Diagnosing distal bile duct obstruction remains challenging. This study aimed to examine the diagnostic ability of artificial intelligence (AI) based on clinical biomarkers in diagnosing malignant distal bile duct obstruction. A total of 206 patients with distal bile duct obstruction were included in this study. Clinical laboratory parameters were collected from the patients and evaluated using AI. All clinical parameters were input into the AI algorithm, and the AI value for malignant distal bile duct obstruction was calculated. The benign and malignant diagnostic capabilities of AI and other factors (alkaline phosphatase [ALP], intrahepatic bile duct [IHBD] diameters, and total bile duct [CBD] diameters) were compared. Benign and malignant bile duct obstruction were diagnosed in 142 and 64 patients, respectively. The median AI value of malignant distal bile duct obstruction was significantly greater than that of benign distal bile duct obstruction (0.991 vs. 0.002, p < 0.001). The area under the receiver operating characteristic curve of AI, ALP, IHBD diameter, and CBD diameter were 0.908, 0.795, 0.794, and 0.775, respectively. AI showed a sensitivity, specificity, and accuracy of 83.1%, 87.2%, and 85.9%. AI-based on clinical biomarkers could serve as an auxiliary for diagnosing malignant bile duct obstruction.
Distal bile duct obstruction may occur not only in benign diseases, such as choledocholithiasis, chronic pancreatitis, and immunoglobulin G4 (IgG4)-related sclerosing cholangitis, but also in malignant diseases, such as distal bile duct and pancreatic cancers1. Distal bile duct obstructions, especially malignant ones, often require surgical resection and chemotherapy. Therefore, an accurate diagnosis of malignant bile duct obstruction is necessary.
Transabdominal ultrasonography (US), computed tomography (CT), magnetic resonance cholangiopancreatography (MRCP), and endoscopic examinations, including endoscopic retrograde cholangiopancreatography (ERCP), are performed to diagnose distal bile duct obstruction. In lower bile duct obstructions, the diagnostic accuracies of US, multi-detector row CT, and plain MRCP for malignant lesions are 36.8%2, 73.0%3, and 78.0%4, respectively. As a result, accurately diagnosing and differentiating benign from malignant duct obstruction using imaging modalities are still challenging. Although a pathological diagnosis is necessary for a definitive diagnosis, the sensitivity for pathological malignancy by cytology with ERCP and endoscopic transpapillary forceps biopsies in patients with distal bile duct obstruction is 30–48%5 and 59%3, respectively. Hence, the accuracy of pathological diagnosis is inadequate. It has also been reported that 15–24% of cases in which biliary tract cancer was suspected were benign6. Therefore, preoperative differentiation of benign and malignant distal bile duct obstruction remains challenging.
In recent years, artificial intelligence (AI) advancements have been remarkable7,8. AI has been used in biliary-pancreatic diseases, relatively common in intrahepatic cholangiocarcinoma, and a wide variety of modalities have been reported, including CT, magnetic resonance imaging (MRI)9, and various AI models such as artificial neural network (ANN)10 and support vector machine (SVM)11. In the case of extrahepatic cholangiocarcinoma, the radiomics models of MRI through random forests showed an area under the curve (AUC) of 0.8 and 0.9 for the detection of differentiation degree and lymph node metastasis12, respectively, but there are no reports on diagnoses of benign and malignant distal bile duct obstruction. As it is known, the diagnosis or differential diagnosis of distal bile duct obstruction is sometimes difficult. It requires much experience and knowledge because of the variety of radiological features of each disease and the overlap of radiological findings between many diseases. Therefore, in this study, we examined the diagnostic ability of AI using clinical laboratory parameters in differentiating malignant from benign distal bile duct obstruction.
