Acute pancreatitis (AP) is one of the most common gastrointestinal problems for hospital admission globally (Hong et al., 2020). While most patients with AP will recover within a week of a mild course and are often self-limiting, 20% of patients progress to severe disease with a historical mortality risk as high as 30% (Trikudanathan et al., 2019). In the absence of specific treatment in the early phase, initial management of severe acute pancreatitis (SAP) focuses on supportive care such as fluid resuscitation, pain control, and nutritional support, aimed to minimize the impact of systemic inflammatory response syndrome (Lee and Papachristou, 2019). Patients with SAP often need to be transferred to the intensive care unit once organ failure occurs. Therefore, it is important to recognize predictors for severe disease in the early phase of AP, to select those patients who would benefit most from enhanced surveillance or early interventions. Early case identification and classification of disease severity could improve the clinical outcomes (Hong et al., 2021).

Many clinical scoring systems have been developed for the prediction of disease severity, such as the Ranson, chronic health evaluation (APACHE-II) score, Pancreatitis Outcome Prediction (POP) Score (Harrison et al., 2007), and Bedside Index Of Severity In Acute Pancreatitis (BISAP) (Wu et al., 2008). However, the existing scoring systems have moderate accuracy in predicting the severity of AP (Mounzer et al., 2012). Recently, Langmead et al. reported that a 5-cytokine panel consisting of angiopoietin-2, hepatocyte growth factor, interleukin-8, resistin, and soluble tumor necrosis factor receptor 1A accurately predicts persistent organ failure early in the disease process and significantly outperforms the prognostic accuracy of existing laboratory tests and clinical scores (Langmead et al., 2021). However, the test of cytokine is not routinely available, resulting in limited use in clinical practice. Several laboratory indexes such as total cholesterol (Hong et al., 2020), low-density lipoprotein cholesterol (Hong et al., 2018), albumin (Hong et al., 2017a), and blood urea nitrogen (BUN) (Lin et al., 2017) have been proposed as single predictors of severity of AP. Recently, Takeda et al. reported that fluid sequestration is a useful parameter in the early identification of SAP (Takeda et al., 2019). Yan et al. described that pleural effusion volume could be a reliable radiologic biomarker in the prediction of severity and clinical outcomes of AP (Yan et al., 2021). Gibor et al. reported that circulating cell-free DNA in patients with acute biliary pancreatitis is associated with disease markers and prolonged hospitalization time (Gibor et al., 2020). However, these single prediction markers are easy to use in practice but lack high accuracy.

Recently, artificial intelligence methods are also being used in the clinical setting for disease prediction or aiding in making decisions. Among several methods, a random forest (RF) is a group of many decision trees, each of which is characterized by a tree-like structure (Genuer and Poggi, 2020). It will randomly choose features and make observations, build a forest of decision trees, and then average out the results (James et al., 2013). RF allows considering qualitative and quantitative explanatory variables together, without preprocessing (Genuer and Poggi, 2020). Random forests are adapted to both supervised classification problems and regression problems (Genuer and Poggi, 2020). In addition, RF can handle datasets with many predictor variables, while also performing very well. Additionally, it can obtain variable importance ranking when used for prediction modeling (Speiser et al., 2019). RF, as a traditional machine learning method, has been shown to outperform other techniques for sets of features in a variety of different settings. RF has recently demonstrated high performance in risk classification and disease prediction (Yu et al., 2021). Lo et al. developed RF model for forecasting allergenic pollen in North America (Lo et al., 2021). Lin et al. reported that using the RF model could predict environmental risk factors in relation to health outcomes among school children from Romania (Lin et al., 2021b). Roguet et al. reported that RF classification with 16S rRNA gene amplicons offers an accurate solution for identifying host microbial signatures (useful in detecting human and animal fecal contamination in environmental samples) (Roguet et al., 2018). Yang et al. provided an RF prediction model for 3-year risk assessment of cardiovascular disease (Yang et al., 2020).

However, to the best of our knowledge, the use of RF model in predicting disease severity in patients with AP has not been performed yet. The aim of this study was to develop an RF model and compare it with a traditional logistic regression (LR) model for prediction of SAP.

