Predictors of 72-h unscheduled return visits with admission in patients presenting to the emergency department with abdominal pain

We extracted the patient profiles from the administrative database of the Emergency Department of Taipei Medical University Hospital (TMUH-ED), which is a university hospital and level I trauma center. The TMUH-ED database preserves copies of all patients’ medical records in electronic form. This study was approved by the Joint Institutional Review Board of Taipei Medical University (approval number: N202103017). Informed consent was waived because all records were either anonymous or de-identified.

We consecutively enrolled patients aged ≥ 20 years who visited the ED and then revisited the ED within 72 h of the index visit during the period between January 1, 2014, and December 31, 2018. To identify a subgroup of patients with URVs related to abdominal pain, we first excluded patients whose diagnoses based on International Classification of Diseases, Tenth Revision (ICD-10) codes during the revisit were not related to abdominal pain. Second, we excluded patients with pelvic diseases such as pelvic inflammatory disease or pregnancy-related problems during the revisit. Subsequently, abdominal-pain-related symptoms and diagnoses made on the basis of ICD-10 codes during the index visit were confirmed by Hui-An Lin and Li-Tsung Lin. Lastly, we excluded patients who were discharged against medical advice or those who had been transferred to the ED from other departments during either the initial or second visits.

The selected variables were broadly classified into 2 categories: patient-related and system-related variables. The patient-related variables comprised sex, age (20–34, 35–49, 50–64, ≥ 65 year), triage score (Levels 1–5), vital signs (body temperature and pulse rate), pain score (based on numeric rating system: 0–3, 4–7, and 8–10), blood pressure (systolic blood pressure [SBP] and diastolic blood pressure [DBP]), tenderness or rebound tenderness, potentially immunocompromised status (cancer, diabetes mellitus, chronic kidney disease), or history of abdominal surgery. The system-related variables comprised duration of ED stay (0–2, ≥ 2 h), number of analgesics used, performance of initial examination (order of laboratory tests and imaging procedures), and history of TMUH visits.

The Taiwan Triage and Acuity Scale [20], which is widely used at hospitals in Taiwan, classifies patients in the ED into 5 levels on the basis of the acuity of illness: Level 1 (resuscitation), Level 2 (emergent), Level 3 (urgent), Level 4 (less urgent), and Level 5 (not urgent). Most patients with abdominal pain were triaged as a level 3; however, the patients who had associated symptoms and signs of severe pain (pain score of 8–10), SBP of < 90 mmHg, pulse rate of > 140 or < 50 beat/min, and fever up to 39 ℃ were escalated to the higher levels of 1 or 2. Accordingly, patients with abdominal pain were categorized into two groups (levels of 1–2 and 3–5, respectively). In this study, we considered the number of analgesics used, regardless of the route of administration, as an indicator of the need for pain control. We also included the decision for laboratory examination or imaging surveys—made by emergency physicians on the basis of their clinical judgments—as an independent factor in our study. These levels, along with system-related variables, were considered to indicate the severity of illness and degree of pain control required for the patients in our study cohort.

We compared URVA and URVNA groups by applying the Chi-square test for categorical variables and the Mann–Whitney U test for continuous variables. Thereafter, univariate and multivariate logistic regression analyses were performed to obtain odds ratios and to identify the significant variables. In the multivariate analysis, we adopted two strategies for the explanatory variables: (1) the first strategy was by putting all clinical meaningful variables which showed statistically significant in univariable analysis (in our models 1 and 2). (2) the second strategy was by selecting the variables through backward elimination (in our models 3 and 4). In addition, the Hosmer–Lemeshow statistics were used to examine the goodness of fit for logistic regression models, and the Akaike information criterion (AIC) was used to determine which model had the best fit for the data. The model with the lowest value of AIC was the most preferred model. We used receiver operating characteristic (ROC) curves to examine the correlation between the identified significant variables and URVAs. The Youden’s index (sensitivity + specificity—1) was used to obtain the optimal cut-off value for age. As a validation analysis, a classification and regression tree (CART)—a predictive model that applies a dichotomous decision process and splits data according to a certain cut-off value—was used to establish clinical guidelines for identifying URVAs. In the CART model, all significant variables in the logistic regression analysis were included and were determined which stratification to conduct at every stage. All statistical analyses were performed using R software 4.1 (R Core Team, 2021) and SAS 9.4 (SAS Institute, Cary, NC, USA). The statistical significance was defined as p < 0.05.

During the 5-year study period, a total of 6829 patients with URVs were identified in our database; of these patients, 702 met the study inclusion criteria. Among the 702 patients with URVs, 249 (35.5%) and 453 (64.5%) had URVAs and URVNAs, respectively. The flowchart of the study procedures, including the patient enrollment and outcome analysis procedures is displayed in Fig. 1. The detailed information of included ICD-10 codes included in present study is demonstrated in Additional file 1: Table S1.

