Journal of Gastrointestinal Surgery (2022) 26:1713–1723 https://doi.org/10.1007/s11605-022-05398-7 ORIGINAL ARTICLE Machine Learning‑Based Analysis in the Management of Iatrogenic Bile Duct Injury During Cholecystectomy: a Nationwide Multicenter Study Victor Lopez‑Lopez1 · Javier Maupoey2 · Rafael López‑Andujar2 · Emilio Ramos3 · Kristel Mils3 · Pedro Antonio Martinez4 · Andres Valdivieso5 · Marina Garcés‑Albir6 · Luis Sabater6 · Luis Díez Valladares7 · Sergio Annese Pérez8 · Benito Flores8 · Roberto Brusadin1 · Asunción López Conesa1 · Valentin Cayuela1 · Sagrario Martinez Cortijo9 · Sandra Paterna10 · Alejando Serrablo10 · Santiago Sánchez‑Cabús11 · Antonio González Gil12 · Jose Antonio González Masía13 · Carmelo Loinaz14 · Jose Luis Lucena15 · Patricia Pastor16 · Cristina Garcia‑Zamora17 · Alicia Calero18 · Juan Valiente19 · Antonio Minguillon20 · Fernando Rotellar21,22 · Jose Manuel Ramia23 · Cándido Alcazar23 · Javier Aguilo24 · Jose Cutillas25 · Christoph Kuemmerli26,27 · Jose A. Ruiperez‑Valiente4 · Ricardo Robles‑Campos1 Received: 24 March 2022 / Accepted: 17 June 2022 / Published online: 5 July 2022 © The Author(s) 2022 Abstract Background Iatrogenic bile duct injury (IBDI) is a challenging surgical complication. IBDI management can be guided by artificial intelligence models. Our study identified the factors associated with successful initial repair of IBDI and predicted the success of definitive repair based on patient risk levels. Methods This is a retrospective multi-institution cohort of patients with IBDI after cholecystectomy conducted between 1990 and 2020. We implemented a decision tree analysis to determine the factors that contribute to successful initial repair and developed a risk-scoring model based on the Comprehensive Complication Index. Results We analyzed 748 patients across 22 hospitals. Our decision tree model was 82.8% accurate in predicting the success of the initial repair. Non-type E (p < 0.01), treatment in specialized centers (p < 0.01), and surgical repair (p < 0.001) were associated with better prognosis. The risk-scoring model was 82.3% (79.0–85.3%, 95% confidence interval [CI]) and 71.7% (63.8–78.7%, 95% CI) accurate in predicting success in the development and validation cohorts, respectively. Surgical repair, successful initial repair, and repair between 2 and 6 weeks were associated with better outcomes. Discussion Machine learning algorithms for IBDI are a novel tool may help to improve the decision-making process and guide management of these patients. Keywords Iatrogenic bile duct injury · Cholecystectomy · Machine learning · Artificial neural network Introduction Iatrogenic bile duct injury (IBDI) is a major complication of cholecystectomy. While the technique for laparoscopic chol- Jose A. Ruiperez-Valiente and Ricardo Robles-Campos are co- ecystectomy (LC) has improved over the years, the incidence senior authors. of IBDI remains unchanged due to miss interpretation of normal hepatobiliary anatomy. In particular, vascular hepatic The paper is based on a previous communication in A Coruña lesions that show involvement of the intrahepatic bile duct (Spain) to the 32th National Meeting of the Spanish Surgical are associated with higher morbidity and mortality such that Association. a liver transplant is required in some cases.1–3 The incidence * Victor Lopez‑Lopez of IBDI during cholecystectomy is influenced by the anat- victorrelopez@gmail.com omy of the patient, nature of the pathology, intraoperative Extended author information available on the last page of the article complications, such as bleeding, and technical errors.4,5 13 Vol.:(0123456789) 1714 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 IBDI is associated with multiple sequelae that require successful. The clinical aspects of the patient included the multiple therapeutic interventions. As such, it must be man- American Society of Anesthesiologist (ASA) scores (≥ 3), aged well to achieve successful results and minimize com- whether the patient had risk factors, whether it was a type E plications.6,7 At the onset, IBDI must be diagnosed immedi- IBDI, and whether the patient had any vascular injury. ately, and the patient must be referred to a specialized center Definitive repair was defined as success when the goal without delay.8,9 The long-term prognosis of IBDI is also of injury repair was reached by endoscopy, interventional dependent on the skill of the surgeon performing the repair, radiologist or surgery. Success of the repair included the and the approach selected for the repair. Among these fac- removal of drains, stent-free, the absence of symptoms of tors, early detection of IBDI and its sequelae are the keys to bile leak infection or the absence of signs of obstruction proper management.10 However, IBDI is poorly documented. (cholestasis, cholangitis or stenosis). We defined special- Most of the studies on IBDI provide incomplete data on the ized centers as those that performed liver transplantation or severity of the condition and fail to identify all of the seque- performed complex hepatobiliary surgery and only referred lae involved. The Comprehensive Complication Index (CCI) those patients who required liver transplantation. is a potential tool