Artificial intelligence-assisted multidisciplinary therapy for a complex case of cholangitis with septic shock: a case report and simulated decision-making analysis

Introduction

Choledocholithiasis complicated by acute cholangitis represents a critical condition in hepatobiliary surgery. For patients with a history of previous biliary surgery and recurrent infections, the condition is highly prone to progression to sepsis, septic shock, and multiple organ dysfunction, leading to a significantly increased mortality rate (1,2). In such patients, biliary obstruction often triggers a dysregulated systemic inflammatory response, accompanied by coagulopathy, thrombocytopenia, and hepatic dysfunction, which collectively increase the difficulty and complexity of clinical management (3). Currently, the management of complex biliary tract infections complicated by sepsis is heavily reliant on a multidisciplinary team (MDT) approach. The core components of this strategy encompass infection control, endoscopic or surgical intervention to relieve obstruction, coagulation management, and organ function support.

However, primary care institutions often encounter challenges when implementing the conventional MDT model. These include insufficient allocation of specialized resources, such as the absence of on-call endoscopic teams during night shifts, as well as decision-making dilemmas like the contradictory timing requirements for antimicrobial versus anticoagulant therapies, thereby hindering the optimization of dynamic clinical decision-making (4). Against this backdrop, the technical advantages of artificial intelligence (AI) in simulating MDT become prominent. AI can integrate multi-dimensional data in real-time, simulate expert decision-making logic, and generate individualized strategies for critical junctures such as the timing of biliary drainage, the therapeutic window for antibiotic administration, and the threshold for correcting coagulopathy, thereby providing decision support for resource-limited settings (5). Currently, clinical reports on the application of AI in such critical illness domains remain scarce. This study reports a case of a patient with recurrent choledocholithiasis and shock. In parallel with real-world MDT management, an AI-MDT decision simulation and comparative analysis were conducted, aiming to provide empirical evidence and conceptual insights for exploring the application of AI technology in the field of critical and emergency care. We present this article in accordance with the CARE reporting checklist (available at https://acr.amegroups.com/article/view/10.21037/acr-2026-0055/rc).

Case presentation

Admission status and initial clinical data

A 52-year-old female patient was admitted on February 1, 2025, with a chief complaint of “abdominal pain for over 3 years, recurrent for 1 day”. The patient had a history of biliary stones, with multiple prior episodes of abdominal pain managed by treatments including endoscopic retrograde cholangiopancreatography (ERCP) for stone extraction and cholecystectomy. Upon this admission, she presented with persistent colicky pain in the right upper quadrant, accompanied by chills, fever, nausea, vomiting, and scleral icterus. Initial assessment led to a diagnosis of choledocholithiasis with low biliary obstruction, septic shock, and coagulopathy. Her past surgical history included an open-heart repair for an atrial septal defect in 1994, with no regular follow-up thereafter.

• Vital signs on admission: temperature 37.9 ℃, heart rate (HR) 156 bpm, respiratory rate 20 bpm, blood pressure (BP) 91/45 mmHg, oxygen saturation 97%.
• Physical examination: an old surgical scar was visible on the abdomen. The abdomen was soft, with tenderness in the epigastric region and no rebound tenderness. Bowel sounds were normal.
• Laboratory investigations: white blood cell count 30.93×10⁹/L, neutrophils 96.20%, hemoglobin 111.00 g/L, platelet count 151.00×10⁹/L. Coagulation profile: prothrombin time (PT) 18.2 s (normal 11–13.5 s), activated partial thromboplastin time (APTT) 45.1 s (normal 25–35 s), international normalized ratio (INR) 1.8. Total bilirubin 99.4 µmol/L, direct bilirubin 85.6 µmol/L, aspartate aminotransferase 132 U/L, alanine aminotransferase 148 U/L, serum amylase 74 U/L.
• Abdominal CT findings: stones in the common hepatic duct and common bile duct causing low biliary obstruction, with minimal pneumobilia in the intrahepatic and extrahepatic biliary ducts; status post cholecystectomy; small amount of ascites; post-surgical changes of the sternum (Figure 1). Justification for CT over bedside ultrasound in this hemodynamically unstable patient: (I) the patient was transiently stabilized with fluid and low-dose norepinephrine (MAP >65 mmHg for 30 minutes); (II) the CT scanner is located within the same critical care unit (20-meter transport with full monitoring); (III) a rapid, non-contrast protocol (<2 minutes) was used; and (IV) initial bedside ultrasound was technically limited by overlying bowel gas.

