Nomogram for predicting opioid-induced nausea and vomiting for cancer pain patients

The study was retrospective. Between June 2020 and June 2022, patients diagnosed with cancer pain who were prescribed opioids due to solid cancer by palliative care specialists were recruited from 3 different hospitals in Tianjin Medical University General Hospital, The First Hospital of Qinhuangdao, and People's Hospital of Inner Mongolia Autonomous Region. Data for 416 patients were available. All patients were randomly divided into the training group (n = 293) and internal validation group (n = 123) according to the proportion of 7:3.

The inclusion criteria were as follows: (1) patients who used either weak or strong opioids; (2) ≥ 18 years old with malignant solid tumor; (3) patients opioid naïve or not. the exclusion criteria were as follows: (1) the patient has serious emotional or mental disorders; (2) the patients have experienced nausea, retching, or vomiting in the 24 h prior to opioid use; (3) the patient had symptomatic primary or metastatic centers nervous system malignancy; (4) the patient had nausea and vomiting caused by intestinal obstruction, electrolyte disturbance, and metabolic abnormalities vomit; (5) the patient has a history of glaucoma and recent episodes are frequent; (6) a patient can be included in the study only once.

OINV diagnostic criteria are defined according to Common Adverse Event Terminology Assessment Standard 5.0 (CTCAE.v5.0). Nausea or vomiting greater than or equal to grade 1 in the CTCAE.v5.0 criteria are defined as the occurrence of opioid-related nausea or vomiting.

The general information of all cases was recorded in detail according to the electronic case data and the contents of the questionnaire collected Data, including gender (male, female), age, tumor type (gastrointestinal cancer, non-gastrointestinal cancer), history of drinking, history of motion sickness, average sleep time, anxiety level, thought to have side effects, first-time use of opioids (or not), drug dose adjustment, presence of CINV in this or previous chemotherapy. It should be noted that the treatment in CINV refers to conventional chemotherapy.

In terms of data quality control, to ensure the scientific rigor and precision of the study, a stringent process will be followed to select research subjects based on inclusion and exclusion criteria. Prior to data collection, healthcare personnel will undergo professional training and adhere to a standardized set of guidelines to explain data entry requirements, thereby ensuring the quality of data collection. When encountering situations where it is unclear whether nausea and vomiting are caused by other factors, we will conduct a differential diagnosis of the factors causing nausea and vomiting, and it will only be included if the diagnosis is OINV. Unable to differentiate between opioid medications and other confounding factors causing nausea and vomiting will be excluded.

Firstly, in the training group, univariate logistic regression analysis was conducted on each variable to select variables with statistically significant differences. Subsequently, the variables with significant associations were incorporated into the multivariate logistic analysis to determine independent risk factors for OINV in patients. Finally, the selected independent risk factors were used to establish a nomogram for predicting the probability of OINV in patients. The nomogram translates each regression coefficient into a score ranging from 0 to 100. The variable with the lowest β coefficient is allocated a score of 100. The points associated with each independent variable are cumulatively summed to derive the overall point total, which is then transformed into predicted probabilities. (C-) index, also known as the Concordance Index, is a statistical measure commonly used to assess the predictive capability of a model. The (C-) index ranges from 0.5 (no discrimination, equivalent to random guessing) to 1.0 (perfect discrimination). The receiver operating characteristic (ROC) curve is a central metric used for evaluating the discriminative performance of predictive models [21]. It plots the true positive rate (sensitivity) against the false positive rate (1-specificity) at various classification thresholds. The area under the ROC curve (AUC) summarizes the overall performance of the model. AUC = 1 indicates perfect discrimination, 0.9 < AUC < 1 indicates excellent discrimination, 0.7 < AUC < 0.9 indicates a good discrimination, 0.5 < AUC < 0.7 indicates a poor discrimination, AUC = 0.5 is equivalent to random guessing, indicating no discrimination. In this study, the (C-)index and ROC curve were used to assess the nomogram’s predictive accuracy or discrimination. In order to enhance the clinical utility of the nomogram, optimal cut-off values for continuous variables were determined based on ROC curves and the Youden index [22]. A calibration plot was drawn to determine the consistency of the probabilities predicted by the nomogram and observation probabilities. Decision Curve Analysis (DCA) is a method used to evaluate the clinical utility or practical usefulness of a predictive model or diagnostic test [23]. It takes into account the benefits (true positives) and harms (false positives) of using a model across different threshold probabilities. So, the clinical benefits of the nomogram were estimated by the DCA.