that can summarize all of the postopera- IBDI was classified according to Strasberg’s classifica- tive complications associated with IBDI.11 The CCI uses a tion. The patient risk factors analyzed included cholecystitis, metric approach to quantify complications. cholangitis, bleeding, scleroatrophic gallbladder carcinoma, Several algorithms have been developed to aid in the med- anatomical abnormalities, Mirizzi syndrome, choledocho- ical decision-making process.12–15 Once machine learning lithiasis, and acute or chronic pancreatitis. The primary reaches a certain threshold, the technology can be applied intervention included surgical, radiologic, or endoscopic to improve IBDI management. Our study aimed to design a procedures. The surgical repair techniques included primary decision tree model to identify the factors that influence the suturing, end-to-end anastomosis, cystic repair, use of the success of the initial repair, as well as develop a risk-scoring Kehr tube, lavage plus drainage, Roux-en-Y hepaticojeju- system based on the CCI data. The quality of both models nostomy, liver resection, or liver transplantation. The radio- was validated using an independent cohort of patients. logical approaches included percutaneous drain and stent placement. The endoscopic interventions included stent placement and papillotomy. Methods Decision Tree and CCI‑Based Risk Score Models Study Design Construction This study was a retrospective multi-institution cohort We proposed two different models in this study. The first study. Patients diagnosed with IBDI in 22 Spanish centers model examined the success of the initial repair (yes or no) between January 1990 and July 2020 were included in our by using a recursive conditional inference tree with the ctree study. Ethical approval for this study was obtained from the package in R. We then incorporated the CCI into this model. Clinical Research and Ethics Committee at the Virgen de la The CCI weighs the severity of all postoperative compli- Arrixaca Clinic and University Regional Hospital Murcia, cations according to the Clavien–Dindo classification and Spain (Internal Protocol Code: 2021–5-3HCUVA). All the expresses morbidity through a continuous numeric scale centers completed a uniform data collection protocol with from 0 (no complications) to 100 (death). clear variable definitions. The second model examined the effect of risk groups on the prognosis of IBDI repair. We grouped CCI scores into Variable Selection ranges and characterized each range as being low (0–25), low-medium (25–50), medium–high (50–75), and high The initial dataset contained 40 variables with data on 748 (75–100) risk. We grouped these scores together because patients. We collected data on the demographic variables, we expected little differentiation in the level of risk within clinical characteristics, and clinical aspects of each patient. each range. We then implemented a random forest regression The demographic variables included age and gender. The model using the caret package in R to obtain the predicted clinical characteristics included the type of approach (open CCI values for our study population. As such, we were able or laparoscopic), intraoperative diagnosis, days to diagnosis, to stratify patients according to their level of risk. whether the hospital was specialized, whether surgical repair The hyperparameters of both models were finetuned by was required, whether the repair was performed in the same implementing tenfold cross-validation, which highlighted hospital, and the time to repair (less than 2 weeks, between 2 the values that provided the best quality metrics while and 6 weeks, or more than 6 weeks). For the second model, avoiding overfitting. Statistical significance was set at we also included data on whether the initial repair was p < 0.05, and statistical analysis was performed using R 13 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 1715 programming language (Version 3.6.1; R Foundation for Results Statistical Computing, Vienna, Austria). We analyzed 748 patients across 22 hospitals in Spain. Our study population comprised 394 (52.7%) women and 354 Validation and Quality Metrics (47.3%) men. The mean age of the patients in this cohort was 57 ± 16.72 years. Figure 1 shows the distribution of the To validate the generalizability of our models, we variables. The mean CCI of patients was 46.1 (± 19) with excluded the data from several hospitals during the a 4.9% of mortality. A total of 15 patients (2%) required a development of the model. As such, both models were liver transplant. The rate of success according to the need developed with the data from 15 hospitals (596 patients) for a first, second, third or fourth repair attempt was 60.6, (development cohort) and validated with the data from 78.3, 68.2, and 72.7%, respectively. The development cohort seven hospitals (152 patients) (validation cohort). included data on 596 patients from 15 hospitals across 12 We determined the quality metrics for the develop- provincial administrative regions. In contrast, the validation ment and validation cohorts. Our first model was a binary cohort included data on 152 patients from seven hospitals model. We evaluated this model using