Figure 1 Imaging findings. (A) Non-contrast abdominal CT shows a hyperdense stone (red arrow) in the distal common bile duct. (B) Non-contrast abdominal CT demonstrates the ENBD tube (red arrow) positioned in the common bile duct after emergency decompression. (C,D) MRCP confirms multiple filling defects (red arrows) in the common hepatic and common bile ducts, consistent with choledocholithiasis. Red arrows in all panels indicate the location of biliary stones or the drainage tube as specified. CT, computed tomography; ENBD, endoscopic nasobiliary drainage; MRCP, magnetic resonance cholangiopancreatography.

Initial emergency intervention and clinical course

On the day of admission (February 1, 2025), an emergency ERCP with endoscopic nasobiliary drainage (ENBD) placement was performed, yielding approximately 3 mL of dark brown, turbid bile, which alleviated the biliary obstruction. The initial choice of ENBD over primary endoscopic retrograde biliary drainage (ERBD) was based on: (I) the patient’s coagulopathy (INR 1.8)—ENBD avoids the bleeding risk of sphincterotomy required for most ERBD placements; (II) the need to monitor bile output and obtain serial bile cultures in septic shock; and (III) the principle of a reversible, non-permanent drain while acute physiology is stabilized. Given the patient’s coagulopathy and high risk of bleeding, aggressive stone removal was not attempted. The initial anti-infective regimen consisted of cefoperazone-sulbactam (2 g, q8h) combined with moxifloxacin (0.4 g, qd). Comprehensive organ support therapy was concurrently administered, including hepatoprotective agents (adenosylmethionine 1 g, qd and magnesium isoglycyrrhizinate), gastric protection with rabeprazole, acid and enzyme suppression with octreotide, and vitamin K1. Norepinephrine was used to maintain hemodynamic stability. Furthermore, frozen plasma and platelet transfusions were given to correct the coagulation disorder. Due to persistent clinical deterioration (worsening hypotension, procalcitonin rising from 2.5 to 18.6 ng/mL) and the institutional antibiogram showing high rates of ESBL-producing Escherichia coli in biliary isolates, antibiotic therapy was escalated empirically to imipenem-cilastatin sodium (1 g, q8h) on February 2 following MDT discussion. Blood and bile cultures were obtained prior to escalation and subsequently grew ESBL-producing Escherichia coli.

During the course of treatment, the patient’s serum amylase levels rose progressively: post-ERCP day 1: 214 U/L; day 2: 486 U/L (peak); day 3: 392 U/L. Considering the suspected possible mechanical obstruction of the pancreatic duct by the ENBD tube (based on the temporal relationship—amylase rise beginning 24 hours after ENBD placement, absence of pancreatic edema or necrosis on repeat CT, and lack of alternative obvious cause), the ENBD was exchanged for an ERBD stent on February 7. Post-exchange, amylase levels normalized within 12 hours (peak 486→78 U/L). Post-ERCP pancreatitis was systematically ruled out using clinical criteria (no new abdominal pain, no vomiting) and CT criteria (no pancreatic edema or peripancreatic fluid). The anti-infective regimen was subsequently de-escalated back to cefoperazone-sulbactam (Figure 2).

Figure 2 Therapeutic algorithm. ENBD, endoscopic nasobiliary drainage; ERCP, endoscopic retrograde cholangiopancreatography; NPO, nil per os; q8h, every 8 hours; qd, every day; TCM, traditional Chinese medicine.