Nomogram construction and validation were performed using the ‘rms’ and ‘car’ packages of R 4.1.1 software (https://​cran.​r-project.​org/​bin/​windows/​base/​old/​4.​1.​1/​R-4.​1.​1-win.​exe). In all statistical analyses, a level of P < 0.05 was adopted to determine statistical significance.

A total of 416 subjects were collected and analyzed in this study (293 for nomogram development and 123 for validation), as shown in Table 1. The demographic and clinical characteristics of the study population are shown in Table 1. There was no statistical difference in most variables between the two data sets due to random allocation by R software. Only the P value of the History of motion sickness in the two groups was less than 0.05. Importantly, in the training set, OINV was reported in 17.7% of patients (52/293). In the validation set, 22% of patients (27/123) experienced OINV.

We first determine the predictive value of each variable through logistic regression analysis. Table 2 describes the univariate logistic regression analysis of OINV. Unifactorial analysis gender (female), age (≤ 60 years old), history of motion sickness, average sleep time < 5 h at night, non-first-time use of opioids, presence of CINV in this or previous chemotherapy, drug dose adjustment and thought to have side effects were significantly associated with OINV. However, no statistically significant relationships (P > 0.05) were observed with a history of drinking, history of morning sickness (only female), anxiety, gastrointestinal cancer, or type of opioid.

Based on the analysis results, we conducted a multivariate logistic regression analysis, as shown in Table 3. History of motion sickness, average sleep time < 5 h at night, non-first-time use of opioids, presence of CINV in this or previous chemotherapy, and drug dose adjustment were significantly associated with OINV, which were used in the construction of the nomogram. Notably, age, gender, and gastrointestinal cancer which were important predictors of OINV in cancer pain patients [7, 16], showed no significant association with the occurrence of OINV in the study.

In order to identify whether there are any other significant risk factors, we conducted both univariate and multivariate logistic regression analyses separately for nausea and vomiting as outcomes in the overall population (Supplementary Table S1–4). Compared to these, the independent risk factors for OINV are more comprehensive.

We established a nomogram based on the logistic regression model developed using data from the training set. Significant risk factors identified by logistic regression analysis were fitted into the model. The final nomogram is shown in Fig. 1. Validation was first performed with the training set. Bootstrapping (1000 replications) was applied and a calibration curve was generated (Fig. 2A). There was no obvious deviation between the risk curve predicted by the model and the actual observed risk, indicating that the model is well-calibrated. The C-index for OINV prediction was 0.839 (95% CI, 0.832–0.845) in this nomogram, the area under the ROC curve (AUC) is 0.839 (95% CI, 0.780–0.897) (Fig. 2C), indicating a good performance of OINV prediction using the combined nomograms. Calculating the Youden index using the ROC curve to define the cut-off threshold was 136, corresponding to 80.8% sensitivity, and 71.4% specificity. When the total score is below 136, patients are less likely to be diagnosed with OINV.

The validation set data of patients (n = 123) were used for model validation. The calibration plot for the probability of OINV revealed good concordance between the nomogram prediction and actual observations (Fig. 2B). The C-index for OINV prediction was 0.802 (95% CI, 0.786–0.819) and the AUC was 0.802 (95% CI, 0.708–0.897), demonstrating that the nomogram was well-fitted (Fig. 2D).

The decision curve analysis (DCA) was subsequently performed to illustrate the clinical decision utility of the combined nomogram. As shown in Fig. 3A, the threshold probabilities for the train set range from 5 to 92%, and the screening strategy based on our nomogram OINV risk estimate results in a superior net benefit compared to the treat-none or treat-all strategies. In the validation set, the threshold probability is between 15 and 78% (Fig. 3B). Consequently, DCA reveals that the nomogram has a superior overall net benefit over a wide range of practical threshold probabilities, indicating a high potential clinical utility.​