classical binary across five provincial administrative regions. classifier metrics, which included the area under the receiver operator characteristic curve (AUC), accuracy, Decision‑Making Model on the Success of the Initial sensitivity, specificity, kappa, F1-score, and precision. Repair The second model was evaluated as follows: we initially performed a regression analysis to determine the root This model provides a visual interpretation of the patients mean square error (RMSE), mean absolute error (MAE), who demonstrated good results because their IBDI were and R-squared values. We then classified our data into adequately managed, irrespective of whether the initial the aforementioned CCI ranges and verified this using repair was successful (Fig. 2). Some examples of the accuracy, kappa, and the confusion matrix. interpretation of this decision tree are as follows: a patient with type A–D IBDI who was managed in a specialized center has a higher probability of a successful initial repair Fig. 1  Chronological evolution of the number of cases for each year collected by the participating centers and a summary of the main variables used for the development and validation of both models 13 1716 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 Fig. 2  Base-tree model on the success of the initial repair. The unbi- simplicity. The higher a variable appears in the tree, the more pre- ased regression tree includes all the variables listed in the base model dictively important the variable was lesion Type E. Decision making in the text. The maximum depth of the tree is set at four layers for according lesions non type E (a) and type E (b) (nodes 7–8) than if the initial repair was performed in a but the failure rate increased if these specialized centers non-specialized center (nodes 5–6). In contrast, a patient did not perform surgical repair (node 26). For cases that with type E IBDI has a better prognosis if the primary were treated in non-specialized centers (nodes 12, 14, and intervention was surgical (nodes 31–34 and 35 for surgical 17–19), the success rate dropped significantly if the patient and non-surgical repair, respectively). However, our model had risk factors or vascular problems. The most important also demonstrated that the success rate fell when the repair variables in this model (Table 1) were the type of injury was done early because the need for early or emergency (E type) (p < 0.001), whether the hospital was specialized repair was associated with more severe IBDI. Our model (p < 0.001), whether the repair was performed in the same also demonstrated that patients with type E IBDI fared bet- hospital where the IBDI occurred (p < 0.001), and whether ter with treatment in specialized centers (nodes 22, 24–26), the repair was surgical (p < 0.001). Table 1  Distribution of the variables included in the decision model related with the probability of initial repair success Variable Total Data No. (%) Development cohort Validation cohort No. (%) p value No. (%) Demographic characteristic Age, mean (SD) 57 (16.72) 57.6 (16.41) 57 (17.38) 0.53 Men 354 (47.3%) 278 (46.6%) 76 (50.0%) 0.39 Clinical characteristic Open approach 183 (24.5%) 137 (22.9%) 46 (30.2%) 0.05 Laparoscopic approach 565 (75.5%) 459 (77.0%) 106 (69.7%) 0.05 Intraoperative Diagnosis 265 (35.4%) 220 (36.9%) 45 (29.6%) 0.02 Days to diagnosis, mean (SD) 18.66 (18.21) 18.22 (18.1) 20.36 (18.34) 0.01 Specialized Hospital 290 (38.8%) 262 (43.9%) 28 (18.4%) < 0.001 No specialized Hospital 458 (61.2%) 334 (56.1%) 124 (81.6%) < 0.001 Repair Same Hospital 600 (80.2%) 464 (77.8%) 136 (89.4%) < 0.001 Surgical Repair 565 (75.5%) 468 (78.5%) 97 (63.8%) < 0.001 Time of Repair < 2 weeks 571 (76.3%) 435 (72.9%) 136 (89.4%) 0.08 Time of Repair 2—6 weeks 113 (15.1%) 101 (16.9%) 12 (7.8%) 0.08 Time of Repair > 6 weeks 64 (8.6%) 60 (10.2%) 4 (2.6%) 0.08 Success first repair 453 (60.6%) 326 (54.7%) 127 (83.5%) ** Patient characteristic ASA > 3 215 (28.7%) 172 (28.8%) 43 (28.2%) < 0.001 Risk Factors 463 (61.9%) 368 (61.7%) 95 (62.5%) 0.084 E type 473 (63.2%) 417 (69.9%) 56 (36.8%) < 0.001 Vascular injury 92 (12.3%) 86 (14.4%) 6 (3.9%) 0.21 13 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 1717 Among patients in the development cohort, our qual- repair was successful, whether the initial repair was surgical, ity metrics analysis demonstrated that this model was 81% whether the repair was performed between 2 and 6 weeks of (range, 77.8–91.1%; 95% confidence interval [CI]) accurate the diagnosis, and whether the repair was performed at the in predicting whether the initial repair would be successful. same hospital as the diagnosis. The remaining variables are Additionally, the model had a sensitivity and specificity of listed in Table 2. 