Development and Validation of a Diagnostic Model for Biliary Tract Cancer Detection in Patients with Benign Biliary Disease: A Multicenter, Retrospective Case-Control Study

Development and Validation of a Diagnostic Model for Biliary Tract Cancer Detection in Patients with Benign Biliary Disease: A Multicenter, Retrospective Case-Control Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Development and Validation of a Diagnostic Model for Biliary Tract Cancer Detection in Patients with Benign Biliary Disease: A Multicenter, Retrospective Case-Control Study Hanul Park, Kibeom Kim, Jonghyun Lee, Sung Yong Han, Chang Min Cho, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8736730/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background/Aims Biliary tract cancer (BTC) is an aggressive malignancy often diagnosed late due to its frequently asymptomatic presentation. Given substantial clinical overlap between BTC and benign biliary tract disease (BTD), early and accurate differentiation remains challenging. This study represents the first effort to develop and validate a non-invasive clinical model capable of identifying patients at especially high risk of BTC within the BTD population, thereby facilitating earlier diagnosis and intervention. Methods In this multicenter observational study, we prospectively collected patient-reported survey data and retrospectively extracted laboratory, diagnostic data from electronic medical records at four tertiary centers in South Korea. A machine learning model was trained to differentiate biliary tract cancer (BTC) from benign biliary tract disease (BTD) using 1,439 patients from three centers and externally validated the model in 245 patients from an independent center. Explainable machine learning quantified biomarker contributions to model predictions. Results Our model demonstrated robust diagnostic performance, achieving an AUROC of 0.893 (95% CI, 0.863–0.920) and sensitivity of 0.784 (0.727–0.827). External validation from an independent center dataset produced consistent results. Notably, the model substantially outperformed CA19-9 and was developed without reliance on this biomarker, enabling applicability in Lewis antigen–negative patients. It also accurately identified BTC cases even among patients with normal CA19-9 levels, supporting its utility in broader clinical populations. Conclusions These results suggest our model’s potential to serve as a non-invasive screening tool, identifying patients at high risk of BTC among those presenting with suspicious biliary pathology, particularly when CA19-9 is uninformative. Biliary tract cancer cholangiocarcinoma machine learning diagnosis prediction Figures Figure 1 Figure 2 Figure 3 Figure 4 Background Biliary tract cancer (BTC) is a rare malignancy with poor prognosis arising from the gallbladder, biliary tree, or, less commonly, the ampulla of Vater [1]. Although its incidence is relatively low compared to other malignancies, the high mortality rate underscores its clinical significance [2,3]. The high case-fatality rate of BTC is largely attributable to late-stage diagnosis, often resulting from the absence of early symptoms [1,4]. Previous studies investigating the risk factors for BTC have identified several benign biliary tract diseases (BTD), including primary sclerosing cholangitis (PSC), hepatolithiasis, liver fluke infestation ( Opisthorchis viverrini and Clonorchis sinensis ), and congenital biliary abnormalities. Beyond sharing common epidemiologic and metabolic risk factors such as chronic inflammation, hepatitis B and C infections, obesity, and diabetes [5-8], BTD itself may serve as a premalignant condition that predisposes patients to subsequent BTC development. However, the accuracy of detecting early-stage BTC in patients with these conditions remains limited as BTC and BTD frequently exhibit indistinguishable symptoms and radiologic features during early disease stages [6-8]. Serologic tumor markers, including carbohydrate antigen 19-9 (CA19-9) and carcinoembryonic antigen (CEA), have been proposed as adjunctive tools for BTC risk stratification. However, their diagnostic performance remains suboptimal, particularly in patients with underlying BTD, due to variable sensitivity, limited specificity, and frequent false-positive elevations associated with benign biliary inflammation [4,6–8]. Consequently, reliance on conventional diagnostic strategies alone is insufficient for identifying early-stage BTC among patients initially diagnosed with BTD. Several studies have examined the risk of BTC within specific BTD cohorts—most notably among patients with PSC [9–11]. More recently, machine learning algorithms have shown promise in detecting BTC, owing to their ability to capture complex interactions between patient characteristics and disease status [12-14]. Nevertheless, key challenges remain: Risk factors for BTC are still unclear; prior studies often relied on narrow cohorts—primarily PSC patients—limiting generalizability; and the overall diagnostic performance of existing tools remains suboptimal. From this perspective, there is a critical unmet need for a systematic approach to identify high-risk subpopulations within BTD patients who may harbor occult or early-stage BTC. Standard diagnosis of BTC requires a high index of suspicion and the integration of clinical, laboratory, endoscopic, and radiologic data [6-8]. However, such approaches remain impractical for routine application and are inherently constrained by the asymptomatic nature of early-stage BTC, financial barriers and patient reluctance to undergo advanced imaging modalities and invasive procedures [15]. Consequently, current diagnostic paradigms are insufficient to reliably distinguish early BTC from BTD at the point of initial presentation. Accurate risk stratification at the time of BTD diagnosis represents a pivotal opportunity to enable selective surveillance and timely, targeted diagnostic evaluation, thereby maximizing the potential for early detection and curative intervention. In this context, we aimed to develop a non-invasive, data-driven prediction model to facilitate the early detection of BTC. Using routinely available demographic and clinical data from patients diagnosed with BTD and BTC, we developed and externally validated a machine learning model designed to screen and identify patients at particularly high risk of BTC within the BTD population. Methods Data source and study cohort This study utilized a hybrid data collection framework in which patient-reported survey data were collected prospectively, while clinical and laboratory variables were retrospectively extracted from electronic medical records. The dataset was curated for the development and external validation of a predictive classification model distinguishing biliary tract cancer (BTC) from benign biliary disease (BTD). BTC subtypes included intrahepatic, perihilar, and distal cholangiocarcinoma, as well as gallbladder cancer. The BTD subtypes are listed in Supplementary Table 1. Data were obtained from four academic medical centers in South Korea between July 2020 and September 2022. Of 2,084 patients initially enrolled across these centers, 31 (1.5%) were excluded due to incomplete survey data and an additional 369 (17.7%) were excluded for missing all blood-serum measurements or having over 40% missing features, in order to ensure data completeness and model robustness. The final training dataset consisted of 1,439 patients remained in the training cohort—548 with BTC and 891 with BTD—drawn from three hospitals (Pusan National University Hospital; Kyungpook National University Hospital; and Kyungpook National University Chilgok Hospital). The external validation cohort comprised 245 patients (81 BTC and 164 BTD) from Pusan National University Yangsan Hospital. The patient-inclusion flowchart is shown in Fig. 1 . This study was approved by the Ethics Committee of PNUH (IRB number: 2009-006-094, 2011-014-097) and KNUH (IRB number: 2020-04-045). Dataset feature selection A total of 692 data features were collected, encompassing a wide range of clinical domains: demographic characteristics; lifestyle factors and family history; medical history (including prior surgeries and medication use); laboratory test results; infectious disease markers (e.g., hepatitis B and C virus, Clonorchis sinensis ); imaging findings; and operative and pathological reports. Data acquisition was standardized across institutions using a unified survey protocol to minimize inter-institutional variability and reduce systematic bias. Under the guidance of clinical experts, we performed feature preprocessing for relevance and interpretability, yielding a selected panel of 57 variables—comprising laboratory biomarkers and clinical characteristics. Additional baseline characteristics not included in the final analytic panel—particularly demographic characteristics, lifestyle factors, and family history—are presented in Supplementary Table 3. The detailed criteria of feature selection are provided in the Supplementary Methods. To ensure model training quality, we applied a 40% missingness threshold for feature exclusion and exceptions were made for clinically critical features (e.g., CA19-9 and CEA), which were retained despite exceeding this threshold based on expert consultation. Remaining missing values within the included features were imputed using the median value across non-missing entries. Furthermore, to improve predictive accuracy and interpretability, model development was preceded by a structured feature-reduction process. From an initial pool of 57 candidate clinical variables, univariate chi‐square tests identified 26 variables with the strongest associations to the diagnostic outcome (Supplementary Table 2); notably, CA19-9 was deliberately excluded from this final set. The BTC risk prediction model was therefore trained without CA19-9, and its impact was subsequently assessed via sensitivity analysis by reintroducing CA19-9 to evaluate changes in diagnostic performance. Machine learning model for discriminating BTC from BTD To distinguish BTC from BTD, we developed a machine learning model based on Tabular Prior-Data Fitted Networks (TabPFN) [ 16 ]. For internal validation, a Monte Carlo cross-validation framework was employed, in which the training cohort was repeatedly and randomly divided into internal training and validation subsets across 50 iterations to assess model performance stability. For external validation, the model was trained on the combined training cohort of the three centers and evaluated using the held-out center cohort from the independent center. Risk factor analysis of BTC using interpretable machine learning SHapley Additive exPlanations (SHAP) [ 17 ] were applied to identify clinical variables that influenced the model's prediction toward BTC or BTD. To further enhance the clinical interpretability of our prediction model, a surrogate decision tree was trained to approximate the behavior of the TabPFN model. This decision tree–based surrogate model distilled its logic into a small number of binary clinical rules, which could be further developed into a clinical trial protocol. Despite its reduced complexity, the surrogate maintained high concordance with the original model’s predictions and replicated key decision boundaries using clinically recognizable thresholds. Several rule-based explanations were derived, reflecting the behavior of the full model in identifying high-risk BTC cases—including those with non-elevated CA19-9 levels—highlighting its potential utility in diagnostically challenging scenarios. Results Cohort Characteristics The detailed characteristics of the training and external validation cohort are presented in Table 1 . Patients with BTC were significantly older than those with BTD in both the training (median age: 73 vs. 66 years; P < .001) and external validation cohorts (73 vs. 62.5 years; P < .001). Across both cohorts, BTC was consistently associated with elevated serum levels of CA19-9 and CEA, as well as markers indicative of hepatobiliary dysfunction, including total and direct bilirubin, alkaline phosphatase (ALP), and gamma-glutamyl transferase (GGT) (all P < .001). BTC patients also exhibited lower hemoglobin, hematocrit, and serum albumin concentrations relative to BTD patients, consistent with systemic inflammation and hepatic dysfunction. Increases in aminotransferases (AST and ALT) and prolonged activated partial thromboplastin time (aPTT) were likewise observed in the BTC groups. While some variables—such as weight, phosphorus, and uric acid—demonstrated inter-cohort variability in their statistical significance, the overall direction and magnitude of key BTC-associated biomarkers remained consistent between the training and external validation cohorts. The distribution of BTC anatomical subtypes is summarized in Supplementary Table 4. Among patients diagnosed with BTC, 118 (21.5%) had intrahepatic cholangiocarcinoma (iCCA), 145 (26.5%) distal cholangiocarcinoma (dCCA), 160 (29.2%) perihilar cholangiocarcinoma (pCCA), and 161 (29.4%) gallbladder cancer (GBC). Subtype information was unavailable for 57 patients due to incomplete responses in the PHR dataset. Additionally, 12 patients were identified as having multiple concurrent BTC subtypes. Table 1 Cohort characteristics of patients in the PHR dataset Training Cohort External Validation Cohort Clinical variables BTD (n = 891) BTC (n = 548) p-value BTD (n = 164) BTC (n = 81) p-value Age, y 66.0 (56.0–75.0) 73.0 (66.0–79.0) < 0.001 62.50 (48.0–74.0) 73.0 (67.0–77.0) < 0.001 Height, cm 163.0 (156.15–169.90) 161.0 (154.0–167.0) < 0.001 162.50 (156.90–169.0) 162.0 (154.0–167.0) 0.221 Weight, kg 63.0 (56.0–71.35) 59.15 (53.0–66.0) < 0.001 63.0 (54.0–73.0) 59.40 (53.0–64.0) 0.009 Waist, cm 32.0 (30.0–34.0) 32.0 (30.0–34.0) 0.238 32.0 (30.0–32.12) 32.0 (30.0–32.0) 0.623 Male 493 (55.3%) 323 (58.9%) 0.198 82 (50.0%) 46 (56.8%) 0.387 Diabetes 216 (24.2%) 163 (29.7%) 0.025 33 (20.1%) 19 (23.5%) 0.664 Alcohol Usage, n (%) < 0.001 < 0.001 Current 318 (35.7%) 131 (23.9%) 62 (37.8%) 15 (18.5%) Former 198 (22.2%) 167 (30.5%) 35 (21.3%) 35 (43.2%) Never 375 (42.1%) 250 (45.6%) 67 (40.9%) 31 (38.3%) Lab Results ANC, x10⁹/L 4.60 (3.29–6.97) 4.60 (3.44–6.64) 0.527 4.39 (3.08–7.28) 4.59 (3.58–7.36) 0.456 Hemoglobin, g/dL 13.10 (12.0–14.30) 12.60 (11.30–13.70) < 0.001 13.10 (12.30–14.40) 12.40 (11.10–13.60) < 0.001 Hematocrit, % 38.90 (35.90–42.30) 37.20 (33.60–40.60) < 0.001 38.10 (35.48–41.20) 35.80 (31.40–40.10) < 0.001 Platelets, x10⁹/L 222.0 (176.0–267.50) 248.50 (197.0–314.0) < 0.001 226.50 (186.75–271.0) 240.0 (185.0–301.0) 0.247 