Patients and feature selection
We retrospectively reviewed all patients whose distal bile duct obstruction or dilated upstream bile duct was detected on CT scan and underwent ERCP from April 2015 to August 2020 at Yokohama Sakae Kyosai Hospital. Bile duct abnormalities were considered to be involved when the radiologists encountered: (1) bile duct dilatation (7 mm for the common bile duct [CBD]), or (2) biliary wall thickening (1 mm), or (3) masses inside or around the affected duct. Diseases of benign bile duct obstruction included choledocholithiasis, IgG4-related sclerosing cholangitis, chronic pancreatitis, and other benign stenoses. Malignant bile duct obstruction diseases included distal bile duct cancer, pancreatic cancer, ampullary cancer, lymphoma, recurrent lymph node of gastric carcinoma, and duodenal carcinoma. The clinical laboratory parameters of the patients that were evaluated in this study were extracted before ERCP. All laboratory biomarkers were first examined before biliary drainage. The parameters were age, sex, weight, height, intrahepatic bile duct (IHBD) diameter, CBD diameter, leukocyte count (WBC), and WBC fraction (neutrophils and lymphocytes). The values of total bilirubin (T-bil), aspartate aminotransferase (AST), alanine aminotransferase (ALT), γ-glutamyl transpeptidase (γ-GTP), alkaline phosphatase (ALP), amylase, albumin, lactate dehydrogenase (LDH), C-reactive protein (CRP), hemoglobin A1c (HbA1c), glucose, triglyceride (TG), low-density lipoprotein (LDL), high-density lipoprotein (HDL), total cholesterol (T-chol), carcinoembryonic antigen (CEA), carbohydrate antigen 19–9 (CA19-9), and concurrent biliary infection were also assessed. Parameters that had more than half of the missing data were excluded beforehand. Extraction of the clinical information and AI analysis was done by different authors, blind to each other.
This study was approved by the Institutional Review Boards of Yokohama Sakae Kyosai Hospital (20,201,019-2), Aichi Cancer Center Hospital (2021-1-044) and performed in accordance with the Declaration of Helsinki13. This retrospective observational study used only medical information and did not compromise the privacy of the participants. All patients received an opt-out form for informed consent. Those who did not agree to participate were excluded. All authors had access to the research data and reviewed and approved the final manuscript.
Definition of final diagnosis
Histopathological examination of the surgical specimens was considered the final diagnosis in cases of surgical resection. The final diagnosis of patients who did not undergo surgery was determined by either biopsy, cytology, clinical data, or imaging with a 1-year follow-up.
Regarding the diagnosis of malignancy, surgical specimens were used in surgical cases, and biopsy specimens or cytology were used in non-surgical cases. The benign disease was diagnosed if there were no malignant findings on surgical specimens, biopsy specimens, or cytology, and there was no change in the imaging results after at least a 1-year follow-up period. The diagnosis of choledocholithiasis was made following Tokyo Guidelines 201814. The diagnosis of IgG4-related sclerosing cholangitis was made based on clinical diagnostic criteria of IgG4-related sclerosing cholangitis 201215. Chronic pancreatitis was diagnosed per evidence-based clinical practice guidelines for chronic pancreatitis 201516.
All machine learning methods used scikit-learn17. For the imputation of missing values, we used the K nearest neighbor algorithms. When predicting a certain value, we used K neighboring values (5 in this case) as a guide18. The synthetic minority oversampling technique19 and the undersampling method20 were used to correct the ratio of benign and malignant tumors. The Boruta method21 was used for feature selection.
For the machine learning method, we used light GBM (Microsoft.com)22, which is an easy-to-use (short training time) machine learning model used for the AI predictive model. Light GBM is based on gradient boosting23, which is an ensemble learning method using decision trees as weak learners24. The leaf-wise method25 was chosen to handle the decision tree. The predictive value of malignant probability using AI was continuous variables from 0 to 1. When the AI value was close to 1, the malignant probability increased. The tenfold cross-validation26 was used to accurately verify the generalization performance of the predictive model.
The primary endpoint was to investigate and compare the diagnostic ability of AI, ALP, IHBD diameter, and CBD diameter using clinical laboratory parameters in differentiating malignant from benign lesions in patients with distal bile duct obstruction. We also investigated the diagnostic sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy of AI.
Statistical analysis was performed using R software (version 4.0.2). Univariate analysis was performed using the Mann–Whitney U test for continuous variables. The Chi-square test or Fisher’s exact test was used for categorical variables. Statistical significance was set at p < 0.05. Receiver operating characteristic (ROC) curve analysis was performed to characterize the diagnostic ability of ALP, IHBD diameter, CBD diameter, and AI in differentiating malignant from benign distal bile duct obstruction. The area under the ROC curve (AUROC) was used to evaluate the accuracy of the quantitative tests.