0.92 and 0.76, respectively. We then tested this model on the validation cohort. Among patients in the validation cohort, Table 2  Importance of each of the variables used in the development this model was 82.8% (range, 0.76–0.89; 95% CI) accurate and validation of the CCI risk-score model in predicting the success of the initial repair. However, the Variable Variable specificity and sensitivity of the model decreased to 0.87 Importance and 0.47, respectively. The quality metrics of this model are Model 2 listed in Supplementary Table 1. Demographic characteristic Age, mean (SD) 31.43 Risk‑Scoring Model Based on the CCI Men – Clinical characteristic The second model was based on the CCI, which utilizes the Open approach – Clavien–Dindo classification to score risk from 0 to 100. Laparoscopic approach 11.25 We built a random regression forest model that predicted the Intraoperative Diagnosis 33.14 potential patient’s CCI. We applied this model to the devel- Days to diagnosis, mean (SD) 36.85 opment cohort, which resulted in MAE and RMSE values Specialized Hospital – of 5.67 and 7.65 points, respectively. We also applied this No specialized Hospital 22.05 model to the validation cohort, which resulted in MAE and Repair Same Hospital 42.69 RMSE values of 9.70 and 13.13 points, respectively. The Surgical Repair 80.64 predicted CCI scores were classified into ranges that corre- Time of Repair < 2 weeks – sponded to different risk categories. Applied to the develop- Time of Repair 2—6 weeks 48.13 ment and validation cohorts, the model was able to predict Time of Repair > 6 weeks 14.84 the CCI range in 82.3% (range, 79.0–85.3%; 95% CI) and Success first repair 100.0 71.7% (range 63.8–78.7%; 95% CI) of cases, respectively. Patient characteristic Figure 3 shows the confusion matrix for the CCI ranges. ASA > 3 24.10 The quality metrics analysis of this model is listed in Sup- Risk Factors 35.89 plementary Table 2. E type 31.02 Our random forest analysis demonstrated that the most Vascular injury 33.83 important variables in this model were whether the initial Fig. 3  Confusion matrix for CCI ranges. We observed the dynamics prediction, therefore the true prediction rate of each category. Moreo- of the model predictions and which cases are more frequently con- ver, we see that those predictions that are wrong, mostly fall in the fused. In each cell we observe the absolute number of cases that fall subsequent category, therefore, the errors might be considered as not there, and also the percentage that number represents in the row. The large and they might be partially influenced by the sampling of each intensity of the color of the cell represents that percentage. Notice bucket that the diagonal of the matrices represents the same reference and 13 1718 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 13 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 1719 ◂Fig. 4  Visualizations of the distribution CCI ranges for successful of not static, meaning that as time goes by, new data can be firs repair (a), surgical repair (b), and specialized or non-specialized included to recalculate the model, facilitating a continu- hospitals (c) ous improvement in the management of iatrogenic bile duct injury over time. Finally, we should take into account Discussion that the model is only taking into account a subset of the characteristics that can affect the success of the surgery, This is the first study that analyzes on the largest cohort but many others, like the emotional status of the patient of patients in the literature with IBDI the potential of or the specific ability of the surgeon are not taken into artificial intelligence models in their management. Our account. Therefore, the final decision needs to be taken study provided two sets of results because we analyzed two holistically based on all the evidence and knowledge gath- models. We initially implemented a decision tree analysis ered by the surgeon, instead of solely relying on the model centered on the success of the initial repair and then pro- probabilities. posed a risk-scoring model based on the CCI. Both models Recent consensus measures have been proposed to mini- demonstrated that patients diagnosed with IBDI who are mize the risk of IBDI in LC,7,16,17 due to nowadays, IBDI managed in specialized hospitals and with surgical repair remains a problem of great clinical and quality of life signifi- achieve the best outcomes and the least rate of complica- cance.18 The incidence of IBDI in this present national reg- tions. In addition, it also shows that not only those more istry has remained more or less stable since 2000. IBDI may serious injuries, but also minor injuries must be referred occur due to a number of reasons. In particular, we noted to specialized centers to improve results. risk factors in 61.9% of our study population; IBDI occurred Machine learning algorithms offer a decision-making most frequently in the setting of acute cholecystitis (68.4%), procedure to help solve complex problems. Decision- scleroatrophic gallbladder (14.4%), and pancreatitis (4.3%). making process after a IBDI continues to be complex due Two of the most important variables identified in our to every patient and scenario are different and there are models were related to the capacity of the hospital where several factors that influence the correct decision. Machine the primary surgery was performed and whether the repair learning algorithms compared to traditional logistic was performed in the same hospital. While 35.4% of the regression models or Cox proportional hazards models are IBDI in this series were diagnosed intraoperatively, most more actionable and interpretable. In addition, these com- cases were