ESR, mm/hour 33.0 (24.0–33.0) 33.0 (33.0–68.0) < 0.001 33.0 (33.0–33.0) 33.0 (33.0–33.0) 0.076 CRP, mg/dL 0.55 (0.10–2.13) 0.82 (0.26–3.34) < 0.001 0.63 (0.60–2.95) 0.81 (0.45–2.58) 0.382 Total Protein, g/dL 7.0 (6.58–7.33) 6.90 (6.50–7.29) 0.016 7.20 (6.88–7.60) 7.10 (6.50–7.40) 0.012 Albumin, g/dL 4.25 (3.97–4.50) 4.05 (3.68–4.30) < 0.001 4.20 (3.90–4.43) 3.80 (3.30–4.30) < 0.001 Total Bilirubin, g/dL 0.80 (0.46–1.79) 1.40 (0.59–9.55) < 0.001 1.0 (0.70–2.40) 1.80 (0.60–11.0) 0.007 Direct Bilirubin, mg/dL 0.74 (0.30–1.0) 0.74 (0.32–7.04) < 0.001 0.74 (0.35–0.74) 0.74 (0.74–6.56) 0.002 AST, U/L 30.0 (21.0–118.50) 67.0 (28.0–145.0) < 0.001 38.50 (22.0–109.0) 61.0 (29.0–173.0) 0.004 ALT, U/L 28.0 (17.0–128.50) 58.50 (20.0–166.0) < 0.001 41.0 (17.75–116.25) 42.0 (20.0–198.0) 0.189 ALP, U/L 90.0 (66.0–161.50) 259.0 (102.75–501.0) < 0.001 110.50 (75.0–221.0) 348.0 (125.0–521.0) < 0.001 GGT, U/L 173.50 (31.50–281.0) 322.50 (102.25–766.25) < 0.001 173.50 (33.0–285.25) 248.0 (106.0–914.0) < 0.001 Phosphorus, mg/dL 3.24 (2.80–3.51) 3.24 (2.90–3.67) 0.057 3.50 (3.0–3.90) 3.20 (2.80–3.60) 0.034 Uric Acid, mg/dL 4.46 (3.98–5.40) 4.40 (3.30–5.10) < 0.001 4.80 (3.80–6.03) 4.30 (3.20–5.60) 0.022 BUN, mg/dL 14.30 (11.30–18.60) 14.80 (11.67–18.62) 0.146 14.55 (12.47–19.0) 14.90 (11.80–21.80) 0.418 Creatinine, mg/dL 0.81 (0.66–0.97) 0.80 (0.67–0.99) 0.762 0.80 (0.64–0.95) 0.85 (0.69–0.98) 0.161 CA19-9, U/mL 12.75 (6.44–25.69) 93.59 (20.72–712.10) < 0.001 14.50 (5.85–45.30) 76.90 (7.05–504.17) < 0.001 CEA, ng/mL 2.30 (1.98–2.30) 2.48 (2.0–4.51) < 0.001 2.30 (1.60–2.70) 2.30 (2.30–4.12) < 0.001 PT, s 12.0 (11.30–12.40) 12.0 (11.38–12.70) 0.068 12.80 (12.30–13.53) 12.40 (11.90–13.10) 0.002 APTT, s 27.20 (25.30–27.90) 27.50 (25.60–30.42) < 0.001 34.60 (31.60–37.12) 33.30 (29.30–35.90) 0.018 Glucose, mg/dL 121.0 (104.0–141.0) 121.0 (107.0–144.25) 0.180 121.0 (121.0–147.75) 121.0 (104.0–145.0) 0.086 HbA1c, % 5.80 (5.80–5.80) 5.80 (5.70–6.10) 0.636 5.80 (5.70–5.80) 5.80 (5.80–5.80) 0.377 HBsAg, n (%) 0.409 0.065 Positive 22 (2.5%) 16 (2.9%) 3 (1.8%) 4 (4.9%) Negative 856 (96.1%) 528 (96.4%) 31 (79.9%) 70 (86.4%) HBsAb, n (%) < 0.001 0.148 Positive 667 (74.9%) 363 (66.2%) 92 (56.1%) 53 (65.4%) Negative 212 (23.8%) 180 (32.8%) 43 (26.2%) 21 (25.9%) HCVAb, n (%) 0.109 0.191 Positive 10 (1.1%) 10 (1.8%) 5 (3.0%) 1 (1.2%) Negative 864 (97.0%) 534 (97.4%) 141 (86.0%) 76 (93.8%) ANC, absolute neutrophil count; ESR, erythrocyte sedimentation rate; CRP, c-reactive protein; AST, alanine aminotransferase; ALT, alanine aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyl transferase; BUN, blood urea nitrogen; CA19-9, carbohydrate antigen 19 − 9; CEA, carcinoembryonic antigen; PT, prothrombin time; APTT, activated partial thromboplastin time; HbA1c, hemoglobin a1c; HBsAg, hepatitis b surface antigen; HBsAb, hepatitis b surface antibody; HCVAb, hepatitis c virus antibody. Data are presented as median with interquartile range (IQR), defined by the first (Q1) and third (Q3) quartiles. Categorical variables are expressed as counts and percentages, n (%). Performance evaluation of our prediction model for distinguishing BTC from BTD Using 26 preselected clinical variables (Supplementary Table 2), the TabPFN model demonstrated superior performance compared with other machine learning algorithms (Supplementary Table 5). Receiver operating characteristic (ROC) analyses confirmed its higher discriminative ability (Fig. 2 a–b), while calibration plots showed close concordance between predicted probabilities and observed outcomes, underscoring the reliability of its risk estimates (Fig. 2 c–d). We further evaluated the contribution of CA19-9, a widely used biomarker for BTC, to the model’s ability to discriminate between BTC and BTD. Two TabPFN models were compared: one including CA19-9, trained on patients with available measurements (n = 885), and another excluding CA19-9, trained on the full cohort (n = 1,439). Unexpectedly, the model without CA19-9 achieved comparable predictive accuracy (see Table 2 ). Given the clinical need for applicability in Lewis antigen–negative patients and in those lacking CA19-9 measurements, we excluded CA19‑9 from the model development. Table 2 Comparison of diagnostic performances between TabPFN with and without the biomarker CA19-9, and the performance of CA19-9 with its two cutoffs. Model AUROC (95% CI) Sensitivity (95% CI) Specificity (95% CI) PPV (95% CI) NPV (95% CI) Internal Validation TabPFN with CA19-9 0.895 (0.85–0.93) 0.824 (0.75–0.90) 0.772 (0.68–0.87) 0.822 (0.77–0.89) 0.777 (0.70–0.86) TabPFN with CA19-9 (OC: 0.628) 0.895 (0.85–0.93) 0.731 (0.66–0.81) 0.867 (0.76–0.95) 0.876 (0.81–0.94) 0.719 (0.66–0.79) TabPFN without CA19-9 0.893 (0.86–0.92) 0.682 (0.59–0.75) 0.910 (0.87–0.94) 0.825 (0.77–0.89) 0.823 (0.78–0.86) TabPFN without CA19-9 (OC: 0.359) 0.893 (0.86–0.92) 0.784 (0.73–0.83) 0.839 (0.79–0.89) 0.752 (0.69–0.82) 0.863 (0.83–0.89) CA19-9 > 37 0.790 (0.76–0.82) 0.640 (0.60–0.68) 0.817 (0.78–0.85) 0.817 (0.78–0.85) 0.639 (0.60–0.68) CA19-9 > 27.7 0.790 (0.76–0.82) 0.690 (0.65–0.73) 0.778 (0.73–0.82) 0.800 (0.76–0.84) 0.662 (0.62–0.70) External Validation TabPFN with CA19-9 0.801 (0.73–0.86) 0.727 (0.63–0.83) 0.643 (0.66–0.80) 0.505 (0.48–0.67) 0.825 (0.79–0.90) TabPFN with CA19-9 (OC: 0.628) 0.801 (0.73–0.86) 0.635 (0.52–0.75) 0.809 (0.74–0.87) 0.625 (0.51–0.74) 0.816 (0.75–0.88) TabPFN without CA19-9 0.855 (0.81–0.90) 0.744 (0.65–0.83) 0.767 (0.70–0.83) 0.612 (0.51–0.70) 0.858 (0.80–0.92) TabPFN without CA19-9 (OC: 0.359) 0.855 (0.81–0.90) 0.829 (0.74–0.90) 0.616 (0.54–0.68) 0.516 (0.43–0.60) 0.879 (0.82–0.94) CA19-9 > 37 0.649 (0.56–0.73) 0.576 (0.46–0.69) 0.727 (0.65–0.80) 0.514 (0.40–0.62) 0.774 (0.70–0.85) CA19-9 > 27.7 0.649 (0.56–0.73) 0.606 (0.48–0.72) 0.674 (0.59–0.75) 0.482 (0.38–0.58) 0.774 (0.70–0.85) AUROC, area under the receiver operating characteristic curve; CI, confidence interval; OC, Youden-index–optimized cutoff; PPV, positive predictive value; NPV, negative predictive value. TabPFN with CA 19 − 9: TabPFN model trained including CA 19 − 9 as a feature, TabPFN without CA 19 − 9: TabPFN model trained excluding CA 19 − 9. “(OC: x)” indicates performance at each model’s Youden-index–derived cutoff: x, rather than the pre‐specified cutoff 0.5. Our final model achieved an AUROC of 0.893 (95% CI, 0.863–0.920) and a sensitivity of 0.784 (95% CI, 0.727–0.827) for detecting BTC (Youden index 0.359). When applied to the fourth center's cohort — which had been excluded from model training — the diagnostic model maintained strong performance (AUROC, 0.855; 95% CI, 0.805–0.899), identifying over 82% of BTC cases (sensitivity, 0.829; 95% CI, 0.743–0.904) (Youden index 0.359) (Table 2 ). Notably, the model demonstrated substantially higher sensitivity for detecting BTC than the conventional tumor biomarker CA19-9 (cutoff of > 37 U/L), both in the internal cohort (78.4% vs. 64.0%) and in external validation (82.9% vs. 57.6%). Even when applying a lower CA19-9 threshold (27.7 U/L, Youden index), the model retained superior sensitivity in both internal (78.4% vs. 69.0%) and external cohorts (82.9% vs. 60.6%). The difference was substantial across other diagnostic metrics, as shown in Table 2 . Subgroup analysis by CA19-9 Level We further evaluated the diagnostic performance of the TabPFN model in subgroups stratified by CA19-9 levels: normal ( 100 U/mL). We selected 37 U/mL as the lower boundary because it represents the established upper limit of normal for CA19-9 [ 18 ]. The 100 U/mL threshold was chosen as the upper boundary based on evidence showing that values below this level are frequently observed in benign biliary obstruction, whereas values exceeding 100 U/mL are reported to significantly increase the likelihood of malignancy [ 7 , 19 ]. In internal validation, AUROCs were similar across subgroups—0.819 (95% CI, 0.806–0.831), 0.821 (95% CI, 0.792–0.847), and 0.831 (95% CI, 0.813–0.848), respectively—indicating that the model’s discrimination was preserved regardless of CA19-9 elevation. External validation showed similarly robust results (Fig. 2 e–f, Supplementary Table 5). Applying subgroup-specific cutoffs to balance sensitivity and specificity yielded clinically relevant trade-offs. In the normal CA19-9 group, lowering the probability threshold to 0.304 increased sensitivity to 85.7% (95% CI, 71.4–96.6%), enabling the model to correctly identify the majority of BTC cases that would otherwise be missed by CA19-9 alone, at the cost of reduced specificity (52.1%). In the indeterminate group (37–100 U/mL), a cutoff of 0.372 achieved perfect sensitivity (100%), ensuring that no BTC cases were missed, while maintaining moderate specificity (66.7%) (Table 3 ). Table 3 Comparison of the TabPFN model diagnostic performances on different CA19-9 level subgroups (~ 37, 37 ~ 100, 100~) for both internal and external validation cohorts. Subgroups of CA19-9 levels AUROC (95% CI) Sensitivity (95% CI) Specificity (95% CI) PPV (95% CI) NPV (95% CI) Internal Validation ~ 37 0.819 (0.81–0.83) 0.540 (0.52–0.56) 0.882 (0.87–0.89) 0.726 (0.70–0.75) 0.768 (0.75–0.78) ~ 37 (OC: 0.304) 0.819 (0.81–0.83) 0.750 (0.73–0.77) 0.720 (0.71–0.74) 0.608 (0.59–0.63) 0.833 (0.82–0.85) 37 ~ 100 0.821 (0.79–0.85) 0.721 (0.69–0.75) 0.761 (0.71–0.81) 0.873 (0.85–0.90) 0.546 (0.50–0.59) 37 ~ 100 (OC: 0.372) 0.821 (0.79–0.85) 0.816 (0.79–0.84) 0.654 (0.60–0.71) 0.843 (0.82–0.87) 0.610 (0.56–0.66) 100~ 0.831 (0.81–0.85) 0.838 (0.82–0.85) 0.503 (0.45–0.56) 0.922 (0.91–0.93) 0.308 (0.27–0.35) 100~ (OC: 0.709) 0.831 (0.81–0.85) 0.719 (0.70–0.74) 0.794 (0.75–0.84) 0.960 (0.95–0.97) 0.289 (0.26–0.32) External Validation ~ 37 0.731 (0.64–0.82) 0.500 (0.32–0.70) 0.750 (0.66–0.84) 0.368 (0.22–0.53) 0.837 (0.76–0.91) ~ 37 (OC: 0.304) 0.731 (0.64–0.82) 0.857 (0.71–0.97) 0.521 (0.42–0.62) 0.343 (0.23–0.46) 0.926 (0.85–0.98) 37 ~ 100 0.907 (0.75–1.00) 0.833 (0.50–1.00) 0.778 (0.58–0.95) 0.556 (0.20–0.89) 0.933 (0.79–1.00) 37 ~ 100 (OC: 0.372) 0.907 (0.75–1.00) 1.000 (1.00–1.00) 0.667 (0.44–0.88) 0.500 (0.20–0.79) 1.000 (1.00–1.00) 100~ 0.821 (0.69–0.93) 0.875 (0.75–0.97) 0.500 (0.27–0.74) 0.757 (0.61–0.89) 0.692 (0.42–0.92) 100~ (OC: 0.709) 0.821 (0.69–0.93) 0.750 (0.59–0.89) 0.778 (0.57–0.95) 0.857 (0.72–0.97) 0.636 (0.43–0.83) AUROC, area under the receiver operating characteristic curve; CI, confidence interval; OC, Youden-index–optimized cutoff; PPV, positive predictive value; NPV, negative predictive value. “(OC: x)” indicates performance at each model’s Youden-index–derived cutoff: x, rather than the pre‐specified cutoff 0.5. These findings highlight the model’s potential utility as a complementary diagnostic tool, particularly for patients with normal or borderline CA19-9 levels, where reliance on CA19-9 alone may fail to detect early disease. Serum Biomarker Contribution to BTC Risk We next applied SHapley Additive exPlanations (SHAP) to quantify the impact of each serum biomarker on the TabPFN model’s prediction of BTC risk. In the training cohort (Fig. 4 a) and external validation cohort (Fig. 4 b), carcinoembryonic antigen (CEA) and alkaline phosphatase (ALP) emerged as the two strongest predictors. Notably, activated partial thromboplastin time (aPTT) also ranked among the top contributors in the external validation set, while the remaining biomarkers demonstrated consistent importance across both cohorts. To illustrate how individual biomarker values influence the BTC risk score, patient-level biomarker contributions are depicted in Supplementary Fig. 1. Using SHAP analyses, we further evaluated biomarker contributions for each anatomical BTC subtype in the training cohort. Gallbladder cancer and intrahepatic cholangiocarcinoma showed similar biomarker risk profiles, with CEA and ALP emerging as the strongest positive contributors to predicted malignancy. In contrast, among patients with extrahepatic cholangiocarcinoma (perihilar and distal subtypes), ALP was identified as the predominant driver of risk, followed by total bilirubin (Supplementary Fig. 2). Rule-Based Risk Stratification for BTC To better understand our TabPFN model’s predictions and enhance its clinical decision-making utility, we derived a simplified surrogate model—a decision tree—that mimics the model’s approach to BTC detection. Applying this decision tree to both the internal and external validation cohorts stratified patients into high- and low-risk groups, highlighting key biomarker thresholds associated with BTC risk. The decision tree is shown in Supplementary Fig. 3. For patients with markedly elevated alkaline phosphatase (ALP > 311.5 U/L), the model predicted a high probability of BTC (~ 77.8%). Among the high-ALP patients the presence of extreme hyperbilirubinemia (total bilirubin > 6.02 mg/dL) further increased the predicted BTC risk to 94.4%. Notably, when ALP levels were in the range of 311.5–490.5 U/L and total bilirubin was only moderately elevated (1.15–6.02 mg/dL), the model assigned a relatively lower BTC risk (34.9%). In contrast, patients with similarly high ALP but near-normal bilirubin (≤ 1.15 mg/dL) exhibited a substantially higher predicted risk (~ 73.0%). This seemingly paradoxical pattern may reflect different disease presentations: intrahepatic cholangiocarcinoma can elevate ALP while bilirubin remains mildly increased or normal, whereas moderate bilirubin elevations in the context of high ALP might sometimes indicate benign partial obstructions [ 20 ]. For patients with lower ALP levels (≤ 311.5 U/L), the decision tree identified serum carcinoembryonic antigen (CEA) as the next most informative splitter. Those with CEA above 3.31 ng/mL formed a higher-risk subgroup, with a predicted BTC risk of ~ 56.7%, and if CEA exceeded 9.69 ng/mL the risk rose to 86.4%. In the subset of patients with moderately elevated CEA (3.31–9.69 ng/mL), a concurrent total bilirubin > 5.28 mg/dL further boosted the BTC risk to 84.8%. For patients with relatively benign biomarker profiles (ALP ≤ 311.5, CEA ≤ 3.31), age emerged as an important stratifier in the decision tree. Specifically, age > 56.5 years delineated a higher-risk branch. In this older subgroup with otherwise unremarkable markers, an isolated marked hyperbilirubinemia (> 8.90 mg/dL) still conferred a high predicted risk of BTC (73.5%). This emphasizes that in an older patient, the presence of significant jaundice alone is a red flag, even if other markers like ALP or CEA are normal. Conversely, if total bilirubin remained below 8.90 mg/dL in these older patients, the model identified an intermediate-risk scenario (≈ 57% BTC probability) only when subtle coagulation changes were present – specifically, an elevated activated partial thromboplastin time (aPTT > 30.95 s) alongside a normal direct bilirubin (≤ 0.27 mg/dL). Discussion In this study, we developed a machine learning–based model to distinguish biliary tract cancer (BTC) from benign biliary tract disease (BTD) using routinely available demographic, clinical and laboratory variables, with an emphasis on seamless integration into everyday workflows. The model showed consistently high discrimination in both internal and independent external cohorts, including clinically challenging subgroups with low CA19-9 levels. Notably, adding CA19-9 conferred only modest incremental benefit, and performance remained robust when CA19-9 was excluded—supporting use in Lewis antigen–negative patients and in settings where CA19-9 is unavailable. In head-to-head comparisons, the model markedly outperformed CA19-9 alone, underscoring the limitations of CA19-9–dependent screening pathways. Prior work has demonstrated promising BTC/CCA prediction across heterogeneous populations using conventional statistical models [ 21 – 23 ]. Hu et al [ 21 ] employed a multivariate linear Cox proportional hazards (CoxPH) model to screen for cholangiocarcinoma (CCA) in a cohort of 1,459 primary sclerosing cholangitis (PSC) patients from the Mayo Clinic, yielding a concordance index(C-index) of 0.69. T. Meister et al. [ 22 ] and X. Shen et al. [ 23 ] both conducted cholangiocarcinoma (CCA) screening using multivariate logistic regression analysis, each applied to the two cohorts: 1,135 patients with bile duct obstruction and 2,269 patients with intrahepatic lymphocytes. Each studies yielded validation performances with area under the receiver operating characteristic curve (AUROC) values of 0.862 and 0.867, respectively. While direct cross-study comparisons are limited by differences in inclusion criteria and biomarker panels, our study advances this literature by encompassing a broader spectrum of malignant and benign biliary conditions and by demonstrating generalizability in an independent external cohort. Using only routine clinical and laboratory variables, the model maintains high discrimination across CA19-9 strata—including patients with normal CA19-9—addressing a key limitation of CA19-9–centered strategies. Moreover, clinician-facing explanations (SHAP) and a compact surrogate decision tree provide transparent thresholds that facilitate implementation without sacrificing performance. Beyond performance, our model’s risk factors and rule-based pathogenesis of BTC support the model’s clinical validity. Among the top contributing biomarkers of our model, ALP and GGT are known for indicators of cholestasis and biliary obstruction, conditions often preceding BTC development [ 7 ]. Our findings of highly elevated ALP associated with BTC is consistent with the cholestatic laboratory profile often seen in biliary malignancies, where ALP levels typically rise multiple-fold above normal [ 20 ]. Bilirubin elevations are known to reflect obstructive processes and liver injury. In our study, we found high bilirubin levels with a high association to BTC, while near-normal bilirubin levels could also pose a risk in some clusters of patients. Overall, the combination of ALP and bilirubin in the decision tree underscores their importance; indeed, extrahepatic BTC classically presents with conjugated bilirubin and ALP elevated on the order of 2–10 times the upper normal limit in other studies [ 20 ]. Inflammatory markers such as CRP and ESR align with the known association between chronic inflammation and biliary carcinogenesis [ 24 , 25 ]. These features are particularly relevant in diseases such as PSC, recurrent cholangitis, or choledochal cysts, where chronic biliary inflammation increases cancer risk [ 26 , 27 ]. Coagulation and liver function parameters—such as prolonged APTT, hypoalbuminemia, and thrombocytopenia—suggest impaired hepatic reserve and portal hypertension, common in advanced liver disease, a known risk factor for BTC [28,29. Viral hepatitis markers (HBsAg, HCV Ab) were also highly ranked, consistent with their known role in hepatobiliary malignancies, particularly intrahepatic cholangiocarcinoma [ 30 ] Additionally, the model identified metabolic indicators—fatty liver and elevated uric acid—as BTC-associated features. These may reflect underlying NAFLD or metabolic syndrome, which have been linked to biliary malignancies through inflammatory and insulin resistance pathways [ 31 ]. Although our findings align with prior work, limitations persist. First, this study was conducted using a demographically specific cohort (South Korean patients). Thereby, the generalizability of the model to populations with different ethnic, geographic, or healthcare backgrounds warrants further validation. The development of biliary tract cancer (BTC) in patients with biomarker values falling within normal reference ranges remains insufficiently understood, underscoring a diagnostic challenge that calls for further investigation in future studies. Furthermore, we imputed missing values using per-feature medians; given the assumptions and limitations of single imputation, the findings should be interpreted with caution and additional external validation. Conclusions Using routinely available demographic and clinical data from patients diagnosed with BTC and BTD, we developed and validated a machine learning model capable of accurately distinguishing malignant from benign biliary tract disease. Leveraging only routinely available clinical and laboratory data, the model provides a non-invasive screening approach to identify ultra-high–risk patients for BTC within the already high-risk biliary disease population. Its reliance on standard data sources facilitates integration into electronic health record systems, where it could support earlier detection and targeted surveillance in routine clinical practice. Declarations 1. Ethics approval and consent to participate This study was approved by the Institutional Review Board (Ethics Committee) of Pusan National University Hospital (IRB No. 2009-006-094 and 2011-014-097) and the Institutional Review Board (Ethics Committee) of Kyungpook National University Hospital (IRB No. 2020-04-045). Informed consent to participate was obtained from all of the participants. All procedures were conducted in accordance with the Declaration of Helsinki and its later amendments. 2. Consent for publication Not applicable 3. Availability of data and materials The data that support the findings of this study are available from Pusan National University Hospital, but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Pusan National University Hospital. 4. Competing interests The authors declare that they have no competing interests 5. Funding This study was supported by a grant (SMF-AI-EJ003-2023) funded by Seegene Medical Foundation, Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (IITP-2025-RS-2025-02219116), Institute of Information & Communications Technology Planning & Evaluation (IITP) under the Artificial Intelligence Convergence Innovation Human Resources Development (IITP-202300254177), the National R&D Program for Cancer Control, Ministry of Health and Welfare, Republic of Korea (HA20C0009), and Biomedical Research Institute Grant (202200140001) Pusan National University Hospital. 6. Authors' contributions Study concept and design: Hanul Park, Dong Uk Kim, Giltae Song Data acquisition: all authors Data analysis and interpretation: Hanul Park, Dong Uk Kim, Giltae Song Study supervision: Dong Uk Kim, Giltae Song Administrative, technical, or material support: all authors Approval of final manuscript: all authors 7. Acknowledgements Not applicable References Vogel A, Bridgewater J, Edeline J, et al. Biliary tract cancer: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol. 2023;34(2):127–140. Bray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68(6):394–424. Florio AA, Ferlay J, Znaor A, et al. Global trends in intrahepatic and extrahepatic cholangiocarcinoma incidence from 1993 to 2012. Cancer. 2020;126(11):2666–2678. Valle JW, Kelley RK, Nervi B, et al. Biliary tract cancer. Lancet. 2021;397(10272):428–444. Clements O, Eliahoo J, Kim JU, et al. Risk factors for intrahepatic and extrahepatic cholangiocarcinoma: A systematic review and meta-analysis. J Hepatol. 2020;72(1):95–103. Razumilava N, Gores GJ. Classification, diagnosis, and management of cholangiocarcinoma. Clin Gastroenterol Hepatol. 2013;11(1):13–21. Ilyas SI, Gores GJ. Pathogenesis, diagnosis, and management of cholangiocarcinoma. Gastroenterology. 2013;145(6):1215–1229. Blechacz B, Komuta M, Roskams T, et al. Clinical diagnosis and staging of cholangiocarcinoma. Nat Rev Gastroenterol Hepatol. 2011;8(9):512–522. Miros M, Kerlin P, Walker N, et al. Predicting cholangiocarcinoma in patients with primary sclerosing cholangitis before transplantation. Gut. 1991;32(11):1369–1373. Khaderi SA, Sussman NL. Screening for malignancy in primary sclerosing cholangitis (PSC). Curr Gastroenterol Rep. 2015;17:1–5. Hu C, Iyer RK, Juran BD, et al. Predicting cholangiocarcinoma in primary sclerosing cholangitis: using artificial intelligence, clinical and laboratory data. BMC Gastroenterol. 2023;23(1):129. J Huang, X Bai, Y Qiu, et al. Application of AI on cholangiocarcinoma. Frontiers in Oncology. 2024;14. Zhou SN, Jv DW, Meng XF, et al. Feasibility of machine learning-based modeling and prediction using multiple centers data to assess intrahepatic cholangiocarcinoma outcomes. Ann Med. 2023;55(1):215–223. Lv Y, Liu H, He P, et al. A novel model for predicting the prognosis of postoperative intrahepatic cholangiocarcinoma patients. Sci Rep. 2023;13(1):19267. CX Chamberlain, E Faust, D Goldschmidt, et al. Burden of illness for patients with cholangiocarcinoma in the United States: a retrospective claims analysis. Journal of gastrointestinal oncology. 2021;12(2):658 Hollmann, N., Müller, S., Purucker, L. et al. Accurate predictions on small data with a tabular foundation model. Nature. 2025;637(8045):319–326 Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30. Zhang Y, Yang J, Li H, et al. Tumor markers CA19-9, CA242 and CEA in the diagnosis of pancreatic cancer: a meta-analysis. International journal of clinical and experimental medicine. 2015;8(7):11683 Liang B, Zhong L, He Q, et al. Diagnostic accuracy of serum CA19-9 in patients with cholangiocarcinoma: a systematic review and meta-analysis. Medical science monitor: international medical journal of experimental and clinical research. 2015;21:3555 Forner A, Vidili G, Rengo M, et al. Clinical presentation, diagnosis and staging of cholangiocarcinoma. Liver International. 2019;39:98-107 Hu C, Iyer RK, Juran BD, et al. Predicting cholangiocarcinoma in primary sclerosing cholangitis: using artificial intelligence, clinical and laboratory data. BMC Gastroenterol. 2023;23(1):129. Meister T, Uphoff MA, Heinecke A, et al. Novel score for prediction of malignant bile duct obstruction based on biochemical and clinical markers. Aliment Pharmacol Ther. 2015;41(9):877–887. Shen X, Zhao H, Jin X, et al. Development and validation of a machine learning-based nomogram for prediction of intrahepatic cholangiocarcinoma in patients with intrahepatic lithiasis. Hepatobiliary Surg Nutr. 2021;10(6):749-765. Nam YH, Park MS, Lee SM. Preoperative C-reactive protein as a prognostic factor for recurrence after surgical resection of biliary tract cancer. Korean J Clin Oncol. 2015;11(2):101–105. Gerhardt T, Milz S, Schepke M, et al. C-reactive protein is a prognostic indicator in patients with perihilar cholangiocarcinoma. World J Gastroenterol. 2006;12(34):5495–5500. Ilyas SI, Eaton JE, Gores GJ, et al. Primary sclerosing cholangitis as a premalignant biliary tract disease: surveillance and management. Clinical Gastroenterology and Hepatology. 2015;13(12):2152-2165 Labib PL, Goodchild G, Pereira SP, et al. Molecular pathogenesis of cholangiocarcinoma. BMC cancer. 2019;19:1-16 Ke X, Jin B, You, W, et al. Comprehensive analysis of coagulation indices for predicting survival in patients with biliary tract cancer. BMC cancer. 2021;21:1–15. Poordad F. Review article: thrombocytopenia in chronic liver disease. Alimentary Pharmacology & Therapeutics. 2007;26:5-11 Zhou Y, Zhao Y, Li B, et al. Hepatitis viruses infection and risk of intrahepatic cholangiocarcinoma: evidence from a meta-analysis. BMC cancer. 2012;12:1–7. Wongjarupong N, Assavapongpaiboon B, Susantitaphong P, et al. Non-alcoholic fatty liver disease as a risk factor for cholangiocarcinoma: a systematic review and meta-analysis. BMC gastroenterology. 