A total of 222 patients were identified; however, 16 were excluded due to missing data. Hence, 206 patients were included in the final analysis (Fig. 1). Benign and malignant bile duct obstructions were diagnosed in 142 and 64 patients, respectively. HbA1c, TG, LDL, HDL, T-chol, CEA, and CA19-9 were excluded beforehand, as these parameters had more than half the missing data. Characteristics of the patients are shown in Table 1. Factors used for the final AI analysis were IHBD diameter, CBD diameter, WBC, WBC fraction, T-bil, AST, ALT, γ-GTP, ALP, amylase, LDH, and CRP levels. The median IHBD diameter, CBD diameter, WBC count, ALP, ALT, γ-GTP, and T-bil of patients with malignant distal bile duct obstruction were significantly higher than those with benign distal bile duct obstruction. However, AST, LDH, amylase, and CRP levels were not significantly different between patients with malignant and benign distal bile duct obstruction.
The final diagnoses of patients are summarized in Table 2. In total, 142 and 64 patients had a final diagnosis of benign (choledocholithiasis, 135 patients; others, 7 patients) and malignant bile duct obstruction (distal bile duct cancer, 25 patients; pancreatic cancer, 26 patients; others, 13 patients), respectively.
ROC curve analysis and comparison of the predictive value of AI in differentiating malignant from benign distal bile duct obstruction
The median AI predictive value for malignant distal bile duct obstruction was significantly higher than that for benign distal bile duct obstruction (0.991 vs. 0.002, p < 0.001) (Fig. 2). ROC analysis was calculated to determine the cutoff value for the diagnosis of malignancy. Using an AI predictive value of 0.609 as the cutoff point resulted in the diagnostic ability of AI with a sensitivity of 83.1%, specificity of 87.2%, PPV of 74.5%, NPV of 92.0%, and accuracy of 85.9% (Table 3).
ROC curve analysis of the malignant predictive value of AI was demonstrated to have an AUROC of 0.908, which was significantly greater than those of ALP (0.795, p = 0.008), IHBD diameter (0.794, p = 0.007), and CBD diameter (0.775, p = 0.003) in diagnosing malignant distal bile duct obstruction (Fig. 3).
AI is a new technology for objective evaluation. The development of AI in the medical field is growing rapidly. Studies of AI tools in gastroenterology have focused primarily on endoscopy, and the number of studies evaluating these models for hepatobiliary indications using clinical parameters is few27. Differentiating malignant from benign bile duct obstruction is still challenging. This study is the first on the benign and malignant diagnosis of distal bile duct obstruction using AI tools with common parameters.
Our AI diagnostic model using general biomarkers showed higher diagnostic performance than blood test parameters or imaging tests. This study is unique because it shows that it may be possible to diagnose bile duct obstruction using only the general parameters performed when the patient is suspected of having bile duct obstruction. AI diagnostic models using general biomarkers might be easier to perform in a clinical setting. Biomarker-based AI may be an effective auxiliary diagnostic tool.
The development of an AI model that can accurately distinguish between benign and malignant biliary tract disorders can significantly impact a patient’s treatment strategy and prognosis. Previous reports have shown that the positive diagnosis rate of the tumor marker CA19-9 for biliary obstruction was 74%28. Although we could not directly compare AI and CA19-9, the AI in this study was better than that in previous reports regarding CA19-9. AI may be more useful than tumor markers in diagnosing malignancy. In addition to its diagnostic usefulness, we plan to investigate whether AI can be used as a prognostic indicator similar to tumor markers.
There are several methods for diagnosing biliary bile duct obstruction. However, contrast-enhanced CT has risks such as renal dysfunction and anaphylaxis29. In addition, ERCP contrast and bile duct biopsy are highly invasive, as they may lead to fatal turning points such as post-ERCP pancreatitis30. MRI is also useful, but there are restrictions such that it cannot be performed on patients with metallic implants31. This study included only general laboratory parameters that are often used clinically and can assist in making a diagnosis without using expensive machines or micro-alleys. Therefore, biomarker-based AI may be a minimally invasive and effective auxiliary diagnostic tool.