diagnosed at an average of 18 days after the chol- putational models usually have much better performance ecystectomy, with some IBDI diagnosed even years after. In than traditional logistic regressions as they can fit the data addition, also showed that most non-specialist IBDI repairs better. The use of artificial intelligence models has the spe- that were performed by the same surgeon responsible for cial advantage of their ability to analyze in more detail the the injury had very low success rates and an increased risk meaning or nature of complex or imprecise data, to extract of mortality. Immediate repair performed by an experienced patterns and detect trends than with other commonly used surgical team and in the absence of sepsis provided the best modeling approaches. Commonly used models in medical long-term results.19,20 literature are often parametric, like the linear regression Our analysis also demonstrated that type E IBDIs were or a naïve bayes, and therefore are limited by the low flex- statistically associated with poorer outcomes. Strasberg–Bis- ibility capabilities of the model. This is not the case with muth types E4 and E5 IBDI are often associated with vas- the non-parametric models used in this study, that include cular lesions.21 In our series, 33.3% of types E4 and E5 decisions trees and random forest. IBDI were associated with vascular lesions. A published These models have deeper flexibility to fit the data as series on IBDI demonstrated that concomitant injury to the they do not have the parametric constraints of the other hepatic artery was observed in 12–40% of patients, espe- models. Therefore, these models can be effectively used cially with the laparoscopic approach.22 These patients also by surgeons as part of the decision process when they need usually have a history of repeated bilioenteric anastomosis to assess the next step of a treatment. For example, given procedures and/or multiple stents and have recurrent chol- the demographic, type of IBDI, clinical and patient char- angitis and/or chronic cholestasis, which increases their risk acteristics, the surgeon board can insert into the model for secondary biliary cirrhosis, hepatic artery lesions, and different procedures to analyze the success of the repair, or eventually fulminant liver failure.23 the need to refer to a specialized hospital, obtaining repair Our risk-scoring model and random forest analysis also success probabilities of those potential decisions based on demonstrated that it was important whether the initial repair the historical data fitting the models. Analogously, this was successful and whether the primary intervention was can also be done to obtain a risk factor based on the CCI surgical. As such, the greater the number of failed attempts model. Moreover, is important to denote that this model is until definitive repair, the greater the risk of developing postoperative complications and recurrent stenosis. This 13 1720 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 was also demonstrated by a large retrospective study pub- centers (33% and 9%, respectively) than in specialized cent- lished by Reuver et al..24 A multidisciplinary approach that ers (20% and 6%, respectively) (Fig. 4c). includes therapeutic endoscopy and interventional radiol- Our study present certain limitations. The specificity of ogy is essential for the proper management of IBDI.25–28 our model decreased when we applied the model to the new However, endoscopic or percutaneous interventions should set of patients in the validation cohort. This may be due to only be considered as an alternative to surgical repair in type our methodology because we did not analyze whether the A IBDI that do not have strictures and are not associated hospitals in the development and validation cohorts were with anatomical abnormalities. In contrast, endoscopic treat- similar. There may have been differences in terms of the ment of more complex lesions, such as stenosis, is associated level of specialization, methods, and data collection in the with failure in 30–59% of cases and higher re-stenting rates. hospitals between both cohorts, which affected our mod- Overall, Roux-en-Y hepaticojejunostomy remains the gold el’s predictive ability. However, the decrease in the MAE standard for the reconstruction of damaged bile ducts during and RMSE values was not significant. On the other hand, cholecystectomy.29,30 the data for this study come from many institutions over We also classified the predicted CCI values into the low a highly evolved period of time with a retrospective data (0–25), low-medium (25–50), medium–high (50–75), and source, which can be associated with a significant amount high (75–100) risk categories. Our model performed sig- of data loss, bias, and precision. Storage of patient’s infor- nificantly worse in the validation cohort, which may also mation was not as accurate in the nineties compared to con- be attributed to differences between the hospitals in both temporary data, and this happens in most of the specialized cohorts. Our confusion matrix provided some insight. In centers if data is not collected prospectively. Even so, we Fig. 3, each cell indicates the absolute number of cases and