Albumin is one of the most important proteins in plasma and plays a role in maintaining osmotic pressure, antioxidants, and scavenging free radicals (Viasus et al., 2013). Albumin has also long been considered a negative acute-phase protein, with reduced production in inflammation, paving the way for inflammatory cytokines (Charlie-Silva et al., 2019). Serum albumin levels undoubtedly decrease in inflammatory states, which may result in shorter half-life and a larger interstitial pool (Barle et al., 2006) as well as capillary leak (Soeters et al., 2019), during the inflammatory process. Excessive oxidative stress is associated with damage to acinar cells which has been observed in cerulein-induced mouse models of AP (Shen et al., 2018). In addition, clinical evidence suggests that oxidative stress is common in the early phase of AP (Hackert and Werner, 2011). Therefore, it was suggested that decreased albumin may reduce the ability to counterwork oxidative stress-induced acinar damage by binding reactive oxygen species in AP (Xu et al., 2020; Belinskaia et al., 2021). Xu et al. has reported that albumin is an independent predictor for SAP and in-hospital mortality in AP patients (Xu et al., 2020). Our previous study also indicated that hypoalbuminemia within 24 h of admission is independently associated with the development of persistent organ failure and mortality in AP (Hong et al., 2017a). Ocskay et al. (Ocskay et al., 2021) reported that the incidence of hypoalbuminemia was 35.7% during hospitalization and it was dose-dependent, associated with severity and mortality in AP. In our current study, the LR analysis indicated that albumin (p<0.001) is an independent predictor of SAP ( Figure 1). Based on the RF model, albumin is also an important predictor of SAP, based on variable importance analysis ( Figure 2). These results are consistent with previous reports.

Creatinine is primarily generated by muscle mass and dietary intake. It is eliminated from the glomerular filtration membrane (Stevens et al., 2006; Earley et al., 2012) and serves as the most widely used functional biomarker of the kidney, which can reflect renal injury in AP (Earley et al., 2012). Apart from renal injury, it also has been reported that the level of serum creatinine is associated with pancreatic necrosis (Muddana et al., 2009; Papachristou et al., 2010; Lipinski et al., 2013). The possible explanation is that necrotic cells release a large number of toxic substances and pro-inflammatory factors to cause renal injury, manifesting the elevation of serum creatinine. Therefore, the rise of creatinine may be attributed to the renal injury and pancreatic necrosis along with SAP. Wilkman et al. reported that increased creatinine levels are independently associated with 90-day mortality in AP patients (Wilkman et al., 2013). Wan et al. suggested that serum creatinine levels within 24 h of admission are effective for predicting persistent organ failure in AP (Wan et al., 2019). In addition, several scoring systems, which take creatinine as an index, are widely used in the clinical settings, including the Acute Physiology and Chronic Health Evaluation (APACHE) II, Sequential Organ Failure Assessment (SOFA) score for predicting the severity of pancreatitis, and modified Marshall scoring system for assessing organ dysfunction occurrence in SAP (Mederos et al., 2021). Our study indicated serum creatinine could be a useful predictor in both the RF and the LR model for predicting SAP ( Figures 1, 2).

The lipoprotein profile, especially HDL, is markedly decreased in inflammation and the accompanying acute-phase (Jahangiri, 2010). The mechanisms causing low serum HDL and LDL levels in the acute phase of AP remain largely unknown (Hong et al., 2017b; Hong et al., 2018). Jahangiri et al. (Jahangiri, 2010) suggested that it was related to a decreased rate of lipoprotein synthesis in the liver, general catabolism, and activation of the inflammatory system in the acute phase of the disease. Another explanation for the low serum HDL levels is that it may be due to increased expression of the Toll-like receptors (TLRs), especially TLR-4 expression (Zhang et al., 2010). It was reported that stimulated TLR-4 expression suppresses HDL levels (Liao et al., 1999). Khan et al. found that serum lipid concentrations such as HDL cholesterol and LDL cholesterol were associated with patients of SAP in all etiologies (Khan et al., 2013). However, Bugdaci et al. (2011) found a significant association between decreased HDL level and severity of the disease only in alcoholic and hypotriglyceridemic pancreatitis. In hypertriglyceridemic status, it is demonstrated that free fatty acids (FFAs) damage acinar cells and cause pancreatitis attack due to premature activation of trypsinogen, by creating an acidic environment (Okura et al., 2004; Guo et al., 2019). It has been reported that HDL takes part in FFA clearance (Asztalos et al., 2007) so that decreased HDL in hypertriglyceridemic AP cases may lead to an increase in FFA, and further damage acinar cells. Therefore, it has been suggested that an increase in HDL may be helpful for recovery from the disease by contributing to antioxidants (Bugdaci et al., 2011) and anti-inflammatory effect (Murphy and Woollard, 2010). On the other hand, in comparison, few studies are available about the pathophysiological mechanism of decreased LDL in SAP. Our study indicated both HDL and LDL were useful predictors for SAP (Figure 2).

