Choledocholithiasis (CL) is a frequent cause of extrahepatic bile duct obstruction that can be diagnosed in up to 15% of patients with cholecystolithiasis.1,2 Given that there can be complications due to the presence of CL (acute cholangitis, acute pancreatitis), its early diagnosis and treatment is of vital importance. Diagnosis is currently based on clinical, radiologic, and laboratory parameters. Considering body mass index, age, sex, the presence of bile duct dilation, and liver function tests, the American Society for Gastrointestinal Endoscopy (ASGE) and the European Society for Gastrointestinal Endoscopy (ESGE) developed predictive scales for CL.2,3

The ASGE criteria classify patients into low risk, intermediate risk, and high risk for CL, with a probability of 10%, 10–50%, and >50%, respectively. However, accuracy varies from 40% to 85% for high risk and 30%–40% for intermediate risk.2 Due to the wide variability in the diagnostic yield of the abovementioned criteria in clinical practice, other methods utilized for predicting CL diagnosis, such as logistic regression and symbolic regression, have been implemented.4

The gold standard for diagnosing CL is endoscopic retrograde cholangiopancreatography (ERCP), with 94% sensitivity and 100% specificity, but this method is not exempt from serious complications and should be performed mainly in the context of therapeutic indications.5

Therefore, we evaluated the diagnostic accuracy of our application through a logistic regression model utilized in patients with an intermediate or high risk for CL.

An analytic, observational, cross-sectional study was conducted at a single center within the time frame of February 1, 2022, and February 1, 2023. As background in 2021, a logistic regression model was obtained through artificial intelligence for predicting the diagnosis of CL. Initially, the model was validated in a retrospective cohort of patients diagnosed with intermediate or high risk for CL that underwent ERCP during 2020 at the gastroenterology service of our hospital. From that precedent, the model was prospectively applied to hospitalized patients with intermediate or high risk for CL, to evaluate its diagnostic yield. Only patients with the presence of a stone during ERCP were considered to have a definitive diagnosis of CL. The model was developed by one of the authors (LM T-T), who is a Doctor in Artificial Intelligence. An established cutoff point of ≥0.6 for discerning whether the model predicted a positive result (the presence of a stone in ERCP) or a negative cutoff point <0.6 (no stone in ERCP) was determined. The App model was applied at the bedside of 148 patients admitted to our hospital. Inclusion criteria were patients ≥18 years of age, with clinical suspicion, and with intermediate or high risk of CL through laboratory or imaging tests. Exclusion criteria were patients at low risk for CL by the ASGE classification (considering that those patients did not require invasive studies before cholecystectomy), patients with previous cholecystectomy, previous ERCP or biliary surgery, pregnant women, patients with cirrhosis of the liver, clinical suspicion of cholangitis, ASA III, and patients that did not complete their follow-up at our hospital. Laboratory tests were carried out at hospital admission and at 24 and 48h after hospitalization. The laboratory tests performed upon hospital admission were utilized to classify patients according to the ASGE criteria.

Logistic regression model

Our logistic regression model included age, sex, time of ERCP to hospital admission, AST, ALT, alkaline phosphatase, total bilirubin, and the diameter of the common bile duct measured by abdominal ultrasound, for their analysis (Fig. 1). Through the logistic regression model, the contribution of each variable in predicting CL was evaluated and utilized for the control of other confounding factors.

Figure 1.

Logistic regression model used to evaluate the clinical, laboratory, and imaging characteristics of 148 patients with intermediate or high risk of choledocholithiasis and to predict the diagnosis of choledocholithiasis.

ALP: alkaline phosphatase; ALT: alanine aminotransferase; AST: aspartate aminotransferase; CBD: common bile duct; F: female; M: male.

(0.65MB).

Three experienced endoscopists carried out all the ERCPs in our study. The results of the model were not taken into consideration for making medical decisions.

A total of 148 patients with either intermediate or high risk for CL were recruited. Median age was 43 years (range:16−85) and 110 (74%) of the patients were women. As mean values, AST was 214.37 ± 184.25 U/l, ALT 288.13 ± 228.17 U/l, alkaline phosphatase 290.49 ± 174.57 U/l, total bilirubin 6.87 ± 5.13 mg/dl, and the diameter of the common bile duct 9.2 ± 5.4 mm. In our cohort, 71 (48%) patients were classified as having an intermediate risk for CL and 77 (52%) as having a high risk for CL. Table 1 shows all laboratory test results. ERCP was performed on 125 (85%) patients ≥48h from hospital admission. CL diagnosis through the identification of a stone by ERCP was made in 102 (69%) patients. ECRP demonstrated the presence of a stone in the common bile duct in 26 (36.6%) patients at intermediate risk for CL and in 41 (53.2%) patients at high risk for CL.

Prediction of choledocholithiasis made by our logistic regression model

The analysis of our model in the cohort revealed an area under the ROC curve (AUC) value of 0.68, indicating moderate predictive capaciVty. The AUC value for patients with intermediate risk was 0.72, with 65% sensitivity, 65% specificity, 70% PPV, 59% NPV, and 71% accuracy. In high-risk patients, the AUC value was 0.78, with 66% sensitivity, 67% specificity, 89% PPV, 32% NPV, and 89% accuracy (Fig. 2).