have included the graph because it is representative of how its corresponding percentage, the more intense the color cases increased with the introduction of laparoscopy and of the cell, the higher the corresponding percentage. The how the percentage of injuries, although it has suffered a diagonal cells in our matrices have the same references and decrease, remains more or less stable. Finally, there is also predictions, which indicate that they represent the true pre- a limitation about the general applicability of these complex diction rate of each category. Moreover, the incorrect predic- modeling systems especially in community settings where tions from our model fall into the neighboring categories; many of these injuries occur and initial management may therefore, the value of the error is likely to be small and only be delayed or misdirected. This delayed care and often poor partially influenced by the sampling of each risk range. The initial therapeutic intervention may make ultimate outcomes matrices also demonstrate that the high-risk category may worse even if this machine learning is utilized as a tool once be problematic because it has the lowest positive prediction referred to a specialty center. The development of a risk rate in both cohorts. calculator where the surgeon could enter the characteristics We further analyzed the variables identified by our of the patient and evaluate the probability of success of the random forest analysis. First, we assessed if there was an possible decisions could help to generate a treatment plan, association between the CCI range and success of the ini- especially when decision making is based on the surgical tial repair. We found low CCI ranges among patients whose judgment of a seasoned biliary surgeon. initial repair was successful (Fig. 4a), which was expected. Second, we assessed if there was an association between the CCI range and type of surgical repair. We knew from Conclusion the first model that surgical repair increases the likelihood of a successful initial repair. However, the second model Management of IBDI injuries is a difficult multidiscipli- demonstrated that surgical repair was observed more often nary task, with heterogeneous results around the world as in the higher CCI ranges (Fig. 4b). We expect worse out- it includes many issues to consider. Therefore, it is almost comes with higher CCI ranges. However, this observation impossible to develop prospective randomized controlled tri- was because patients who needed surgical repair tended to als. Many of the conclusions have been obtained over time have more severe IBDI. Therefore, while the surgical repair through case series. The development of an algorithm and technique was better, we found higher CCI ranges because a risk scoring model represents a new approach based on these patients had more severe IBDI at the onset, in addi- learning matching that aims to improve the results in these tion to possible complications during repair, which would patients while being aware of the treatment of these patients. have also increased their CCI scores. Third, we assessed for Our analysis stood out above other factors that delays in an association between the CCI range and level of hospital definitive repair (i.e., due to attempts at non-surgical repair) specialization and found that there were more patients from and delayed specialized center remission resulted in poorer the medium–high and high CCI groups in non-specialized outcomes. Furthermore, not only patients with complex type E lesions but also those with less severe lesions will benefit 13 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 1721 the most from being referred to a specialized center as soon 6. Dominguez-Rosado I, Sanford DE, Liu J, Hawkins WG, Mercado as possible to begin the decision-making process. Our risk- MA. Timing of Surgical Repair After Bile Duct Injury Impacts Postoperative Complications but Not Anastomotic Patency. scoring model can then be used to estimate the patients’ Annals of surgery. 2016;264(3):544-53. level of risk based on the initial presentation and IBDI type. 7. Iwashita Y, Hibi T, Ohyama T, , Umezawa A, Takada T, Strasberg The patient’s risk category may be used to determine the SM, et al. Delphi consensus on bile duct injuries during laparo- appropriate management for IBDI. Surgical departments can scopic cholecystectomy: an evolutionary cul-de-sac or the birth pangs of a new technical framework? Journal of hepato-biliary- benefit from these models because they are based on the pancreatic sciences. 2017;24(11):591-602. actual historical data and can aid in decision-making. As 8. El Nakeeb A, Sultan A, Ezzat H, et al. Impact of referral pattern data are further collected and generated through the models, and timing of repair on surgical outcome after reconstruction of the models can be continuously improved, as well. post-cholecystectomy bile duct injury: A multicenter study. Hepa- tobiliary Pancreat Dis Int. 2021;20(1):53-60. 9. Rystedt JM, Montgomery AK. Quality-of-life after bile duct Abbreviations ASA: American Society of Anesthesiologist; AUC injury: intraoperative detection is crucial. A national case-control : Area under curve; CCI: Comprehensive Complication Index; study. HPB (Oxford). 