2017;17:1–8. Additional Declarations No competing interests reported. Supplementary Files SupplementaryTables15.docx SupplementaryMethods.docx SupplementaryFigure1.png SupplementaryFigure2.png SupplementaryFigure3.png Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8736730","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":587661204,"identity":"5e30c244-87d3-4e45-8897-389e1354c0c3","order_by":0,"name":"Hanul Park","email":"","orcid":"","institution":"Pusan National University","correspondingAuthor":false,"prefix":"","firstName":"Hanul","middleName":"","lastName":"Park","suffix":""},{"id":587661205,"identity":"a38da135-d673-4a48-b316-c510534381a3","order_by":1,"name":"Kibeom Kim","email":"","orcid":"","institution":"Pusan National University","correspondingAuthor":false,"prefix":"","firstName":"Kibeom","middleName":"","lastName":"Kim","suffix":""},{"id":587661206,"identity":"b90112d6-2d07-414a-bdc0-612ae0377b36","order_by":2,"name":"Jonghyun Lee","email":"","orcid":"","institution":"Pusan National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jonghyun","middleName":"","lastName":"Lee","suffix":""},{"id":587661207,"identity":"9dacbf85-8744-4549-b33d-3bc0ebd8df3f","order_by":3,"name":"Sung Yong Han","email":"","orcid":"","institution":"Pusan National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sung","middleName":"Yong","lastName":"Han","suffix":""},{"id":587661208,"identity":"5ef2a1f7-7084-48fb-afc8-eddf392b3157","order_by":4,"name":"Chang Min Cho","email":"","orcid":"","institution":"Kyungpook National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chang","middleName":"Min","lastName":"Cho","suffix":""},{"id":587661210,"identity":"c6dc40e7-be2c-4628-a64e-73ce71fd6a71","order_by":5,"name":"Dong Wook Lee","email":"","orcid":"","institution":"Kyungpook National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Dong","middleName":"Wook","lastName":"Lee","suffix":""},{"id":587661212,"identity":"cc47146e-6ae3-48e0-ac79-c884619ec2d5","order_by":6,"name":"Jun Heo","email":"","orcid":"","institution":"Kyungpook National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jun","middleName":"","lastName":"Heo","suffix":""},{"id":587661213,"identity":"694b4291-a10d-4986-828e-c38ccd98bb47","order_by":7,"name":"Min Kyu Jung","email":"","orcid":"","institution":"Kyungpook National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Min","middleName":"Kyu","lastName":"Jung","suffix":""},{"id":587661214,"identity":"a128ae22-c0d6-41da-9ce5-edd8f387c25a","order_by":8,"name":"Hyeong Seok Nam","email":"","orcid":"","institution":"Dong-Eui Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hyeong","middleName":"Seok","lastName":"Nam","suffix":""},{"id":587661220,"identity":"d688aef0-6d6a-4265-a592-b6f975ff21bc","order_by":9,"name":"Jin Kyu Gahm","email":"","orcid":"","institution":"Pusan National University","correspondingAuthor":false,"prefix":"","firstName":"Jin","middleName":"Kyu","lastName":"Gahm","suffix":""},{"id":587661221,"identity":"c87180ff-6ce8-41d7-998b-9702b50c665e","order_by":10,"name":"Giltae Song","email":"","orcid":"","institution":"Pusan National University","correspondingAuthor":false,"prefix":"","firstName":"Giltae","middleName":"","lastName":"Song","suffix":""},{"id":587661225,"identity":"3084e961-863a-4785-99be-7c1bc9359f80","order_by":11,"name":"Dong Uk Kim","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwklEQVRIiWNgGAWjYDCCM4wNzH94bBgMINwE4rQw8MikkaQFiHlsDpOghe/M4dYNEjnn5c3ZDzB++MGQlk9Qi+TZxrYbBmduG+7sSWCW7GHIsWwgpMXgPGPbjcSe2wkGBxIYpBkYKgwI2gLWcvDfuQSD8w+YfxOnBeiwmw08BxIMbiSwAW3JIaxF8szBttsMPMmGG248bLPsMUgjrIXvTPozoBY7eYPzyYdv/KhIJqwFCQDjlIEkDaNgFIyCUTAKcAIAb8U/igrdHC0AAAAASUVORK5CYII=","orcid":"","institution":"Gumi Medical Center, CHA University","correspondingAuthor":true,"prefix":"","firstName":"Dong","middleName":"Uk","lastName":"Kim","suffix":""}],"badges":[],"createdAt":"2026-01-30 03:54:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8736730/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8736730/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102377256,"identity":"e7c3f465-69f8-408c-89fa-35051203b78e","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":408524,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe flow of study cohort inclusion.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/bd39251fcfca9c031ef15b4a.png"},{"id":102377257,"identity":"dd14b88b-7e6c-4a0b-8ff5-5f56a7446ffd","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":608873,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDiagnostic performance and calibration of TabPFN, XGBoost, and logistic regression models. \u003c/strong\u003e(a-b) Receiver operating characteristic (ROC) curves comparing the classification performance of TabPFN, XGBoost, and logistic regression in the (a) internal and (b) external validation cohorts. (c-d) Calibration plots assessing the agreement between predicted probabilities and observed outcomes for each model in the (c) internal and (d) external validation cohorts. (e-f) Subgroup ROC analyses of the TabPFN model stratified by CA19-9 levels in the (e) internal and (f) external validation cohorts. Subgroups include low, intermediate, and high CA19-9 value ranges.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/49d6c1346f64d251691b2a73.png"},{"id":102377263,"identity":"76adc941-8334-43d0-ba90-eb07ae7099e5","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":348096,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparison of diagnostic performance of TabPFN trained including and excluding CA19-9 and CA19-9 itself.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/4e2065b56c9c70ee7160dac0.png"},{"id":102377259,"identity":"e693b417-0828-4e3b-8b53-0e777abd7f04","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":352846,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP summary plots illustrating the contribution of clinical features to BTC diagnosis prediction. \u003c/strong\u003e(a, b) SHAP (SHapley Additive exPlanations) plots displaying the impact of 26 clinical variables on model output, ranked in descending order of importance, for the (a) internal and (b) external validation cohorts. Each point represents an individual patient, with color indicating the feature value (red: high, blue: low) and position along the x-axis reflecting the magnitude and direction of the feature’s influence on the prediction of biliary tract cancer.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/ecc33da1707affccbd768689.png"},{"id":105734959,"identity":"5a1e7f8a-6911-48cf-964e-4dc287e2de89","added_by":"auto","created_at":"2026-03-30 11:48:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3049697,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/b3b3dc8e-669d-4499-9675-febaf2602258.pdf"},{"id":102377255,"identity":"175e2367-4cd3-433a-b2a1-75ded11d755a","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":42887,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTables15.docx","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/2fcc30085933765785b331c1.docx"},{"id":102377262,"identity":"0bf39060-39f8-434a-a3d9-18f0c6f75064","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":16749,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMethods.docx","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/d3d7fd7528b89796f54237ad.docx"},{"id":102377261,"identity":"4beb0586-5880-4709-9943-ace7695bdd01","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":953863,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/b8560833bd738500f46beac4.png"},{"id":102377258,"identity":"52f6143d-2510-4ab0-aa78-950eae5f3bf6","added_by":"auto","created_at":"2026-02-11 05:47:09","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":533476,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/bc564e6c40a62979df29ddf4.png"},{"id":102397760,"identity":"a80c032b-1c48-4885-9c95-dafcf77f0ffa","added_by":"auto","created_at":"2026-02-11 10:19:38","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":1303743,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8736730/v1/4df08f774887243007ad0558.png"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eDevelopment and Validation of a Diagnostic Model for Biliary Tract Cancer Detection in Patients with Benign Biliary Disease: A Multicenter, Retrospective Case-Control Study\u003c/p\u003e","fulltext":[{"header":"Background","content":"\u003cp\u003eBiliary tract cancer (BTC) is a rare malignancy with poor prognosis arising from the gallbladder, biliary tree, or, less commonly, the ampulla of Vater [1]. Although its incidence is relatively low compared to other malignancies, the high mortality rate underscores its clinical significance [2,3]. The high case-fatality rate of BTC is largely attributable to late-stage diagnosis, often resulting from the absence of early symptoms [1,4].\u003c/p\u003e\n\u003cp\u003ePrevious studies investigating the risk factors for BTC have identified several benign biliary tract diseases (BTD), including primary sclerosing cholangitis (PSC), hepatolithiasis, liver fluke infestation (\u003cem\u003eOpisthorchis viverrini\u003c/em\u003e and \u003cem\u003eClonorchis sinensis\u003c/em\u003e), and congenital biliary abnormalities. Beyond sharing common epidemiologic and metabolic risk factors such as chronic inflammation, hepatitis B and C infections, obesity, and diabetes [5-8], BTD itself may serve as a premalignant condition that predisposes patients to subsequent BTC development. However, the accuracy of detecting early-stage BTC in patients with these conditions remains limited as BTC and BTD frequently exhibit indistinguishable symptoms and radiologic features during early disease stages [6-8].\u003c/p\u003e\n\u003cp\u003eSerologic tumor markers, including carbohydrate antigen 19-9 (CA19-9) and carcinoembryonic antigen (CEA), have been proposed as adjunctive tools for BTC risk stratification. However, their diagnostic performance remains suboptimal, particularly in patients with underlying BTD, due to variable sensitivity, limited specificity, and frequent false-positive elevations associated with benign biliary inflammation [4,6\u0026ndash;8]. Consequently, reliance on conventional diagnostic strategies alone is insufficient for identifying early-stage BTC among patients initially diagnosed with BTD.\u003c/p\u003e\n\u003cp\u003eSeveral studies have examined the risk of BTC within specific BTD cohorts\u0026mdash;most notably among patients with PSC [9\u0026ndash;11]. More recently, machine learning algorithms have shown promise in detecting BTC, owing to their ability to capture complex interactions between patient characteristics and disease status [12-14]. Nevertheless, key challenges remain: Risk factors for BTC are still unclear; prior studies often relied on narrow cohorts\u0026mdash;primarily PSC patients\u0026mdash;limiting generalizability;\u0026nbsp;and the overall diagnostic performance of existing tools remains suboptimal.\u003c/p\u003e\n\u003cp\u003eFrom this perspective, there is a critical unmet need for a systematic approach to identify high-risk subpopulations within BTD patients who may harbor occult or early-stage BTC. Standard diagnosis of BTC requires a high index of suspicion and the integration of clinical, laboratory, endoscopic, and radiologic data [6-8]. However, such approaches remain impractical for routine application and are inherently constrained by the asymptomatic nature of early-stage BTC, financial barriers and patient reluctance to undergo advanced imaging modalities and invasive procedures [15]. Consequently, current diagnostic paradigms are insufficient to reliably distinguish early BTC from BTD at the point of initial presentation.\u003c/p\u003e\n\u003cp\u003eAccurate risk stratification at the time of BTD diagnosis represents a pivotal opportunity to enable selective surveillance and timely, targeted diagnostic evaluation, thereby maximizing the potential for early detection and curative intervention. In this context, we aimed to develop a non-invasive, data-driven prediction model to facilitate the early detection of BTC. Using routinely available demographic and clinical data from patients diagnosed with BTD and BTC, we developed and externally validated a machine learning model designed to screen and identify patients at particularly high risk of BTC within the BTD population.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eData source and study cohort\u003c/h2\u003e \u003cp\u003eThis study utilized a hybrid data collection framework in which patient-reported survey data were collected prospectively, while clinical and laboratory variables were retrospectively extracted from electronic medical records. The dataset was curated for the development and external validation of a predictive classification model distinguishing biliary tract cancer (BTC) from benign biliary disease (BTD). BTC subtypes included intrahepatic, perihilar, and distal cholangiocarcinoma, as well as gallbladder cancer. The BTD subtypes are listed in Supplementary Table\u0026nbsp;1.\u003c/p\u003e \u003cp\u003eData were obtained from four academic medical centers in South Korea between July 2020 and September 2022. Of 2,084 patients initially enrolled across these centers, 31 (1.5%) were excluded due to incomplete survey data and an additional 369 (17.7%) were excluded for missing all blood-serum measurements or having over 40% missing features, in order to ensure data completeness and model robustness.\u003c/p\u003e \u003cp\u003eThe final training dataset consisted of 1,439 patients remained in the training cohort\u0026mdash;548 with BTC and 891 with BTD\u0026mdash;drawn from three hospitals (Pusan National University Hospital; Kyungpook National University Hospital; and Kyungpook National University Chilgok Hospital). The external validation cohort comprised 245 patients (81 BTC and 164 BTD) from Pusan National University Yangsan Hospital. The patient-inclusion flowchart is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. This study was approved by the Ethics Committee of PNUH (IRB number: 2009-006-094, 2011-014-097) and KNUH (IRB number: 2020-04-045).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDataset feature selection\u003c/h3\u003e\n\u003cp\u003eA total of 692 data features were collected, encompassing a wide range of clinical domains: demographic characteristics; lifestyle factors and family history; medical history (including prior surgeries and medication use); laboratory test results; infectious disease markers (e.g., hepatitis B and C virus, \u003cem\u003eClonorchis sinensis\u003c/em\u003e); imaging findings; and operative and pathological reports. Data acquisition was standardized across institutions using a unified survey protocol to minimize inter-institutional variability and reduce systematic bias. Under the guidance of clinical experts, we performed feature preprocessing for relevance and interpretability, yielding a selected panel of 57 variables\u0026mdash;comprising laboratory biomarkers and clinical characteristics. Additional baseline characteristics not included in the final analytic panel\u0026mdash;particularly demographic characteristics, lifestyle factors, and family history\u0026mdash;are presented in Supplementary Table\u0026nbsp;3. The detailed criteria of feature selection are provided in the Supplementary Methods.\u003c/p\u003e \u003cp\u003eTo ensure model training quality, we applied a 40% missingness threshold for feature exclusion and exceptions were made for clinically critical features (e.g., CA19-9 and CEA), which were retained despite exceeding this threshold based on expert consultation. Remaining missing values within the included features were imputed using the median value across non-missing entries.\u003c/p\u003e \u003cp\u003eFurthermore, to improve predictive accuracy and interpretability, model development was preceded by a structured feature-reduction process. From an initial pool of 57 candidate clinical variables, univariate chi‐square tests identified 26 variables with the strongest associations to the diagnostic outcome (Supplementary Table\u0026nbsp;2); notably, CA19-9 was deliberately excluded from this final set. The BTC risk prediction model was therefore trained without CA19-9, and its impact was subsequently assessed via sensitivity analysis by reintroducing CA19-9 to evaluate changes in diagnostic performance.\u003c/p\u003e\n\u003ch3\u003eMachine learning model for discriminating BTC from BTD\u003c/h3\u003e\n\u003cp\u003eTo distinguish BTC from BTD, we developed a machine learning model based on Tabular Prior-Data Fitted Networks (TabPFN) [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. For internal validation, a Monte Carlo cross-validation framework was employed, in which the training cohort was repeatedly and randomly divided into internal training and validation subsets across 50 iterations to assess model performance stability. For external validation, the model was trained on the combined training cohort of the three centers and evaluated using the held-out center cohort from the independent center.