Figure 2.

ROC curve analysis to predict the definitive diagnosis of choledocholithiasis, in high and intermediate-risk patients, respectively.

(0.16MB).

The development of different noninvasive, economic tools (neural networks, machine learning) for predicting the presence of CL is vitally important, given that current methods (magnetic resonance imaging, ERCP) for its diagnosis are costly, not exempt from risks for the patient, and not always widely available. Unfortunately, the development of CL prediction tools continues to produce heterogeneous results.

Currently, the ASGE criteria define patients as having low, intermediate, and high risks for CL, but when applied to different populations, these definitions vary greatly. Matt Ridley expressed this concept in his book, The Agile Gene.6 He wrote that persons are similar because they are different and different because they are similar. In such a context, the applicability of any score or criterion becomes quite difficult.

In 2017, Narváez et al. applied the ASGE criteria to patients at the Hospital Universitario “Dr. José Eleuterio González”. Those authors reported a diagnostic accuracy of CL in the high-risk patients of 59%, with 85% sensitivity and 24% specificity. In the intermediate-risk patients, accuracy was 41%, with 14% sensitivity and 75% specificity, indicating unnecessary ERCP in almost half the patients.7 With those data in mind, we developed a logistic regression model to be prospectively applied as an App at the bedside of the patient suspected of presenting with CL. In our model, the PPV was 89% and accuracy was 89% for high-risk patients. Our results were similar to those reported by Dalai et al., who analyzed 270 patients at high risk for CL, utilizing artificial intelligence. Those authors described 91% sensitivity, 25% specificity, 87% PPV, 33% NPV, and 81% accuracy.8 Additionally, in 2014, a published prospective study by Jovanovic et al. utilized an artificial neuronal network to determine the risk for ERCP indication in patients with suspected CL. The AUC value was 0.88 (95% CI 83–93%), significantly higher than that of our model, but showed a similar PPV of 92% in patients at high risk of CL. That model correctly classified 92% of patients who needed an ERCP.9

A study by Steinway et al., utilizing a machine learning-based method for predicting CL in 1,378 patients, compared the diagnostic accuracy of the gradient boosting machine-learning method versus the 2019 ASGE and ESGE criteria, finding accuracy of 71%, 62%, and 62%, respectively, results not significantly different from ours.4

In a study on 1,171 patients, He et al. reported that the specificity of the ASGE criteria for CL in high-risk patients was 74% (95% CI 72–77%) and the PPV was 64% (95% CI 61–675). Even though the high-risk criteria demonstrated a probability above 50% of presenting with CL, more than one-third of the patients underwent diagnostic ERCP.10 In a retrospective article by Jagtap et al. that compared the ASGE and ESGE criteria for CL in high-risk patients, a higher PPV was obtained with the ESGE criteria.11

Table 2 shows the results of a group of studies on patients at high risk of CL, describing sensitivity from 61% to 91%. In addition, specificity varied from 24% to 97%. All studies were conducted on populations with a different prevalence of choledocholithiasis.7,8,10–13

Our predictive model applied to the patients with intermediate risk for CL had 65% sensitivity, 65% specificity, 70% PPV, 59% NPV, and 71% accuracy for CL diagnosis. Our results are similar to those reported in the literature (Table 3).7,11,12

Most of the prediction models reviewed in the literature have good diagnostic yield in patients at high risk for CL. These data can help confirm the indication for therapeutic ERCP or the necessity for bile duct examination (Table 4).

Nevertheless, the medical decision in patients in the intermediate-risk group continues to be a theme of interest because the PPV of our model in that group of patients was 70%, which was a marginal value, similar to that of several other studies. Therefore, an additional imaging study, such as magnetic resonance imaging or endoscopic ultrasound, is still necessary for confirming the diagnosis of CL.14

It should be highlighted that our logistic regression model showed improvement in diagnostic accuracy, with respect to the ASGE criteria. However, evaluating its performance in a different population is necessary because of the high prevalence variability of bile duct stones.

The limitations of our study include the fact that it was conducted at a single center, ERCP was performed ≥ 48h from hospital admission, we did not have laboratory test data from the same day as the ERCP, the majority of our patients were young women with a mean age under 40 years, and our sample size was based on consecutive sampling of patients with intermediate or high risk of CL during one year. Most likely, the diagnostic yield of our model would be more statistically robust if the 3 CL risk groups had been evaluated. Nevertheless, we only selected patients at intermediate or high risk for CL, because in low-risk patients ERCP is not needed and cholecystectomy is indicated.

The strong point of our article is that the model was developed utilizing adequate methodology and it was prospectively applied for diagnosing CL. All ERCPs were performed by expert endoscopists, and the model was applied at the patient’s bedside.

Our logistic expression model showed an improvement in diagnostic accuracy, with respect to the ASGE criteria. We believe our findings can be useful for guiding the physician in his/her clinical decision-making in patients with suspected choledocholithiasis.

No financial support was received in relation to this study/article.