Acute cholangitis due to cholelithiasis (CIAC) is an acute inflammatory disease with poor prognosis

This study aims to create machine learning (ML) models to predict outcomes in patients with CIAC

In this retrospective cohort and ML study, according to the International Classification of Diseases (ICD ) revised to the 9th edition, meeting the diagnosis of "cholangitis" and "gallbladder or bile duct stones" at the same time, or meeting the diagnosis of "bile duct stones with acute cholangitis with or without obstruction" during a single hospitalization according to the 10th edition of the ICD Patients were included in the Critical Care Medicine Information Mart database, which records patient admissions to Beth Israel Deaconess Medical Center in Massachusetts, USA, between June 1, 2001, and November 16, 2022.

The study focused on three primary endpoints, including in-hospital mortality, readmission within 30 days of discharge, and mortality within 180 days of discharge

To mitigate bias arising from missing data, factors exhibiting more than 20% missing values ​​were excluded during the data collection stage

Multiple imputation (MI) techniques were employed to account for missing values ​​in the remaining variables for analysis.

Variables independently associated with outcome were identified by univariate logistic regression in a training subset of patients with CIAC

Patients who were admitted for non-emergency admission, who did not undergo biliary drainage within 24 hours after admission, who were younger than 18 years old, and whose information loss exceeded 20% were excluded.

Nine machine learning methods were used to predict patients' in-hospital mortality, readmission rate within 30 days after discharge, and mortality within 180 days after discharge.

Select patients who were treated at Zhongda Hospital Affiliated to Southeast University from January 1, 2019 to July 30, 2023 as the external validation set

The area under the receiver operating characteristic curve is the main indicator for model performance evaluation.

A total of 1156 patients were included to construct the model

Analyzes were stratified for all patients, patients admitted to the intensive care unit (ICU), and patients who underwent biliary drainage during ICU treatment

Screen out 13~16 features from 186 variables for model training

The XGBoost method showed the best prediction performance, with training set AUROCs of 0.996 (in-hospital mortality rate), 0.886 (readmission rate within 30 days after discharge) and 0.988 (mortality rate within 180 days after discharge)

0.998, 0.933 and 0.988 (patients admitted to ICU)

0.987, 0.908 and 0.982 (patients undergoing biliary drainage during ICU treatment)

In the internal validation set, the AUROC for all-patient in-hospital mortality reached 0.967, the AUROC for readmission rate within 30 days after discharge reached 0.589, and the AUROC for mortality within 180 days after discharge reached 0.857

The AUROC of patients admitted to ICU reached 0.963, 0.668, 0.864

The AUROC of patients who underwent biliary drainage during ICU treatment reached 0.961, 0.669, and 0.828

The AUROC values ​​of the external validation set composed of 61 patients were 0.741, 0.812, and 0.848 respectively.

The XGBoost model is expected to become a tool for predicting the prognosis of CIAC patients and has good clinical applicability.

A series of user-friendly online prediction platforms were built based on the XGBoost model, which can track multiple short-term or long-term clinical outcomes of CIAC patients, which requires further validation with multi-center cohorts and larger sample sizes.

Includes baseline characteristics of all patients

Different results correspond to different patient categories

Best machine learning algorithms for predicting various outcomes corresponding to patient categories

Link to web tool predicting different outcomes for patients with CIAC

Overall flow chart of the study (a)

Algorithm diagram of the study (b)

Feature selection based on SelectFromModel algorithm

For all patients admitted to the ICU and patients who received biliary drainage during ICU treatment, 16, 13, 15 and 15, 14, and 16 variables were selected respectively when constructing the prediction model for in-hospital mortality and readmission within 30 days after admission.

After comprehensive analysis of all patients, it was found that 13 key variables have the greatest impact on the occurrence of in-hospital death.

The horizontal axis represents the name of each variable, and the vertical axis represents the importance of each variable.

Outcomes for all patients (a-c): in-hospital mortality, readmission within 30 days of discharge, and mortality within 180 days of discharge

Outcomes of patients admitted to ICU (d–f)

Outcomes of patients who underwent biliary drainage during ICU treatment (g-i)

ROC curves and PR curves of nine models

Decision curve analysis of XGBoost and the model with the most controversial prediction effect except XGBoost

Outcomes for all patients: in-hospital mortality (a), readmission within 30 days of discharge (b), mortality within 180 days of discharge (c)

Web tool usage examples

In-hospital mortality prediction by entering 15 clinical parameters from individual patients undergoing biliary drainage during ICU care indicating poor prognosis

Receiver operating characteristic curves (a-f) and precision-recall curves (g-l) of the nine models

Decision curve analysis of XGBoost and the model with the most controversial prediction effect except XGBoost

Receiver operating characteristic curve of XGBoost in the external validation set

Baseline characteristics of patients admitted to ICU (n=652)

Baseline characteristics of patients who underwent surgery during ICU care (n=614)

Summary table of conversion relationship between variable data units collected by Zhongda Hospital and existing data units in MIMIC database

Analysis of factors leading to different outcomes across all patients (n=1156)

Differences in characteristics between patients with different prognostic outcomes are detailed in Supplementary Table S4

Baseline characteristics of patients readmitted within 30 days of discharge (n=262)

Factor analysis of whether patients readmitted within 30 days died within 180 days of discharge (univariate logistic regression, n=262)

Factor analysis of whether patients who were readmitted within 30 days died within 180 days of discharge (multivariate logistic regression analysis, n=262)

Analyze whether patients readmitted within 30 days die within 180 days of discharge (excluding in-hospital deaths, n=1104)

Analysis of factors leading to different outcomes in patients admitted to ICU (n=652)

Analysis of influencing factors on different outcomes of patients undergoing surgery during ICU treatment (n=614)

Variables classified according to the type of data distribution corresponding to various patient categories

Different patient categories correspond to influencing factors of different outcomes (P<0.1)

Summary of variables finally incorporated into each model after feature selection

Performance of machine learning algorithms (predicting in-hospital mortality for all patients)

Performance of machine learning algorithms (predicting readmission within 1 month of discharge for all patients)

Performance of machine learning algorithms (predicting mortality within 180 days of discharge for all patients)

Performance of machine learning algorithms (predicting in-hospital mortality in patients admitted to ICU)

(Predicting readmission within 1 month of discharge for patients admitted to ICU)

(Predicting mortality within 180 days of discharge for patients admitted to ICU)

Performance of machine learning algorithms (predicting in-hospital mortality in patients undergoing surgery during ICU care)

(Predicting readmission within 1 month of discharge for patients who underwent surgery during ICU care)

(Predicting mortality within 180 days of discharge in patients undergoing surgery during ICU care)

Variables of patients who received biliary drainage during ICU treatment in Zhongda Hospital (n=61)