Pleural effusion occurs in 3–50% patients with AP, based on a previous study (Basran et al., 1987; Kumar et al., 2019; Peng et al., 2020). The effusion can be asymptomatic and often hemorrhagic, usually resolving as pancreatitis subsides (Basran et al., 1987). Several mechanisms of pleural effusion in pancreatitis have been proposed, such as the trans-diaphragmatic lymphatic blockage, the pancreatic pleural fistula caused by the rupture of the pancreatic duct, and the fluid exudation from the sub-pleural diaphragmatic vessels into the pleural cavity (Kumar et al., 2019). Pleural effusion is reported to be associated with a severe course for initial risk assessment severity in AP and a sign of SAP (Heller et al., 1997; Tenner et al., 2013; Lankisch et al., 2015). Yan et al. reported that pleural effusion volume quantified on chest CT was positively associated with the duration of hospitalization (Yan et al., 2021). As a prognostic factor, pleural effusion has been incorporated in SAP severity predictive systems such as the Bedside Index for Severity in Acute Pancreatitis (BISAP) score (Gao et al., 2015), the Panc 3 score (Brown et al., 2007), and the Extra Pancreatic Inflammation on CT (EPIC) score (De Waele et al., 2007). Following the above outcomes, the present study suggested that pleural effusion (OR 5.11, 95%CI 2.38-10.94) was an independent risk factor for SAP (Figure 1).

The mechanism of BUN elevation in AP is thought to be based on the loss of intravascular volume, caused by interstitial extravasations owing to the systemic inflammatory response syndrome and an AP promoted direct renal injury mechanism. It has been reported that BUN, as a single predictor, had moderate accuracy in predicting persistent organ failure in AP (Mounzer et al., 2012). Koutroumpakis et al. reported that the rise in BUN at 24 h was the most accurate in predicting persistent organ failure and pancreatic necrosis (Koutroumpakis et al., 2015). Li et al. suggested that BUN was an independent risk factor to predict in-hospital mortality (Li et al., 2020). Valverde-Lopez et al. indicated that BUN was the best predictor of SAP after 48 h (Valverde-Lopez et al., 2017). BUN is also included in many scoring systems for AP, such as BISAP, JSS, and Glasgow score (Mounzer et al., 2012). Consistent with these reports, our study shows that BUN is the most important predictor of the RF model based on variable important analysis (Figure 2).

Decreased levels of serum calcium are commonly seen in critical illness, and hypocalcemia is significantly more frequent in patients with SAP (Peng et al., 2017). The mechanisms of hypocalcemia in SAP may be multi-factorial, such as abnormalities of parathyroid hormone secretion and action as well as vitamin D deficiency, binding of calcium in areas of fat necrosis, likely to contribute to the medication side effects (Weir et al., 1975; Steele et al., 2013). Serum calcium levels are closely related to the severity of the disease and its complications in AP. It has been incorporated in several clinical scoring systems as such as Pancreatitis Outcome Prediction (POP) Score, and Simple Prognostic Score (Harrison et al., 2007; Gonzálvez-Gasch et al., 2009). Mentula et al. suggested that serum calcium was the best single marker in predicting organ failure in AP after 24 h of symptom onset (Mentula et al., 2005). He et al. indicated that serum calcium was one of the independent predictors of the severity of AP in elderly patients (He et al., 2021). Serum calcium was also considered a significant factor in predicting early death in SAP (Shinzeki et al., 2008). As expected, our study indicated that calcium could be a useful predictor of SAP in the RF model (Figure 2).