This study had several limitations. First, it was a single-center retrospective study. Second, this study had a small sample size (N = 206). Therefore, internal validation for AI diagnosis was used for malignancy (tenfold cross-validation). When assessing the diagnostic ability, it was necessary to increase the number of training and test data while separating them from each other. Aligning benign and malignant cases would decrease the number of cases. To increase the number of training data to improve the performance of AI, benign and malignant cases were not aligned. The number of benign and malignant cases differed, which may have affected the results. The apparent accuracy may have been high due to the large number of benign cases. Still, we tried to compensate for the differences as much as possible by using the synthetic minority oversampling technique and the undersampling method when creating the model for AI analysis. Third, many missing values were excluded, such as HbA1c, TG, LDL, HDL, T-chol, CEA, and CA19-9. These parameters were excluded because more than half of them had missing values. Notably, CEA and CA19-9 are important markers when determining if a lesion is benign or malignant and should have been included in this analysis. We plan to include these markers in prospective studies to further improve the diagnostic performance of benign and malignant biliary obstruction in the future. However, we propose that AI can be used as a more accurate diagnostic tool and prognostic predictor.
In conclusion, AI based on clinical biomarkers may be useful for objective evaluation prior to invasive examination and treatment of distal bile duct obstruction. This method could serve as an auxiliary for diagnosing malignant bile duct obstruction.
The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.
Coucke, E. M., Akbar, H., Kahloon, A. & Lopez, P. P. in StatPearls (StatPearls Publishing Copyright © 2021, StatPearls Publishing LLC., 2021).
Albu, S. et al. Diagnosis and treatment of extrahepatic cholangiocarcinoma: Results in a series of 124 patients. Rom. J. Gastroenterol. 14, 33–36 (2005).
Heinzow, H. S. et al. Comparative analysis of ERCP, IDUS, EUS and CT in predicting malignant bile duct strictures. World J. Gastroenterol. 20, 10495–10503 (2014).
Sun, N., Xu, Q., Liu, X., Liu, W. & Wang, J. Comparison of preoperative evaluation of malignant low-level biliary obstruction using plain magnetic resonance and coronal liver acquisition with volume acceleration technique alone and in combination. Eur. J. Med. Res. 20, 92 (2015).
Navaneethan, U. et al. Comparative effectiveness of biliary brush cytology and intraductal biopsy for detection of malignant biliary strictures: A systematic review and meta-analysis. Gastrointest. Endosc. 81, 168–176 (2015).
Wakai, T. et al. Clinicopathological features of benign biliary strictures masquerading as biliary malignancy. Am. Surg. 78, 1388–1391 (2012).
Kurita, Y. et al. Diagnostic ability of artificial intelligence using deep learning analysis of cyst fluid in differentiating malignant from benign pancreatic cystic lesions. Sci. Rep. 9, 6893. https://doi.org/10.1038/s41598-019-43314-3 (2019).
Kuwahara, T. et al. Usefulness of deep learning analysis for the diagnosis of malignancy in intraductal papillary mucinous neoplasms of the pancreas. Clin. Transl. Gastroenterol 10, 1–8 (2019).
Liu, X. et al. Can machine learning radiomics provide pre-operative differentiation of combined hepatocellular cholangiocarcinoma from hepatocellular carcinoma and cholangiocarcinoma to inform optimal treatment planning?. Eur. Radiol. 31, 244–255 (2021).
Urman, J. M. et al. Pilot multi-omic analysis of human bile from benign and malignant biliary strictures: A machine-learning approach. Cancers (Basel) 12, 1644 (2020).
Xu, L. et al. A radiomics approach based on support vector machine using MR images for preoperative lymph node status evaluation in intrahepatic cholangiocarcinoma. Theranostics 9, 5374–5385. https://doi.org/10.7150/thno.34149 (2019).
Yang, C. et al. Radiomics model of magnetic resonance imaging for predicting pathological grading and lymph node metastases of extrahepatic cholangiocarcinoma. Cancer Lett. 470, 1–7. https://doi.org/10.1016/j.canlet.2019.11.036 (2020).