2016;18(12):1010-6. IBDI: Iatrogenic bile duct injury; LC: Laparoscopic cholecystectomy; 10. Jajja MR, Laboe A, Hashmi S, Nadeem SO, Sayed BA, Sarmiento MAE : Mean absolute error; RMSE : Root mean square error JM. Standardizing Diagnostic and Surgical Approach to Manage- ment of Bile Duct Injuries After Cholecystectomy: Long-Term Supplementary Information The online version contains supplemen- Outcomes of Patients Treated at a High-Volume HPB Center. tary material available at https://d​ oi.o​ rg/1​ 0.1​ 007/s​ 11605-0​ 22-0​ 5398-7. Journal of gastrointestinal surgery : official journal of the Society for Surgery of the Alimentary Tract. 2021;25(11):2796-2805. 11. Clavien PA, Vetter D, Staiger RD, Slankamenac K, Mehra T, Graf Funding Open Access funding provided thanks to the CRUE-CSIC R. The Comprehensive Complication Index (CCI®): Added Value agreement with Springer Nature. and Clinical Perspectives 3 Years "Down the Line". Annals of surgery. 2017;265(6):1045-50. Declarations 12. Garcia-Vidal C, Sanjuan G, Puerta-Alcalde P, Moreno-García E, Soriano A. Artificial intelligence to support clinical decision- Conflict of Interest The authors declare no competing interests. making processes. EBioMedicine. 2019;46:27-9. 13. Chen JH, Asch SM. Machine Learning and Prediction in Medicine - Beyond the Peak of Inflated Expectations. The New England Open Access This article is licensed under a Creative Commons Attri- journal of medicine. 2017;376(26):2507-9. bution 4.0 International License, which permits use, sharing, adapta- 14. Nagendran M, Chen Y, Lovejoy CA, Gordon AC, Komorowski tion, distribution and reproduction in any medium or format, as long M, Harvey H. Artificial intelligence versus clinicians: systematic as you give appropriate credit to the original author(s) and the source, review of design, reporting standards, and claims of deep learning provide a link to the Creative Commons licence, and indicate if changes studies. BMJ (Clinical research ed). 2020;368:m689. were made. The images or other third party material in this article are 15. Tokuyasu T, Iwashita Y, Matsunobu Y, Kamiyama T, Ishikake M, included in the article’s Creative Commons licence, unless indicated Sakaguchi S, et al. Development of an artificial intelligence system otherwise in a credit line to the material. If material is not included in using deep learning to indicate anatomical landmarks during laparo- the article’s Creative Commons licence and your intended use is not scopic cholecystectomy. Surgical endoscopy. 2020 ;35(4):1651-1658. permitted by statutory regulation or exceeds the permitted use, you will 16. Brunt LM, Deziel DJ, Telem DA, Strasberg SM, Aggarwal R, need to obtain permission directly from the copyright holder. To view a Asbun H, et al. Safe Cholecystectomy Multi-society Practice copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. Guideline and State of the Art Consensus Conference on Pre- vention of Bile Duct Injury During Cholecystectomy. Annals of surgery. 2020;272(1):3-23. 17. Strasberg SM. A three-step conceptual roadmap for avoiding bile duct injury in laparoscopic cholecystectomy: an invited per- References spective review. Journal of hepato-biliary-pancreatic sciences. 2019;26(4):123-7. 1. Booij KAC, de Reuver PR, van Dieren S, van Delden OM, Rauws 18. Ejaz A, Spolverato G, Kim Y, Dodson R, Sicklick JK, Pitt HA, EA, Busch OR, et al. Long-term Impact of Bile Duct Injury on et al. Long-term health-related quality of life after iatrogenic bile Morbidity, Mortality, Quality of Life, and Work Related Limita- duct injury repair. Journal of the American College of Surgeons. tions. Annals of surgery. 2018;268(1):143-50. 2014;219(5):923-32.e10. 2. Parrilla P, Robles R, Varo E, Jiménez C, Sánchez-Cabús S, Pareja 19. Kohn JF, Trenk A, Kuchta K, et al. Characterization of common Eet al. Liver transplantation for bile duct injury after open and bile duct injury after laparoscopic cholecystectomy in a high- laparoscopic cholecystectomy. The British journal of surgery. volume hospital system. Surgical endoscopy. 2018;32(3):1184-91. 2014;101(2):63-8. 20. Wang X, Yu WL, Fu XH, et al. Early Versus Delayed Surgical 3. Koppatz H, Sallinen V, Mäkisalo H, Nordin A. Outcomes and Repair and Referral for Patients With Bile Duct Injury: A Systematic quality of life after major bile duct injury in long-term follow-up. Review and Meta-analysis. Annals of surgery. 2020;271(3):449-59. Surg Endosc. 2021;35(6):2879-88. 21. Li J, Frilling A, Nadalin S,et al. Management of concomitant 4. Cohen JT, Charpentier KP, Beard RE. An Update on Iatrogenic hepatic artery injury in patients with iatrogenic major bile duct Biliary Injuries: Identification, Classification, and Management. injury after laparoscopic cholecystectomy. The British journal of The Surgical clinics of North America. 2019;99(2):283-99. surgery. 2008;95(4):460-5. 5. Ahmad DS, Faulx A. Management of Postcholecystectomy Bil- 22. Strasberg SM, Helton WS. An analytical review of vasculobiliary iary Complications: A Narrative Review. The American journal injury in laparoscopic and open cholecystectomy. HPB (Oxford). of gastroenterology. 2020;115(8):1191-8. 2011;13(1):1-14. 