The definitive management plan for the residual stones (intentionally left during the initial ERCP) is elective, delayed choledochoscopy with laser lithotripsy via the future percutaneous transhepatic cholangial drainage (PTCD) tract, scheduled for 6–8 weeks after discharge once the patient has fully recovered from septic shock and coagulopathy has been definitively corrected. The patient remains on low-dose ursodeoxycholic acid as a bridge.

All procedures performed in this case were in accordance with the ethical standards of the institutional and/or national research committee(s) and with the Declaration of Helsinki and its subsequent amendments. Written informed consent was obtained from the patient for publication of this case report and accompany images. A copy of the written consent is available for review by the editorial office of this journal.

Through the comprehensive management outlined above, the infection was brought under control and the patient’s condition gradually stabilized. She was transferred to a traditional Chinese medicine hospital on February 13, 2025, for subsequent rehabilitation and recuperation. During the follow-up period, the patient remained afebrile and did not experience recurrence of abdominal pain or jaundice over the subsequent 1 year.

AI simulation methodology

To evaluate the auxiliary potential of AI in complex clinical decision-making concurrently with the real-world MDT process, we employed two large language models (DeepSeek and ChatGPT-5) to simulate MDT decision-making. The following verbatim prompt was provided to both models on February 3, 2025 (query date). Model versions: DeepSeek (DeepSeek-V3) and ChatGPT-5 (GPT-5-turbo).

Prompt: *“You are an artificial intelligence serving as a multidisciplinary team (MDT) simulator for a complex case of septic shock secondary to choledocholithiasis. A 52-year-old female presents with: BP 91/45 mmgH, HR 156 bpm, temperature 37.9 ℃, white blood cell 30.93×10⁹/L, INR 1.8, total bilirubin 99.4 µmol/L, direct bilirubin 85.6 µmol/L, amylase 74 U/L. CT shows common bile duct stone with pneumobilia. She has a history of prior ERCP and cholecystectomy. An emergency ERCP with ENBD has just been performed, but the patient remains hypotensive requiring norepinephrine. Amylase is rising (peak 486 U/L). Coagulopathy persists (INR 1.6). Please provide your recommendations from the perspectives of: (I) infectious diseases; (II) ultrasonography/interventional radiology; (III) gastroenterology, and (IV) hepatobiliary surgery. Provide specific, actionable recommendations with brief reasoning”.**

The AI outputs were recorded verbatim and compared to the real-world MDT decisions. Table 1 presents this comparative analysis.

Discussion

This case demonstrates an extremely complex pathophysiological process involving biliary stones leading to septic shock, coagulopathy, and pancreatitis. Its successful management exemplifies the core value of a patient-centered MDT approach (1,6). The real-world MDT successfully interrupted the vicious cycle of disease progression by precisely timing the critical intervention of biliary drainage and dynamically adjusting the anti-infective and supportive treatment regimens (7).

This study innovatively introduced large AI models into MDT decision simulation. The results demonstrated that both DeepSeek and ChatGPT-5 showed high consistency with the real MDT regarding core therapeutic principles, uniformly recommending immediate biliary decompression and escalation of antibiotic therapy to potent broad-spectrum agents (8). This finding strongly demonstrates that AI, through training on vast corpora of medical literature and clinical guidelines, has developed the capability to effectively internalize and apply the core principles central to sepsis resuscitation (9,10). The rapid knowledge integration and structured output capabilities demonstrated by AI indicate its potential to serve as an efficient “auxiliary cognitive support system” for clinicians, particularly for primary care physicians in emergency situations, aiding in the rapid anchoring of the correct diagnostic and therapeutic trajectory (11,12).