The characteristics of the study population are shown in Table 1. The median age of the 702 patients included in this study was 47 years (in a range of 20 to 103 years), and 42.9% of the included patients were men. The majority of the patients’ visits were not urgent (triage Levels 3–5; 93.4%). The laboratory test and imaging execution rates were 61.0% and 58.4%, respectively. A comparison of the URVA and URVNA groups revealed that the URVA group was older (median age: 52 vs 42 year; P < 0.001), had a greater history of prior hospital visits (94.0% vs 87.9%; P = 0.010), had a longer ED duration (median hours: 2.2 vs 1.5 h; P = 0.001), had higher triage scores (Levels 1–2: 10.0% vs 4.6%; P = 0.006), had higher laboratory test execution rates (81.1% vs 49.9%; P < 0.001), had higher imaging execution rates (71.5% vs 51.2%; P < 0.001), and was prescribed more analgesics (≥ 2: 11.2% vs 4.2%; P < 0.001). Moreover, a comparison of the URVA and URVNA groups indicated that the URVA group exhibited lower SBP levels (SBP ≥ 130 mmHg: 41.0% vs 53.2%; P = 0.002) and more comorbidities (history of cancer: 16.1% vs 8.4%, P = 0.002; history of abdominal surgery: 30.5% vs 19.9%, P = 0.001; and history of diabetes mellitus: 15.3% vs 7.9%, P = 0.003). The 2 groups did not differ significantly in terms of other variables, namely gender, body temperature, heart rate, DBP, pain score, chronic kidney disease, tenderness (including rebound tenderness), or sonography performance.

The results of comparative analyses between the URVA and URVNA groups are presented in Table 2. In our logistic regression analysis, age was employed as the continuous variable (models 1 and 3) and categorical variable (models 2 and 4). To obtain the best cut-off value for age, we found the age at ≥ 40 vs < 40 year had the highest Youden’s index (Additional file 1: Table S2). All models attained the criteria of stable goodness of fit in the Hosmer–Lemeshow test. For comparison of models, the second strategy of backward elimination was more favored than the first strategy due to the lower AIC values. Moreover, the model 4 with age as the categorical variable was selected as the best model due to the lowest AIC value of 55.17.

In model 4, older age (≥ 40 vs < 40 year: AOR, 2.10; 95% CI 1.47–2.99), execution of laboratory tests (yes vs no: AOR, 3.70; 95% CI 2.54–5.39), and administration of multiple analgesics (≥ 2 vs < 2: AOR, 2.06; 95% CI 1.09–3.87) were associated with increased odds of URVA. Both models 3 and 4 revealed consistent results.

The ROC curves of risk factors for URVA are displayed in Fig. 2. The areas under the curves (AUCs) for triage score, administration of multiple analgesics, age (serving as a continuous variable), and execution of laboratory tests were 0.527 (P = 0.236), 0.535 (P = 0.122), 0.619 (P < 0.001), and 0.656 (P < 0.001), respectively. Through these 4 risk factors, the model could achieve acceptable discrimination performance (AUC = 0.716; P < 0.001) in the prediction of URVAs.

Our logistic regression analysis results indicated that age ( ≥ 40 year), triage score (Levels 1–2), execution of laboratory tests, and number of analgesics used (≥ 2) were significant variables predicting URVAs after the index visit to the ED (Table 3). Cutoff points for these variables were derived by using the maximum value of the Youden's index on the corresponding ROC curves (Additional file 2: Fig. S1). Thus, these variables were used to establish a new scoring system for predicting URVAs. A score of 1 was assigned to age (≥ 40 year), triage score (Levels 1–2), and number of analgesics used (≥ 2), and a score of 2 was assigned to execution of laboratory tests; thus, the total score for the system ranged from 0 to 5. On the basis of the maximum value of Youden's index on ROC curves, the optimal cut-off point for the variables predicting URVAs was 3, with the corresponding sensitivity and specificity being 63.1% and 68.0%, respectively (Fig. 3). Accordingly, the established system achieved acceptable discrimination performance in predicting URVAs (AUC = 0.709; 95% CI 0.672–0.746) and acceptable calibration (Hosmer–Lemeshow test; P = 0.483).

To explore influential predictors of URVAs, a CART model was established to predict URVAs and URVNAs (Fig. 4). The 4 variables included in this model were execution of laboratory tests (yes vs no), age (≥ 40 vs < 40 year), triage score (Levels 1–2 vs Levels 3–5), and number of analgesics administered (≥ 2 vs < 2). Execution of laboratory tests was the first determining predictor in this model. Among 274 patients who did not undergo laboratory tests, only 47 (17%) had URVAs. Of 428 patients who underwent laboratory tests, age (≥ 40 vs < 40 year) was a predictor of URVAs. Patients aged < 40 years (n = 148) were less likely to have URVAs (n = 55; 37.2%) compared with other patients. Among patients aged > 40 years, those with a triage score of Levels 1–2 (n = 25) were more likely to have URVAs (n = 18; 72.0%) compared with other patients. Among patients with a triage score of Levels 3–5 (n = 255), number of analgesics administered was a predictor of subsequent branching. Patients who received 2 or more analgesics (n = 27) were more likely to have URVAs (n = 18; 66.7%) compared with other patients.