13 1722 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 23. Booij KAC, Coelen RJ, de Reuver PR, et al. Long-term follow-up 28. Pitt HA, Sherman S, Johnson MS, et al. Improved outcomes and risk factors for strictures after hepaticojejunostomy for bile of bile duct injuries in the 21st century. Annals of surgery. duct injury: An analysis of surgical and percutaneous treatment 2013;258(3):490-9. in a tertiary center. Surgery. 2018;163(5):1121-7. 29. Post cholecystectomy bile duct injury: early, intermediate or late 24. de Reuver PR, Grossmann I, Busch OR,et al. Referral pattern repair with hepaticojejunostomy - an E-AHPBA multi-center and timing of repair are risk factors for complications after study. HPB (Oxford). 2019;21(12):1641-7. reconstructive surgery for bile duct injury. Annals of surgery. 30. Ismael HN, Cox S, Cooper A, et al. The morbidity and mortality 2007;245(5):763-70. of hepaticojejunostomies for complex bile duct injuries: a multi- 25. Haidar H, Manasa E, Yassin K, Suissa A, Kluger Y, Khamaysi I. institutional analysis of risk factors and outcomes using NSQIP. Endoscopic treatment of post-cholecystectomy bile leaks: a ter- HPB (Oxford). 2017;19(4):352-8. tiary center experience. Surg Endosc. 2021;35(3):1088-92. 26. Pioche M, Ponchon T. Management of bile duct leaks. J Visc Surg. Publisher's Note Springer Nature remains neutral with regard to 2013;150(3 Suppl):S33-8. jurisdictional claims in published maps and institutional affiliations. 27. Lillemoe KD, Martin SA, Cameron JL, et al. Major bile duct injuries during laparoscopic cholecystectomy. Follow-up after combined surgical and radiologic management. Ann Surg. 1997;225(5):459–68; discussion 68–71. 13 Journal of Gastrointestinal Surgery (2022) 26:1713–1723 1723 Authors and Affiliations Victor Lopez‑Lopez1 · Javier Maupoey2 · Rafael López‑Andujar2 · Emilio Ramos3 · Kristel Mils3 · Pedro Antonio Martinez4 · Andres Valdivieso5 · Marina Garcés‑Albir6 · Luis Sabater6 · Luis Díez Valladares7 · Sergio Annese Pérez8 · Benito Flores8 · Roberto Brusadin1 · Asunción López Conesa1 · Valentin Cayuela1 · Sagrario Martinez Cortijo9 · Sandra Paterna10 · Alejando Serrablo10 · Santiago Sánchez‑Cabús11 · Antonio González Gil12 · Jose Antonio González Masía13 · Carmelo Loinaz14 · Jose Luis Lucena15 · Patricia Pastor16 · Cristina Garcia‑Zamora17 · Alicia Calero18 · Juan Valiente19 · Antonio Minguillon20 · Fernando Rotellar21,22 · Jose Manuel Ramia23 · Cándido Alcazar23 · Javier Aguilo24 · Jose Cutillas25 · Christoph Kuemmerli26,27 · Jose A. Ruiperez‑Valiente4 · Ricardo Robles‑Campos1 1 14 Department of Surgery and Transplantation, Virgen de La Department of General Surgery, Digestive Tract Arrixaca Clinic and University Hospital, Murcian Institute and Abdominal Organ Transplantation, Hospital of Biosanitary Research (IMIB), Ctra. Madrid-Cartagena, Universitario, 12 de Octubre, Madrid, Spain 30120 s/nEl Palmar, Murcia, Spain 15 Department of Surgery, Puerta del Hierro University 2 Department of Hepatobiliary Surgery and Transplants, Hospital, Madrid, Spain Hospital Universitario La Fe, Valencia, Spain 16 Department of Surgery, Reina Sofía University Hospital, 3 Department of Hepatobiliary Surgery and Liver Transplant Murcia, Spain Unit, Hospital Universitari Bellvitge, University 17 Department of Surgery, Rafael Mendez Hospital, Murcia, of Barcelona, Barcelona, Spain Spain 4 Department of Information and Communication Engineering, 18 Department of General Surgery, Elche University Hospital, Universidad de Murcia, Murcia, Spain University Miguel Hernández of Elche, Alicante, Spain 5 Hepatobiliary Surgery and Liver Transplant Unit, Cruces 19 Department of General Surgery, Hellin Hospital, Albacete, University Hospital, Bilbao, Spain Spain 6 Department of Surgery, Hospital Clínico, University 20 Department of General Surgery, Obispo Polanco Hospital, of Valencia, Biomedical Research Institute INCLIVA, Teruel, Spain Valencia, Spain 21 7 HPB and Liver Transplant Unit, Department of General Department of Surgery, Hepatopancreatobiliary Unit, Surgery, Clinica Universidad de Navarra, Universidad de Hospital Clínico San Carlos, Madrid, Spain Navarra, Pamplona, Spain 8 Department of Surgery, Morales University Hospital, 22 Institute of Health Research of Navarra (IdisNA), Pamplona, Madrid, Spain Spain 9 Department of Surgery, Alcorcon University Hospital, 23 HPB Surgery and Liver Transplantation, Dr. Balmis General Madrid, Spain University Hospital, and Alicante Institute for Health 10 Department of Surgery, Miguel Servet University Hospital, and Biomedical Research (ISABIAL), Alicante, Spain Saragossa, Spain 24 Department of General Surgery, Hospital Lluís Alcanyís 11 Hepatobiliopancreatic Surgery Unit, Hospital de La Santa Hospital, Xàtiva, Valencia, Spain Creu I Sant Pau, Universitat Autònoma de Barcelona, 25 Department of General Surgery, Hospital Francesc de Borja, Barcelona, Spain Gandía, Valencia, Spain 12 Department of Surgery, Los Arcos del Mar Menor University 26 Department of Surgery and Transplantation, University Hospital, Murcia, Spain Hospital Zurich, Zurich, Switzerland 13 Department of Surgery, General University Hospital, 27 Department of Surgery, Clarunis - University Centre Albacete, Spain for Gastrointestinal and Liver Diseases, Basel, Switzerland 13