inflammatory drugs. This case underscores the insidious nature of CoG; it sho...