World Medical Association Declaration of Helsinki. Ethical principles for medical research involving human subjects. JAMA 310, 2191–2194. https://doi.org/10.1001/jama.2013.281053 (2013).
Kiriyama, S. et al. Tokyo Guidelines 2018: Diagnostic criteria and severity grading of acute cholangitis (with videos). J. Hepatobiliary Pancreat. Sci. 25, 17–30. https://doi.org/10.1002/jhbp.512 (2018).
Ohara, H. et al. Clinical diagnostic criteria of IgG4-related sclerosing cholangitis 2012. J. Hepatobiliary Pancreat. Sci. 19, 536–542. https://doi.org/10.1007/s00534-012-0521-y (2012).
Ito, T. et al. Evidence-based clinical practice guidelines for chronic pancreatitis 2015. J. Gastroenterol. 51, 85–92. https://doi.org/10.1007/s00535-015-1149-x (2016).
Kramer, O. Machine Learning for Evolution Strategies 45–53 (Springer, 2016).
Bertsimas, D., Pawlowski, C. & Zhuo, Y. D. From predictive methods to missing data imputation: An optimization approach. The Journal of Machine Learning Research 18, 7133–7171 (2017).
Zhu, T., Lin, Y. & Liu, Y. Synthetic minority oversampling technique for multiclass imbalance problems. Pattern Recognit. 72, 327–340 (2017).
Mohammed, R., Rawashdeh, J. & Abdullah, M. in 2020 11th international conference on information and communication systems (ICICS). 243–248 (IEEE).
Kursa, M. B. & Rudnicki, W. R. Feature selection with the Boruta package. J. Stat. Softw. 36, 1–13 (2010).
Ke, G. et al. Lightgbm: A highly efficient gradient boosting decision tree. Adv. Neural Inf. Process. Syst. 30, 3149–3157 (2017).
Friedman, J. H. Greedy function approximation: a gradient boosting machine. Ann. Stat. 29, 1189–1232 (2001).
Safavian, S. R. & Landgrebe, D. A survey of decision tree classifier methodology. IEEE Trans. Syst. Man Cybern. 21, 660–674 (1991).
Shi, H. Best-first Decision Tree Learning Master of Science (MSc) thesis, University of Waikato. The University of Waikato (2007).
Fushiki, T. Estimation of prediction error by using K-fold cross-validation. Stat. Comput. 21, 137–146 (2011).
Le Berre, C. et al. Application of artificial intelligence to gastroenterology and hepatology. Gastroenterology 158, 76-94.e72. https://doi.org/10.1053/j.gastro.2019.08.058 (2020).
Liu, W. et al. Differential diagnostic roles of the serum CA19-9, total bilirubin (TBIL) and the ratio of CA19-9 to TBIL for benign and malignant. J. Cancer 9, 1804–1812. https://doi.org/10.7150/jca.25093 (2018).
Christiansen, C., Pichler, W. J. & Skotland, T. Delayed allergy-like reactions to X-ray contrast media: Mechanistic considerations. Eur. Radiol. 10, 1965–1975. https://doi.org/10.1007/s003300000543 (2000).
Kochar, B. et al. Incidence, severity, and mortality of post-ERCP pancreatitis: A systematic review by using randomized, controlled trials. Gastrointest. Endosc. 81, 143-149.e149. https://doi.org/10.1016/j.gie.2014.06.045 (2015).
Ahmed, F. Z. et al. Not all pacemakers are created equal: MRI conditional pacemaker and lead technology. J. Cardiovasc. Electrophysiol. 24, 1059–1065. https://doi.org/10.1111/jce.12238 (2013).
The authors are grateful to S. Watanabe, T. Shimizu, Y. Takai, and E. Sakai from the Yokohama Sakae Kyosai Hospital for their assistance with this study.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
About this article
Cite this article
Sugimoto, Y., Kurita, Y., Kuwahara, T. et al. Diagnosing malignant distal bile duct obstruction using artificial intelligence based on clinical biomarkers. Sci Rep 13, 3262 (2023). https://doi.org/10.1038/s41598-023-28058-5