Despite alignment on core principles, a notable divergence emerged between the AI and the real-world MDT regarding specific implementation strategies, most prominently in the selection of biliary drainage modalities. The AI models exhibited a tendency to recommend PTCD or present multiple options based on theoretical risk-benefit models. In contrast, the real-world MDT integrated a complex array of practical considerations. To provide a granular “micro-reality” analysis: (I) an experienced therapeutic endoscopist was immediately available at our tertiary center, with a personal volume of >200 ERCPs/year; (II) the patient’s biliary anatomy was favorable for repeat ERCP (no Roux-en-Y, no periampullary diverticulum); (III) the MDT considered the logistical burden of managing an external PTCD tube for 6–8 weeks prior to definitive choledochoscopy; (IV) the patient expressed a strong preference to avoid an external drain for cultural and cosmetic reasons; and (V) while PTCD is often cited in guidelines for coagulopathy, our interventional radiology team judged that ultrasound-guided PTCD through a dilated tract still carried a non-trivial bleeding risk (INR 1.6 at the time of decision). This comprehensive situational analysis led to a more forward-looking decision: exchanging the ENBD for an ERBD stent. This discrepancy underscores a core limitation of current general-purpose AI: its deficiency in deep perception and quantitative assessment of “clinical reality feasibility” (13,14). It excels at being “evidence-based” but falls short in being “context-adapted”, and cannot fully replicate the integrated art of trade-off that human experts exercise in specific clinical scenarios (15).

The “Decision Simulation Analysis” methodology employed in this study elevates the case analysis beyond traditional descriptive narratives, advancing it into a new, quantifiable, and comparative dimension of decision analysis. This approach enables the systematic deconstruction of the constituent elements of clinical reasoning, clearly distinguishing between universal principles and contextualized trade-offs. It holds significant value for advancing both clinical reasoning education and the assessment of decision-making quality (16,17). Furthermore, this case provides a reference for a standardized protocol for managing such critical and emergency conditions in primary care hospitals. It also demonstrates the utility of AI tools in offering initial decision support, a capability of particular significance in resource-limited settings (11,18).

This study also has several limitations. First, as a single case report, the conclusions require validation through studies with larger sample sizes. Second, the AI simulation was based on a single prompt and did not reflect the multi-round, dynamic decision-making process inherent in clinical practice. Third, and importantly, we acknowledge that more sophisticated prompting techniques—such as Chain-of-Thought (CoT) prompting, self-consistency checks, and iterative multi-turn conversational interfaces—might have allowed the AI to better approximate human MDT reasoning. Our single-prompt design therefore represents a conservative, lower-bound estimate of AI capability. Future research should employ dynamic, conversational AI interfaces to better simulate real clinical deliberation. Most importantly, general-purpose large language models lack integration of “micro-reality” data, such as regional antimicrobial resistance profiles and local healthcare resource availability. Future research should focus on developing specialized and localized clinical decision support systems. By deeply integrating regional data and conducting rigorous clinical validation, such efforts can bridge the gap between AI capabilities and real-world clinical scenarios.

Conclusions

This case confirms that the MDT approach remains the gold standard for managing complex biliary septic shock. The collective wisdom and clinical experience of human experts are irreplaceable in integrating dynamic information and weighing practical feasibility. Large AI models demonstrate significant potential as powerful auxiliary tools. The principled decision support they provide is of high quality and rapid in response, indicating important prospects for clinical application. However, current general-purpose AI lacks deep perception of “clinical reality feasibility” and cannot fully replicate human contextual adaptation. Advancing the deep integration of AI with clinical practice to construct an optimized diagnostic and therapeutic paradigm based on human-AI collaboration represents a critical direction for enhancing the management of critical and emergency illnesses in the future.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the CARE reporting checklist. Available at https://acr.amegroups.com/article/view/10.21037/acr-2026-0055/rc

Peer Review File: Available at https://acr.amegroups.com/article/view/10.21037/acr-2026-0055/prf

