Acute pancreatitis (AP) is a common cause of gastroenterology-related hospitalizations, with a morbidity rate of 34 per 100,000 individuals globally (Zhou et al., 2022). Although AP is inclined to be self-limiting, around 20% of patients will progress to severe AP (SAP), with persistent organ failure (POF) and poor prognosis (Wu et al., 2021). Therefore, early detection of SAP and early treatment such as fluid resuscitation are dispensable for reducing the morbidity and mortality of SAP.

Conventional scoring systems such as the RANSON score, bedside index of severity in acute pancreatitis (BISAP), modified computed tomography severity index (MCTSI), and Acute Physiology and Chronic Health Evaluation (APACHE) II have been generally applied to assess the severity of AP (Bollen et al., 2012; Mounzer et al., 2012). Some novel point systems, such as SABP (Hong et al., 2019), the pancreatic activity scoring system (PASS), and the Chinese simple scoring system (CSSS) (Wu et al., 2021), have been proposed in recent years. However, the traditional scores are relatively complicated for clinical use, and the novel scores are not generalized, whose ability to predict SAP varies and accuracy ranges from 0.70 to 0.95 (Hong et al., 2019; Paragomi et al., 2021; Wu et al., 2021).

Machine learning (ML) applied in medicine, both supervised and unsupervised, is becoming increasingly popular based on its efficient computing algorithms to learn from massive clinical data (Deo, 2015). Previous studies (Han et al., 2022; Liu et al., 2022; Xu et al., 2021; Yang et al., 2021; Zhang et al., 2021; Zhao et al., 2021; Zhou et al., 2021) have confirmed that ML has great potential in building models for disease diagnosis, prognosis prediction, survival analysis, etc. Traditional ML includes logistic regression (LR), support vector machine (SVM), random forest, etc. A novel ML called automated machine learning (AutoML) intelligently selects from various algorithms and hyperparameters to create models customized to target data. It takes less time to develop more accurate models using intelligent early stopping, cross validation, regularization, and hyperparameter optimization when compared to traditional ML.

Our study aims to train, validate, and test a series of ML models for early prediction of SAP within 72-h hospitalization using the H2O AutoML platform in multiple centers. Additionally, traditional logistic regression (LR) analysis is also developed, as well as four existing scoring systems.

In this study, we developed and tested several AutoML models to early identify who would progress to SAP. These models were all superior to existing scoring systems such as BISAP, RANSON, and MCTSI. Additionally, the GBM model obtained the highest value of AUC above 0.90, with specificity and accuracy all above 0.95.

Early prediction of SAP patients is essential for determining which patients require appropriate management such as intensive care, rapid fluid resuscitation, and early enteral nutrition (Gliem et al., 2021). Up to now, various scoring systems have been developed for early risk stratification of AP patients. Hong et al. (2019) built a prediction score called SABP, consisting of SIRS, albumin, blood urea nitrogen (BUN), and PE, which was trained and validated on 700 and 194 patients from two hospitals. The AUC of SABP on the external validation cohort in their study was superior to that in our study (0.873 vs. 0.673). This difference may be partly explained by the fact that BUN was not a routine examination within 24 h of admission in our hospital. Data deficiency may decrease the efficacy of the SABP score due to the reduction of sample size. Typical models such as RANSON, MCTSI, and BISAP in our study achieved inferior accuracy to the models we built. Besides the probability mentioned above, the other possible explanation may be that relevant indicators of those scores were categorical variables while were converted into continuous variables in our study, which may increase the probability of false-positive results due to the decrease in threshold value.

Compared with traditional univariate and sequent multivariate analyses, AutoML greatly improved work efficiency due to its less time consumption and higher accuracy. Additionally, ensemble models combined various machine-learning algorithms, utilizing multiclassifiers to predict the target outcome via taking a vote of individual predictions, which could enhance the overall performance (Goh et al., 2021). In this study, we selected four models built by five types of AutoML algorithms (GBM, XGBoost, DRF, GLM, and DL) for predicting the risk of SAP. All models, among which the GBM model ranking first in AUC and XGBoost and in accuracy on the test dataset, yielded satisfactory results. AUC gives a more feasible method to settle the problem of unbalanced data by putting the same weight on both classes in contrast to accuracy (Janitza et al., 2013). Additionally, since our aim is to early detect high-risk AP patients who would progress to SAP, sensitivity is a better choice, which is calculated as the ratio of subjects predicted positive with our proposed models to patients who are actually positive. Therefore, the GBM model was the best one in our study.