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eRisk factor analysis of BTC using interpretable machine learning\u003c/h2\u003e \u003cp\u003eSHapley Additive exPlanations (SHAP) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] were applied to identify clinical variables that influenced the model's prediction toward BTC or BTD. To further enhance the clinical interpretability of our prediction model, a surrogate decision tree was trained to approximate the behavior of the TabPFN model. This decision tree\u0026ndash;based surrogate model distilled its logic into a small number of binary clinical rules, which could be further developed into a clinical trial protocol. Despite its reduced complexity, the surrogate maintained high concordance with the original model\u0026rsquo;s predictions and replicated key decision boundaries using clinically recognizable thresholds. Several rule-based explanations were derived, reflecting the behavior of the full model in identifying high-risk BTC cases\u0026mdash;including those with non-elevated CA19-9 levels\u0026mdash;highlighting its potential utility in diagnostically challenging scenarios.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eCohort Characteristics\u003c/h2\u003e \u003cp\u003eThe detailed characteristics of the training and external validation cohort are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Patients with BTC were significantly older than those with BTD in both the training (median age: 73 vs. 66 years; P \u0026lt; .001) and external validation cohorts (73 vs. 62.5 years; P \u0026lt; .001). Across both cohorts, BTC was consistently associated with elevated serum levels of CA19-9 and CEA, as well as markers indicative of hepatobiliary dysfunction, including total and direct bilirubin, alkaline phosphatase (ALP), and gamma-glutamyl transferase (GGT) (all P \u0026lt; .001). BTC patients also exhibited lower hemoglobin, hematocrit, and serum albumin concentrations relative to BTD patients, consistent with systemic inflammation and hepatic dysfunction. Increases in aminotransferases (AST and ALT) and prolonged activated partial thromboplastin time (aPTT) were likewise observed in the BTC groups. While some variables—such as weight, phosphorus, and uric acid—demonstrated inter-cohort variability in their statistical significance, the overall direction and magnitude of key BTC-associated biomarkers remained consistent between the training and external validation cohorts. The distribution of BTC anatomical subtypes is summarized in Supplementary Table\u0026nbsp;4. Among patients diagnosed with BTC, 118 (21.5%) had intrahepatic cholangiocarcinoma (iCCA), 145 (26.5%) distal cholangiocarcinoma (dCCA), 160 (29.2%) perihilar cholangiocarcinoma (pCCA), and 161 (29.4%) gallbladder cancer (GBC). Subtype information was unavailable for 57 patients due to incomplete responses in the PHR dataset. Additionally, 12 patients were identified as having multiple concurrent BTC subtypes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCohort characteristics of patients in the PHR dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eTraining Cohort\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c8\" namest=\"c5\"\u003e \u003cp\u003eExternal Validation Cohort\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClinical variables\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBTD (n = 891)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBTC (n = 548)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBTD (n = 164)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e \u003cp\u003eBTC (n = 81)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, y\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e66.0 (56.0–75.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e73.0 (66.0–79.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e62.50 (48.0–74.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e73.0 (67.0–77.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeight, cm\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e163.0 (156.15–169.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e161.0 (154.0–167.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e162.50 (156.90–169.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e162.0 (154.0–167.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.221\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWeight, kg\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e63.0 (56.0–71.35)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e59.15 (53.0–66.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e63.0 (54.0–73.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e59.40 (53.0–64.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.009\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWaist, cm\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32.0 (30.0–34.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e32.0 (30.0–34.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.238\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e32.0 (30.0–32.12)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e32.0 (30.0–32.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.623\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e493 (55.3%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e323 (58.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.198\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82 (50.0%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e46 (56.8%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.387\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e216 (24.2%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e163 (29.7%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.025\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e33 (20.1%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e19 (23.5%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.664\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlcohol Usage, n (%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCurrent\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e318 (35.7%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e131 (23.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e62 (37.8%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e15 (18.5%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFormer\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e198 (22.2%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e167 (30.5%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e35 (21.3%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e35 (43.2%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNever\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e375 (42.1%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e250 (45.6%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e67 (40.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e31 (38.3%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLab Results\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANC, x10⁹/L\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.60 (3.29–6.97)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.60 (3.44–6.64)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.527\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.39 (3.08–7.28)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.59 (3.58–7.36)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.456\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHemoglobin, g/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13.10 (12.0–14.30)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.60 (11.30–13.70)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e13.10 (12.30–14.40)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e12.40 (11.10–13.60)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHematocrit, %\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38.90 (35.90–42.30)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e37.20 (33.60–40.60)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e38.10 (35.48–41.20)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e35.80 (31.40–40.10)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePlatelets, x10⁹/L\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e222.0 (176.0–267.50)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e248.50 (197.0–314.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e226.50 (186.75–271.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e240.0 (185.0–301.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.247\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eESR, mm/hour\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e33.0 (24.0–33.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e33.0 (33.0–68.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e33.0 (33.0–33.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e33.0 (33.0–33.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.076\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCRP, mg/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.55 (0.10–2.13)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.82 (0.26–3.34)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.63 (0.60–2.95)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.81 (0.45–2.58)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.382\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal Protein, g/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7.0 (6.58–7.33)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.90 (6.50–7.29)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.016\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7.20 (6.88–7.60)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e7.10 (6.50–7.40)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.012\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlbumin, g/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.25 (3.97–4.50)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.05 (3.68–4.30)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.20 (3.90–4.43)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.80 (3.30–4.30)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal Bilirubin, g/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.80 (0.46–1.79)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.40 (0.59–9.55)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.0 (0.70–2.40)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.80 (0.60–11.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.007\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDirect Bilirubin, mg/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.74 (0.30–1.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.74 (0.32–7.04)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.74 (0.35–0.74)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.74 (0.74–6.56)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAST, U/L\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e30.0 (21.0–118.50)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e67.0 (28.0–145.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e38.50 (22.0–109.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e61.0 (29.0–173.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eALT, U/L\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28.0 (17.0–128.50)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e58.50 (20.0–166.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e41.0 (17.75–116.25)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e42.0 (20.0–198.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.189\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eALP, U/L\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e90.0 (66.0–161.50)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e259.0 (102.75–501.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e110.50 (75.0–221.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e348.0 (125.0–521.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGGT, U/L\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e173.50 (31.50–281.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e322.50 (102.25–766.25)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e173.50 (33.0–285.25)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e248.0 (106.0–914.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePhosphorus, mg/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.24 (2.80–3.51)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.24 (2.90–3.67)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.057\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.50 (3.0–3.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.20 (2.80–3.60)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.034\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUric Acid, mg/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.46 (3.98–5.40)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.40 (3.30–5.10)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.80 (3.80–6.03)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.30 (3.20–5.60)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.022\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBUN, mg/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.30 (11.30–18.60)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14.80 (11.67–18.62)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.146\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e14.55 (12.47–19.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e14.90 (11.80–21.80)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.418\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCreatinine, mg/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.81 (0.66–0.97)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.80 (0.67–0.99)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.762\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.80 (0.64–0.95)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.85 (0.69–0.98)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.161\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA19-9, U/mL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.75 (6.44–25.69)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e93.59 (20.72–712.10)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e14.50 (5.85–45.30)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e76.90 (7.05–504.17)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCEA, ng/mL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.30 (1.98–2.30)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.48 (2.0–4.51)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.30 (1.60–2.70)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.30 (2.30–4.12)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePT, s\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.0 (11.30–12.40)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.0 (11.38–12.70)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.068\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e12.80 (12.30–13.53)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e12.40 (11.90–13.10)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAPTT, s\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e27.20 (25.30–27.