Funding: This work was supported by the National Natural Science Foundation of China (Nos. 81802874 and 82260548) and the Natural Science Foundation of Guangxi Province of China (No. 2024GXNSFAA010347).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://acr.amegroups.com/article/view/10.21037/acr-2026-0055/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. All procedures performed in this case were in accordance with the ethical standards of the institutional and/or national research committee(s) and with the Declaration of Helsinki and its subsequent amendments. Written informed consent was obtained from the patient for publication of this case report and accompany images. A copy of the written consent is available for review by the editorial office of this journal.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Smith SE. Management of Acute Cholangitis and Choledocholithiasis. Surg Clin North Am 2024;104:1175-89. [Crossref] [PubMed]
2. Gadiyaram S, Nachiappan M. The second ‘gallbladder operation’. J Minim Access Surg 2022;18:596-602. [Crossref] [PubMed]
3. Belyaev AM, Thwaite P, Rossaak J, et al. The Use of Inflammatory Markers for Treatment Response Monitoring in Acute Cholangitis: A Retrospective Cohort Study. J Surg Res 2024;293:14-21. [Crossref] [PubMed]
4. Katz SJ. Staff and resident perceptions on the introduction of a team based multi-specialty resident night shift system. PLoS One 2022;17:e0268569. [Crossref] [PubMed]
5. Huang S, Zhou Y, Liang Y, et al. Machine-learning-derived online prediction models of outcomes for patients with cholelithiasis-induced acute cholangitis: development and validation in two retrospective cohorts. EClinicalMedicine 2024;76:102820. [Crossref] [PubMed]
6. Matsubara S. Clinical Advances in the Diagnosis and Treatment of Biliary Tract Diseases. J Clin Med 2023;12:2282. [Crossref] [PubMed]
7. See TC. Acute biliary interventions. Clin Radiol 2020;75:398.e9-398.e18. [Crossref] [PubMed]
8. ASGE Standards of Practice Committee. American Society for Gastrointestinal Endoscopy guideline on the role of therapeutic EUS in the management of biliary tract disorders: summary and recommendations. Gastrointest Endosc 2024;100:967-79. [Crossref] [PubMed]
9. Li F, Wang S, Gao Z, et al. Harnessing artificial intelligence in sepsis care: advances in early detection, personalized treatment, and real-time monitoring. Front Med (Lausanne) 2024;11:1510792. [Crossref] [PubMed]
10. Desposito L, Bascara C. Review: sepsis guidelines and core measure bundles. Postgrad Med 2024;136:702-11. [Crossref] [PubMed]
11. Kareemi H, Yadav K, Price C, et al. Artificial intelligence-based clinical decision support in the emergency department: A scoping review. Acad Emerg Med 2025;32:386-95. [Crossref] [PubMed]
12. Fast D, Adams LC, Busch F, et al. Autonomous medical evaluation for guideline adherence of large language models. NPJ Digit Med 2024;7:358. [Crossref] [PubMed]
13. Hager P, Jungmann F, Holland R, et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat Med 2024;30:2613-22. [Crossref] [PubMed]
14. Naeem A, Khan O, Baqir SM, et al. Language Artificial Intelligence Models as Pioneers in Diagnostic Medicine? A Retrospective Analysis on Real-Time Patients. J Clin Med 2025;14:1131.
15. Wu J, Guyatt G, Yao L. Efficiency must not compromise trustworthiness in rating certainty and formulating recommendations in AI era. BMJ 2025;389:r1105. [Crossref] [PubMed]
16. Deschênes MF, Fernandez N, Lechasseur K, et al. Transformation and Articulation of Clinical Data to Understand Students’ and Health Professionals’ Clinical Reasoning: Protocol for a Scoping Review. JMIR Res Protoc 2023;12:e50797. [Crossref] [PubMed]
17. Araújo B, Gomes SF, Ribeiro L. Critical thinking pedagogical practices in medical education: a systematic review. Front Med (Lausanne) 2024;11:1358444. [Crossref] [PubMed]
18. Wang C, Wang F, Li S, et al. Patient Triage and Guidance in Emergency Departments Using Large Language Models: Multimetric Study. J Med Internet Res 2025;27:e71613. [Crossref] [PubMed]