The SHAP analysis demonstrated that the occurrence of PE at admission was the most important feature for the GBM model. Yan et al. (2021) conducted research on pleural effusion volume (PEV) for predicting the severity of acute pancreatitis, with an AUC of 0.8158. Similarly, a study from Peng et al. in 2019 (Peng et al., 2020) also revealed that PEV holds a high accuracy (AUC = 0.839) for predicting the occurrence of SAP. In our study, PE was a common important feature selected not only by GBM but also by LASSO, BISAP, MCTSI, and SABP, indicating that PE and PEV were indeed reliable radiologic biomarkers in the prediction of SAP. Albumin has been proven as an independent risk factor for SAP according to previous studies (Hong et al., 2017; Hong et al., 2019; Wu et al., 2020; Xu et al., 2020). Hong et al. (2017) concluded that hypoalbuminemia within 24 h of hospital admission was greatly associated with increased probability of occurrence of POF and death in AP patients. A large-scale retrospective study analyzed the two open-access ICU databases to reveal the predictive significance of serum albumin in patients with AP (Xu et al., 2020). Chen and colleagues (Chen et al., 2021) carried out a subanalysis in hypertriglyceridemia pancreatitis populations for exploring the association between albumin and severity of AP. It was generally believed that elevated level of TG would drive the occurrence of SAP due to toxic effects on pancreatic acinar cell (Chen et al., 2021). The free fatty acids, hydrolyzed by pancreatic lipase from TGs, can bind to albumin in the serum and thus stimulate the inflammatory process. Therefore, Chen’s study effectively ruled out the confounding effect of TG and demonstrated that decrease of albumin was indeed an independent predictive factor. In our GBM model, albumin contributed the most to the predictive model, and TG ranked the fourth.

Zhang et al. (Peng et al., 2017) reported a significant correlation between the decrease of serum Ca²⁺ and the incidence of POF by triggering the SIRS process that recruits neutrophils and leads to further release of reactive oxygen species and organ damage. Another study from Gravito-Soares et al. (2018) proposed that RDW, a marker reflecting inflammation status, showed great predictive performance of AP severity with AUC of >0.810 and mortality with AUC of >0.842. Additionally, this study further suggested that RCR was an excellent predictor of AP severity with AUC value of 0.973. Han et al. (2021) also discovered a positive correlation between a high level of RCR and a poor prognosis for patients with AP. Consistent with the aforesaid studies, our study illustrated that Ca²⁺, RCR, neutrophil count, and SIRS were among the top 10 important variables in the GBM model.

A new scoring system for predicting organ failure in AP was proposed by Wu et al. (2020) in 2020, consisting of LDH, creatinine, albumin and Ca²⁺. LDH, also included in the typical RANSON score, is not commonly seen in recent proposed models. In 2008, Gurda-Duda et al. (2008) recommended LDH activity (within 12 h from disease onset) as a biomarker for early predicting prognosis of AP, with sensitivity of 63.6% and a specificity of 89.6%. It is well known that hypertriglyceridemia and hyperglycemia are correlated with severity of AP and organ failure (Gurda-Duda et al., 2008; Park et al., 2020; Chen et al., 2021). Gurda-Duda et al. (2008) suggested that blood glucose concentration (within 36 h from disease onset) could be a complementary measurement, with sensitivity of 72.7% and a specificity of 75.8%. Park et al. (2020) investigated the association between the TyG index (=ln [fasting TG (mg/dl) × fasting plasma glucose (mg/dl)]/2) and the severity of AP in 373 patients. The results showed that the TyG index not only accurately predicted SAP but also increased the predictive value of traditional models. The underlying mechanism might be explained by insulin resistance, which activated proinflammatory molecules accelerating the progression of SAP. PT, a parameter of coagulation state, was among the top 10 important features in our GBM model. However, a multivariate logistic regression analysis performed by Radenkovic et al. (2009) and three machine learning algorithms performed by Qiu et al. (2019) did not include these parameters into the final models.

Here, we built six predictive models using traditional logistic regression and AutoML, with high AUC of >0.860 and high accuracy of >0.896. Furthermore, it is more convenient and efficient to get the predictive probability for SAP using AutoML. Additionally, our study was a multicenter hospital-based research, which is a common way of efficiently evaluating a new technique and may provide a better foundation for the subsequent generalization of our models. However, there are some limitations in our study. Firstly, we divided AP patients into non-SAP and SAP instead of mild AP, moderate SAP, and SAP, which might decrease the sensitivity of our models. Secondly, our study is a retrospective study which might affect the performance of our models in a prospective clinical study. More prospective research needs to be conducted for external validation of our models. Thirdly, the online deployment website needs maintenance, and more data need to be inputted to improve the generalizability and performance of our models.

We developed a series of effective models for early prediction of SAP based on AutoML platform, and these models outperformed the existing scoring systems, which might offer insights into AutoML applications in future medical studies. Additionally, the GBM model demonstrated practicable performance in early prediction better than LR and existing scoring systems.

The data presented in this study are available on request from the corresponding authors.

This study was approved by the ethics committee of the first affiliated hospital of Soochow University. Written informed consent from the patients/ participants was not required to participate in this study in accordance with the national legislation and the institutional requirements.

MY and RZ contributed to data collection and writing. MY was responsible for statistical analysis. ZZ assisted in data collection and statistical analysis. LL, JG, and CY contributed to data cleaning and creating charts. JL and WX assisted in computer programming. XL and JZ contributed to revising this dissertation. JZ and CX managed this project and provided the funding. All authors listed have made a substantial, direct, and intellectual contribution to the work and approved it for publication.

Science and Technology Plan of Suzhou City (SKY2021038) sponsored by CX. Youth Program of Suzhou Health Committee (KJXW2019001) sponsored by JZ.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer SL declared a shared parent affiliation with the author(s) MY, RZ, ZZ, LL, JG, WX, CY, JL, XLL, CX, JZ to the handling editor at the time of review.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.