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e27.50 (25.60–30.42)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e34.60 (31.60–37.12)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e33.30 (29.30–35.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.018\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGlucose, mg/dL\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e121.0 (104.0–141.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e121.0 (107.0–144.25)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.180\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e121.0 (121.0–147.75)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e121.0 (104.0–145.0)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.086\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHbA1c, %\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.80 (5.80–5.80)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.80 (5.70–6.10)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.636\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5.80 (5.70–5.80)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.80 (5.80–5.80)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.377\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHBsAg, n (%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.409\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.065\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e22 (2.5%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16 (2.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3 (1.8%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4 (4.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e856 (96.1%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e528 (96.4%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e31 (79.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e70 (86.4%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHBsAb, n (%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt; 0.001\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.148\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e667 (74.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e363 (66.2%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e92 (56.1%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e53 (65.4%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e212 (23.8%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e180 (32.8%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e43 (26.2%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e21 (25.9%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHCVAb, n (%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.109\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e0.191\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 (1.1%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10 (1.8%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5 (3.0%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1 (1.2%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e864 (97.0%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e534 (97.4%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e141 (86.0%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e76 (93.8%)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"8\"\u003eANC, absolute neutrophil count; ESR, erythrocyte sedimentation rate; CRP, c-reactive protein; AST, alanine aminotransferase; ALT, alanine aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyl transferase; BUN, blood urea nitrogen; CA19-9, carbohydrate antigen 19 − 9; CEA, carcinoembryonic antigen; PT, prothrombin time; APTT, activated partial thromboplastin time; HbA1c, hemoglobin a1c; HBsAg, hepatitis b surface antigen; HBsAb, hepatitis b surface antibody; HCVAb, hepatitis c virus antibody.\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colspan=\"8\"\u003eData are presented as median with interquartile range (IQR), defined by the first (Q1) and third (Q3) quartiles. Categorical variables are expressed as counts and percentages, n (%).\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e \u003cp\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003ePerformance evaluation of our prediction model for distinguishing BTC from BTD\u003c/h2\u003e \u003cp\u003eUsing 26 preselected clinical variables (Supplementary Table\u0026nbsp;2), the TabPFN model demonstrated superior performance compared with other machine learning algorithms (Supplementary Table\u0026nbsp;5). Receiver operating characteristic (ROC) analyses confirmed its higher discriminative ability (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea–b), while calibration plots showed close concordance between predicted probabilities and observed outcomes, underscoring the reliability of its risk estimates (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec–d).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe further evaluated the contribution of CA19-9, a widely used biomarker for BTC, to the model’s ability to discriminate between BTC and BTD. Two TabPFN models were compared: one including CA19-9, trained on patients with available measurements (n = 885), and another excluding CA19-9, trained on the full cohort (n = 1,439). Unexpectedly, the model without CA19-9 achieved comparable predictive accuracy (see Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Given the clinical need for applicability in Lewis antigen–negative patients and in those lacking CA19-9 measurements, we excluded CA19‑9 from the model development.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of diagnostic performances between TabPFN with and without the biomarker CA19-9, and the performance of CA19-9 with its two cutoffs.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInternal Validation\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN with CA19-9\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.895 (0.85–0.93)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.824 (0.75–0.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.772 (0.68–0.87)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.822 (0.77–0.89)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.777 (0.70–0.86)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN with CA19-9\u003c/p\u003e \u003cp\u003e(OC: 0.628)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.895 (0.85–0.93)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.731 (0.66–0.81)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.867 (0.76–0.95)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.876 (0.81–0.94)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.719 (0.66–0.79)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN without CA19-9\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.893 (0.86–0.92)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.682 (0.59–0.75)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.910 (0.87–0.94)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.825 (0.77–0.89)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.823 (0.78–0.86)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN without CA19-9\u003c/p\u003e \u003cp\u003e(OC: 0.359)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.893 (0.86–0.92)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.784 (0.73–0.83)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.839 (0.79–0.89)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.752 (0.69–0.82)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.863 (0.83–0.89)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA19-9 \u0026gt; 37\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.790 (0.76–0.82)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.640 (0.60–0.68)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.817 (0.78–0.85)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.817 (0.78–0.85)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.639 (0.60–0.68)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA19-9 \u0026gt; 27.7\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.790 (0.76–0.82)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.690 (0.65–0.73)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.778 (0.73–0.82)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.800 (0.76–0.84)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.662 (0.62–0.70)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExternal Validation\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN with CA19-9\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.801 (0.73–0.86)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.727 (0.63–0.83)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.643 (0.66–0.80)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.505 (0.48–0.67)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.825 (0.79–0.90)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN with CA19-9\u003c/p\u003e \u003cp\u003e(OC: 0.628)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.801 (0.73–0.86)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.635 (0.52–0.75)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.809 (0.74–0.87)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.625 (0.51–0.74)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.816 (0.75–0.88)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN without CA19-9\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.855 (0.81–0.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.744 (0.65–0.83)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.767 (0.70–0.83)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.612 (0.51–0.70)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.858 (0.80–0.92)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN without CA19-9\u003c/p\u003e \u003cp\u003e(OC: 0.359)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.855 (0.81–0.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.829 (0.74–0.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.616 (0.54–0.68)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.516 (0.43–0.60)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.879 (0.82–0.94)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA19-9 \u0026gt; 37\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.649 (0.56–0.73)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.576 (0.46–0.69)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.727 (0.65–0.80)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.514 (0.40–0.62)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.774 (0.70–0.85)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA19-9 \u0026gt; 27.7\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.649 (0.56–0.73)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.606 (0.48–0.72)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.674 (0.59–0.75)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.482 (0.38–0.58)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.774 (0.70–0.85)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"6\"\u003eAUROC, area under the receiver operating characteristic curve; CI, confidence interval; OC, Youden-index–optimized cutoff; PPV, positive predictive value; NPV, negative predictive value.\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colspan=\"6\"\u003eTabPFN with CA 19 − 9: TabPFN model trained including CA 19 − 9 as a feature, TabPFN without CA 19 − 9: TabPFN model trained excluding CA 19 − 9. “(OC: x)” indicates performance at each model’s Youden-index–derived cutoff: x, rather than the pre‐specified cutoff 0.5.\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e \u003cp\u003e\u003c/p\u003e \u003cp\u003eOur final model achieved an AUROC of 0.893 (95% CI, 0.863–0.920) and a sensitivity of 0.784 (95% CI, 0.727–0.827) for detecting BTC (Youden index 0.359). When applied to the fourth center's cohort — which had been excluded from model training — the diagnostic model maintained strong performance (AUROC, 0.855; 95% CI, 0.805–0.899), identifying over 82% of BTC cases (sensitivity, 0.829; 95% CI, 0.743–0.904) (Youden index 0.359) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Notably, the model demonstrated substantially higher sensitivity for detecting BTC than the conventional tumor biomarker CA19-9 (cutoff of \u0026gt; 37 U/L), both in the internal cohort (78.4% vs. 64.0%) and in external validation (82.9% vs. 57.6%). Even when applying a lower CA19-9 threshold (27.7 U/L, Youden index), the model retained superior sensitivity in both internal (78.4% vs. 69.0%) and external cohorts (82.9% vs. 60.6%). The difference was substantial across other diagnostic metrics, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eSubgroup analysis by CA19-9 Level\u003c/h2\u003e \u003cp\u003eWe further evaluated the diagnostic performance of the TabPFN model in subgroups stratified by CA19-9 levels: normal (\u0026lt; 37 U/mL), indeterminate (37–100 U/mL), and elevated (\u0026gt; 100 U/mL). We selected 37 U/mL as the lower boundary because it represents the established upper limit of normal for CA19-9 [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The 100 U/mL threshold was chosen as the upper boundary based on evidence showing that values below this level are frequently observed in benign biliary obstruction, whereas values exceeding 100 U/mL are reported to significantly increase the likelihood of malignancy [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. In internal validation, AUROCs were similar across subgroups—0.819 (95% CI, 0.806–0.831), 0.821 (95% CI, 0.792–0.847), and 0.831 (95% CI, 0.813–0.848), respectively—indicating that the model’s discrimination was preserved regardless of CA19-9 elevation. External validation showed similarly robust results (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ee–f, Supplementary Table\u0026nbsp;5).\u003c/p\u003e \u003cp\u003eApplying subgroup-specific cutoffs to balance sensitivity and specificity yielded clinically relevant trade-offs. In the normal CA19-9 group, lowering the probability threshold to 0.304 increased sensitivity to 85.7% (95% CI, 71.4–96.6%), enabling the model to correctly identify the majority of BTC cases that would otherwise be missed by CA19-9 alone, at the cost of reduced specificity (52.1%). In the indeterminate group (37–100 U/mL), a cutoff of 0.372 achieved perfect sensitivity (100%), ensuring that no BTC cases were missed, while maintaining moderate specificity (66.7%) (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of the TabPFN model diagnostic performances on different CA19-9 level subgroups (~ 37, 37 ~ 100, 100~) for both internal and external validation cohorts.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSubgroups of\u003c/p\u003e \u003cp\u003eCA19-9 levels\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003cp\u003e(95% CI)\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInternal Validation\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e~ 37\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.819 (0.81–0.83)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.540 (0.52–0.56)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.882 (0.87–0.89)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.726 (0.70–0.75)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.768 (0.75–0.78)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e~ 37 (OC: 0.304)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.819 (0.81–0.83)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.750 (0.73–0.77)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.720 (0.71–0.74)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.608 (0.59–0.63)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.833 (0.82–0.85)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e37 ~ 100\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.821 (0.79–0.85)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.721 (0.69–0.75)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.761 (0.71–0.81)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.873 (0.85–0.90)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.546 (0.50–0.59)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e37 ~ 100 (OC: 0.372)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.821 (0.79–0.85)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.816 (0.79–0.84)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.654 (0.60–0.71)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.843 (0.82–0.87)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.610 (0.56–0.66)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e100~\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.831 (0.81–0.85)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.838 (0.82–0.85)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.503 (0.45–0.56)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.922 (0.91–0.93)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.308 (0.27–0.35)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e100~ (OC: 0.709)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.831 (0.81–0.85)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.719 (0.70–0.74)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.794 (0.75–0.84)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.960 (0.95–0.97)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.289 (0.26–0.32)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExternal Validation\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e~ 37\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.731 (0.64–0.82)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.500 (0.32–0.70)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.750 (0.66–0.84)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.368 (0.22–0.53)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.837 (0.76–0.91)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e~ 37 (OC: 0.304)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.731 (0.64–0.82)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.857 (0.71–0.97)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.521 (0.42–0.62)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.343 (0.23–0.46)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.926 (0.85–0.98)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e37 ~ 100\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.907 (0.75–1.00)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.833 (0.50–1.00)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.778 (0.58–0.95)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.556 (0.20–0.89)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.933 (0.79–1.00)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e37 ~ 100 (OC: 0.372)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.907 (0.75–1.00)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.000 (1.00–1.00)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.667 (0.44–0.88)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.500 (0.20–0.79)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.000 (1.00–1.00)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e100~\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.821 (0.69–0.93)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.875 (0.75–0.97)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.500 (0.27–0.74)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.757 (0.61–0.89)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.692 (0.42–0.92)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e100~ (OC: 0.709)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.821 (0.69–0.93)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.750 (0.59–0.89)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.778 (0.57–0.95)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.857 (0.72–0.97)\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.636 (0.43–0.83)\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"6\"\u003eAUROC, area under the receiver operating characteristic curve; CI, confidence interval; OC, Youden-index–optimized cutoff; PPV, positive predictive value; NPV, negative predictive value.\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colspan=\"6\"\u003e“(OC: x)” indicates performance at each model’s Youden-index–derived cutoff: x, rather than the pre‐specified cutoff 0.5.\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e \u003cp\u003e\u003c/p\u003e \u003cp\u003eThese findings highlight the model’s potential utility as a complementary diagnostic tool, particularly for patients with normal or borderline CA19-9 levels, where reliance on CA19-9 alone may fail to detect early disease.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eSerum Biomarker Contribution to BTC Risk\u003c/h2\u003e \u003cp\u003eWe next applied SHapley Additive exPlanations (SHAP) to quantify the impact of each serum biomarker on the TabPFN model’s prediction of BTC risk. In the training cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea) and external validation cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb), carcinoembryonic antigen (CEA) and alkaline phosphatase (ALP) emerged as the two strongest predictors. Notably, activated partial thromboplastin time (aPTT) also ranked among the top contributors in the external validation set, while the remaining biomarkers demonstrated consistent importance across both cohorts. To illustrate how individual biomarker values influence the BTC risk score, patient-level biomarker contributions are depicted in Supplementary Fig.\u0026nbsp;1.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eUsing SHAP analyses, we further evaluated biomarker contributions for each anatomical BTC subtype in the training cohort. Gallbladder cancer and intrahepatic cholangiocarcinoma showed similar biomarker risk profiles, with CEA and ALP emerging as the strongest positive contributors to predicted malignancy. In contrast, among patients with extrahepatic cholangiocarcinoma (perihilar and distal subtypes), ALP was identified as the predominant driver of risk, followed by total bilirubin (Supplementary Fig.\u0026nbsp;2).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eRule-Based Risk Stratification for BTC\u003c/h2\u003e \u003cp\u003eTo better understand our TabPFN model’s predictions and enhance its clinical decision-making utility, we derived a simplified surrogate model—a decision tree—that mimics the model’s approach to BTC detection. Applying this decision tree to both the internal and external validation cohorts stratified patients into high- and low-risk groups, highlighting key biomarker thresholds associated with BTC risk. The decision tree is shown in Supplementary Fig.\u0026nbsp;3.\u003c/p\u003e \u003cp\u003eFor patients with markedly elevated alkaline phosphatase (ALP \u0026gt; 311.5 U/L), the model predicted a high probability of BTC (~ 77.8%). Among the high-ALP patients the presence of extreme hyperbilirubinemia (total bilirubin \u0026gt; 6.02 mg/dL) further increased the predicted BTC risk to 94.4%. Notably, when ALP levels were in the range of 311.5–490.5 U/L and total bilirubin was only moderately elevated (1.15–6.02 mg/dL), the model assigned a relatively lower BTC risk (34.9%). In contrast, patients with similarly high ALP but near-normal bilirubin (≤ 1.15 mg/dL) exhibited a substantially higher predicted risk (~ 73.0%). This seemingly paradoxical pattern may reflect different disease presentations: intrahepatic cholangiocarcinoma can elevate ALP while bilirubin remains mildly increased or normal, whereas moderate bilirubin elevations in the context of high ALP might sometimes indicate benign partial obstructions [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor patients with lower ALP levels (≤ 311.5 U/L), the decision tree identified serum carcinoembryonic antigen (CEA) as the next most informative splitter. Those with CEA above 3.31 ng/mL formed a higher-risk subgroup, with a predicted BTC risk of ~ 56.7%, and if CEA exceeded 9.69 ng/mL the risk rose to 86.4%. In the subset of patients with moderately elevated CEA (3.31–9.69 ng/mL), a concurrent total bilirubin \u0026gt; 5.28 mg/dL further boosted the BTC risk to 84.8%.\u003c/p\u003e \u003cp\u003eFor patients with relatively benign biomarker profiles (ALP ≤ 311.5, CEA ≤ 3.31), age emerged as an important stratifier in the decision tree. Specifically, age \u0026gt; 56.5 years delineated a higher-risk branch. In this older subgroup with otherwise unremarkable markers, an isolated marked hyperbilirubinemia (\u0026gt; 8.90 mg/dL) still conferred a high predicted risk of BTC (73.5%). This emphasizes that in an older patient, the presence of significant jaundice alone is a red flag, even if other markers like ALP or CEA are normal. Conversely, if total bilirubin remained below 8.90 mg/dL in these older patients, the model identified an intermediate-risk scenario (≈ 57% BTC probability) only when subtle coagulation changes were present – specifically, an elevated activated partial thromboplastin time (aPTT \u0026gt; 30.95 s) alongside a normal direct bilirubin (≤ 0.27 mg/dL).\u003c/p\u003e \u003c/div\u003e "},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed a machine learning–based model to distinguish biliary tract cancer (BTC) from benign biliary tract disease (BTD) using routinely available demographic, clinical and laboratory variables, with an emphasis on seamless integration into everyday workflows. The model showed consistently high discrimination in both internal and independent external cohorts, including clinically challenging subgroups with low CA19-9 levels. Notably, adding CA19-9 conferred only modest incremental benefit, and performance remained robust when CA19-9 was excluded—supporting use in Lewis antigen–negative patients and in settings where CA19-9 is unavailable. In head-to-head comparisons, the model markedly outperformed CA19-9 alone, underscoring the limitations of CA19-9–dependent screening pathways.\u003c/p\u003e\u003cp\u003ePrior work has demonstrated promising BTC/CCA prediction across heterogeneous populations using conventional statistical models [\u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e–\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Hu et al [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] employed a multivariate linear Cox proportional hazards (CoxPH) model to screen for cholangiocarcinoma (CCA) in a cohort of 1,459 primary sclerosing cholangitis (PSC) patients from the Mayo Clinic, yielding a concordance index(C-index) of 0.69. T. Meister et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] and X. Shen et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] both conducted cholangiocarcinoma (CCA) screening using multivariate logistic regression analysis, each applied to the two cohorts: 1,135 patients with bile duct obstruction and 2,269 patients with intrahepatic lymphocytes. Each studies yielded validation performances with area under the receiver operating characteristic curve (AUROC) values of 0.862 and 0.867, respectively. While direct cross-study comparisons are limited by differences in inclusion criteria and biomarker panels, our study advances this literature by encompassing a broader spectrum of malignant and benign biliary conditions and by demonstrating generalizability in an independent external cohort. Using only routine clinical and laboratory variables, the model maintains high discrimination across CA19-9 strata—including patients with normal CA19-9—addressing a key limitation of CA19-9–centered strategies. Moreover, clinician-facing explanations (SHAP) and a compact surrogate decision tree provide transparent thresholds that facilitate implementation without sacrificing performance.\u003c/p\u003e\u003cp\u003eBeyond performance, our model’s risk factors and rule-based pathogenesis of BTC support the model’s clinical validity. Among the top contributing biomarkers of our model, ALP and GGT are known for indicators of cholestasis and biliary obstruction, conditions often preceding BTC development [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Our findings of highly elevated ALP associated with BTC is consistent with the cholestatic laboratory profile often seen in biliary malignancies, where ALP levels typically rise multiple-fold above normal [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Bilirubin elevations are known to reflect obstructive processes and liver injury. In our study, we found high bilirubin levels with a high association to BTC, while near-normal bilirubin levels could also pose a risk in some clusters of patients. Overall, the combination of ALP and bilirubin in the decision tree underscores their importance; indeed, extrahepatic BTC classically presents with conjugated bilirubin and ALP elevated on the order of 2–10 times the upper normal limit in other studies [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Inflammatory markers such as CRP and ESR align with the known association between chronic inflammation and biliary carcinogenesis [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. These features are particularly relevant in diseases such as PSC, recurrent cholangitis, or choledochal cysts, where chronic biliary inflammation increases cancer risk [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Coagulation and liver function parameters—such as prolonged APTT, hypoalbuminemia, and thrombocytopenia—suggest impaired hepatic reserve and portal hypertension, common in advanced liver disease, a known risk factor for BTC [28,29. Viral hepatitis markers (HBsAg, HCV Ab) were also highly ranked, consistent with their known role in hepatobiliary malignancies, particularly intrahepatic cholangiocarcinoma [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] Additionally, the model identified metabolic indicators—fatty liver and elevated uric acid—as BTC-associated features. These may reflect underlying NAFLD or metabolic syndrome, which have been linked to biliary malignancies through inflammatory and insulin resistance pathways [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eAlthough our findings align with prior work, limitations persist. First, this study was conducted using a demographically specific cohort (South Korean patients). Thereby, the generalizability of the model to populations with different ethnic, geographic, or healthcare backgrounds warrants further validation. The development of biliary tract cancer (BTC) in patients with biomarker values falling within normal reference ranges remains insufficiently understood, underscoring a diagnostic challenge that calls for further investigation in future studies. Furthermore, we imputed missing values using per-feature medians; given the assumptions and limitations of single imputation, the findings should be interpreted with caution and additional external validation.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eUsing routinely available demographic and clinical data from patients diagnosed with BTC and BTD, we developed and validated a machine learning model capable of accurately distinguishing malignant from benign biliary tract disease. Leveraging only routinely available clinical and laboratory data, the model provides a non-invasive screening approach to identify ultra-high–risk patients for BTC within the already high-risk biliary disease population. Its reliance on standard data sources facilitates integration into electronic health record systems, where it could support earlier detection and targeted surveillance in routine clinical practice.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003e1. Ethics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Institutional Review Board (Ethics Committee) of Pusan National University Hospital (IRB No. 2009-006-094 and 2011-014-097) and the Institutional Review Board (Ethics Committee) of Kyungpook National University Hospital (IRB No. 2020-04-045). Informed consent to participate was obtained from all of the participants. All procedures were conducted in accordance with the Declaration of Helsinki and its later amendments.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2. Consent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3. Availability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are available from Pusan National University Hospital, but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Pusan National University Hospital.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4. Competing interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5. Funding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by a grant (SMF-AI-EJ003-2023) funded by Seegene Medical Foundation, Institute of Information \u0026amp; communications Technology Planning \u0026amp; Evaluation (IITP) grant funded by the Korea government (MSIT) (IITP-2025-RS-2025-02219116), Institute of Information \u0026amp; Communications Technology Planning \u0026amp; Evaluation (IITP) under the Artificial Intelligence Convergence Innovation Human Resources Development (IITP-202300254177), \u0026nbsp;the National R\u0026amp;D Program for Cancer Control, Ministry of Health and Welfare, Republic of Korea (HA20C0009), and Biomedical Research Institute Grant (202200140001) Pusan National University Hospital.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6. Authors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStudy concept and design: Hanul Park, Dong Uk Kim, Giltae Song\u003c/p\u003e\n\u003cp\u003eData acquisition: all authors\u003c/p\u003e\n\u003cp\u003eData analysis and interpretation: Hanul Park, Dong Uk Kim, Giltae Song\u003c/p\u003e\n\u003cp\u003eStudy supervision: Dong Uk Kim, Giltae Song\u003c/p\u003e\n\u003cp\u003eAdministrative, technical, or material support: all authors\u003c/p\u003e\n\u003cp\u003eApproval of final manuscript: all authors\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7. Acknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e"},{"header":"References ","content":"\u003col\u003e\n\u003cli\u003eVogel A, Bridgewater J, Edeline J, et al. Biliary tract cancer: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol. 2023;34(2):127\u0026ndash;140.\u003c/li\u003e\n\u003cli\u003eBray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68(6):394\u0026ndash;424.\u003c/li\u003e\n\u003cli\u003eFlorio AA, Ferlay J, Znaor A, et al. Global trends in intrahepatic and extrahepatic cholangiocarcinoma incidence from 1993 to 2012. Cancer. 2020;126(11):2666\u0026ndash;2678.\u003c/li\u003e\n\u003cli\u003eValle JW, Kelley RK, Nervi B, et al. Biliary tract cancer. Lancet. 2021;397(10272):428\u0026ndash;444.\u003c/li\u003e\n\u003cli\u003eClements O, Eliahoo J, Kim JU, et al. Risk factors for intrahepatic and extrahepatic cholangiocarcinoma: A systematic review and meta-analysis. J Hepatol. 2020;72(1):95\u0026ndash;103.\u003c/li\u003e\n\u003cli\u003eRazumilava N, Gores GJ. Classification, diagnosis, and management of cholangiocarcinoma. Clin Gastroenterol Hepatol. 2013;11(1):13\u0026ndash;21.\u003c/li\u003e\n\u003cli\u003eIlyas SI, Gores GJ. Pathogenesis, diagnosis, and management of cholangiocarcinoma. Gastroenterology. 2013;145(6):1215\u0026ndash;1229.\u003c/li\u003e\n\u003cli\u003eBlechacz B, Komuta M, Roskams T, et al. Clinical diagnosis and staging of cholangiocarcinoma. Nat Rev Gastroenterol Hepatol. 2011;8(9):512\u0026ndash;522.\u003c/li\u003e\n\u003cli\u003eMiros M, Kerlin P, Walker N, et al. Predicting cholangiocarcinoma in patients with primary sclerosing cholangitis before transplantation. Gut. 1991;32(11):1369\u0026ndash;1373.\u003c/li\u003e\n\u003cli\u003eKhaderi SA, Sussman NL. Screening for malignancy in primary sclerosing cholangitis (PSC). Curr Gastroenterol Rep. 2015;17:1\u0026ndash;5.\u003c/li\u003e\n\u003cli\u003eHu C, Iyer RK, Juran BD, et al. Predicting cholangiocarcinoma in primary sclerosing cholangitis: using artificial intelligence, clinical and laboratory data. BMC Gastroenterol. 2023;23(1):129. \u003c/li\u003e\n\u003cli\u003eJ Huang, X Bai, Y Qiu, et al. Application of AI on cholangiocarcinoma. Frontiers in Oncology. 2024;14.\u003c/li\u003e\n\u003cli\u003eZhou SN, Jv DW, Meng XF, et al. Feasibility of machine learning-based modeling and prediction using multiple centers data to assess intrahepatic cholangiocarcinoma outcomes. Ann Med. 2023;55(1):215\u0026ndash;223.\u003c/li\u003e\n\u003cli\u003eLv Y, Liu H, He P, et al. A novel model for predicting the prognosis of postoperative intrahepatic cholangiocarcinoma patients. Sci Rep. 2023;13(1):19267. \u003c/li\u003e\n\u003cli\u003eCX Chamberlain, E Faust, D Goldschmidt, et al. Burden of illness for patients with cholangiocarcinoma in the United States: a retrospective claims analysis. Journal of gastrointestinal oncology. 2021;12(2):658\u003c/li\u003e\n\u003cli\u003eHollmann, N., M\u0026uuml;ller, S., Purucker, L. et al. Accurate predictions on small data with a tabular foundation model. Nature. 2025;637(8045):319\u0026ndash;326 \u003c/li\u003e\n\u003cli\u003eLundberg SM, Lee S-I. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30.\u003c/li\u003e\n\u003cli\u003eZhang Y, Yang J, Li H, et al. Tumor markers CA19-9, CA242 and CEA in the diagnosis of pancreatic cancer: a meta-analysis. International journal of clinical and experimental medicine. 2015;8(7):11683\u003c/li\u003e\n\u003cli\u003eLiang B, Zhong L, He Q, et al. Diagnostic accuracy of serum CA19-9 in patients with cholangiocarcinoma: a systematic review and meta-analysis. Medical science monitor: international medical journal of experimental and clinical research. 2015;21:3555\u003c/li\u003e\n\u003cli\u003eForner A, Vidili G, Rengo M, et al. Clinical presentation, diagnosis and staging of cholangiocarcinoma. Liver International. 2019;39:98-107\u003c/li\u003e\n\u003cli\u003eHu C, Iyer RK, Juran BD, et al. Predicting cholangiocarcinoma in primary sclerosing cholangitis: using artificial intelligence, clinical and laboratory data. BMC Gastroenterol. 2023;23(1):129.\u003c/li\u003e\n\u003cli\u003eMeister T, Uphoff MA, Heinecke A, et al. Novel score for prediction of malignant bile duct obstruction based on biochemical and clinical markers. Aliment Pharmacol Ther. 2015;41(9):877\u0026ndash;887.\u003c/li\u003e\n\u003cli\u003eShen X, Zhao H, Jin X, et al. Development and validation of a machine learning-based nomogram for prediction of intrahepatic cholangiocarcinoma in patients with intrahepatic lithiasis. Hepatobiliary Surg Nutr. 2021;10(6):749-765.\u003c/li\u003e\n\u003cli\u003eNam YH, Park MS, Lee SM. Preoperative C-reactive protein as a prognostic factor for recurrence after surgical resection of biliary tract cancer. Korean J Clin Oncol. 2015;11(2):101\u0026ndash;105.\u003c/li\u003e\n\u003cli\u003eGerhardt T, Milz S, Schepke M, et al. C-reactive protein is a prognostic indicator in patients with perihilar cholangiocarcinoma. World J Gastroenterol. 2006;12(34):5495\u0026ndash;5500.\u003c/li\u003e\n\u003cli\u003eIlyas SI, Eaton JE, Gores GJ, et al. Primary sclerosing cholangitis as a premalignant biliary tract disease: surveillance and management. Clinical Gastroenterology and Hepatology. 2015;13(12):2152-2165\u003c/li\u003e\n\u003cli\u003eLabib PL, Goodchild G, Pereira SP, et al. Molecular pathogenesis of cholangiocarcinoma. BMC cancer. 2019;19:1-16\u003c/li\u003e\n\u003cli\u003eKe X, Jin B, You, W, et al. Comprehensive analysis of coagulation indices for predicting survival in patients with biliary tract cancer. BMC cancer. 2021;21:1\u0026ndash;15.\u003c/li\u003e\n\u003cli\u003ePoordad F. Review article: thrombocytopenia in chronic liver disease. Alimentary Pharmacology \u0026amp; Therapeutics. 2007;26:5-11\u003c/li\u003e\n\u003cli\u003eZhou Y, Zhao Y, Li B, et al. Hepatitis viruses infection and risk of intrahepatic cholangiocarcinoma: evidence from a meta-analysis. BMC cancer. 2012;12:1\u0026ndash;7.\u003c/li\u003e\n\u003cli\u003eWongjarupong N, Assavapongpaiboon B, Susantitaphong P, et al. Non-alcoholic fatty liver disease as a risk factor for cholangiocarcinoma: a systematic review and meta-analysis. BMC gastroenterology. 2017;17:1\u0026ndash;8.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Biliary tract cancer, cholangiocarcinoma, machine learning, diagnosis prediction","lastPublishedDoi":"10.21203/rs.3.rs-8736730/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8736730/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground/Aims \u003c/strong\u003eBiliary tract cancer (BTC) is an aggressive malignancy often diagnosed late due to its frequently asymptomatic presentation. Given substantial clinical overlap between BTC and benign biliary tract disease (BTD), early and accurate differentiation remains challenging. This study represents the first effort to develop and validate a non-invasive clinical model capable of identifying patients at especially high risk of BTC within the BTD population, thereby facilitating earlier diagnosis and intervention.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods \u003c/strong\u003eIn this multicenter observational study, we prospectively collected patient-reported survey data and retrospectively extracted laboratory, diagnostic data from electronic medical records at four tertiary centers in South Korea. A machine learning model was trained to differentiate biliary tract cancer (BTC) from benign biliary tract disease (BTD) using 1,439 patients from three centers and externally validated the model in 245 patients from an independent center. Explainable machine learning quantified biomarker contributions to model predictions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults \u003c/strong\u003eOur model demonstrated robust diagnostic performance, achieving an AUROC of 0.893 (95% CI, 0.863–0.920) and sensitivity of 0.784 (0.727–0.827). External validation from an independent center dataset produced consistent results. Notably, the model substantially outperformed CA19-9 and was developed without reliance on this biomarker, enabling applicability in Lewis antigen–negative patients. It also accurately identified BTC cases even among patients with normal CA19-9 levels, supporting its utility in broader clinical populations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions \u003c/strong\u003eThese results suggest our model’s potential to serve as a non-invasive screening tool, identifying patients at high risk of BTC among those presenting with suspicious biliary pathology, particularly when CA19-9 is uninformative.\u003c/p\u003e","manuscriptTitle":"Development and Validation of a Diagnostic Model for Biliary Tract Cancer Detection in Patients with Benign Biliary Disease: A Multicenter, Retrospective Case-Control Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-11 05:47:04","doi":"10.21203/rs.3.rs-8736730/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3d3ef038-9e25-4e2d-887f-1c6d09288ef3","owner":[],"postedDate":"February 11th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-30T11:47:40+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-11 05:47:04","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8736730","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8736730","identity":"rs-8736730","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}