Clinical evidence shows hyperglycemia is the common early feature of AP and abnormal glucose metabolism is present in almost 40% of AP patients (Banks et al., 2013; Chen et al., 2021). According to the traditional view, the mechanism is that the damage of organisms caused by AP activate the neuroendocrine system and lead to the secretion of many stress hormones (Binker and Cosen-Binker, 2014; Sun et al., 2019; Lu et al., 2021). Meanwhile, it is also related to the damage of the endocrine pancreas caused by SAP attacks. The association between hyperglycemia and adverse clinical outcomes in critically ill patients has been demonstrated in several observational studies, which suggest that high levels of glucose during the progression of AP can promote the release of inflammatory cytokines. These, in turn, influence disease progression (Sun et al., 2019; Chen et al., 2021). Sun et al. has suggested that the level of glucose in serum is positively correlated with the APACHE II scores, TNF-α, and CRP in AP (Sun et al., 2019). However, transient stress hyperglycemia in critically ill patients is considered harmless in some studies, indicating that the body has normal immune regulation ability (Lu et al., 2021), the subsequent derangement of glucose homeostasis could cause damage to the body (Pendharkar et al., 2016). Blood glucose-related indicators are associated with in-hospital mortality in critically ill patients with AP (Lu et al., 2021). Our LR model also shows that glucose is a useful predictor of SAP (Figure 1).

Machine learning has been extensively used for the prediction of severity or complication of AP (Zhou et al., 2022). Thapa et al. has reported that an XGBoost model could predict which patients would require treatment for SAP (Thapa et al., 2022). Early prediction of SAP using machine learning has also been attempted (Thapa et al., 2022). Jin et al. reported that the multilayer perception-artificial neural network (MPL-ANN) model based on routine blood and serum biochemical indexes could reliably predict disease severity in patients with AP (Jin et al., 2021). Choi et al. combined clinical (i.e., APACHE-II and BISAP scores) and radiologic (i.e., Balthazar grade and EPIC score) scoring systems by classification tree analysis for predicting SAP (Choi et al., 2018). Xu et al. reported that adaptive boosting algorithm (AdaBoost) could predict development of multiple organ failure, complicated by moderately severe or severe AP (Xu et al., 2021). However, the above models were limited due to lack of individualized prediction on the test sample. Implementation on such data remains challenging because of the low interpretability of results of machine learning (Yu et al., 2021). Our study indicated that, compared to the LR model and BISAP score, RF exhibited the highest discriminatory performance for the prediction of SAP on both training and test samples (Figures 3, 4, 6, 7). Using the RF model, we could illustrate key features and establish a prediction model, with high accuracy in patients with SAP. The LIME plot could provide a visual illustration of the individualized interpretation of the importance of different features, which might help clinical doctors to understand results of the RF model (Figure 8). The LR model (nomogram) achieved a sensitivity of 93.8%, acceptable specificity of 79.3%, and diagnostic accuracy of 80.8% (Table 3). Though the diagnostic performance of the LR model (nomogram) is inferior to the RF model, it is simple and intuitive to calculate the prediction probability of a result, which makes it valuable in predicting SAP (Figure 1).

To the best of our knowledge, this is the first study to develop an interpretable RF model for SAP prediction. The strength of this study is a large sample size, which enables a strong statistical power. Both patients in ICU and the general ward were enrolled in this study, thus reducing selection bias. However, our study has some limitations, even if it has been internally validated by tenfold cross-validation technique and test set, testing the performance of our RF model in an external/other independent data set is necessary. In addition, even if effective, RF models are sophisticated and difficult to understand, and thus, comparable to a ‘black box’. We have, therefore, demonstrated that by utilizing Lime plots, the results could be more easily interpreted (Al’Aref et al., 2020). At last, we did not evaluate the RF model and single predictors for other clinical outcomes such as patient survival and organ failure occurrence, intensive care unit (ICU) admission, and SAP recurrent rate. It would be interesting to carry out a large-sample prospective study to determine whether our model and other variables such as serum creatinine, albumin, BUN, HDL, LDL, calcium, and glucose play a significant role in predicting these clinical outcomes.

In conclusion, an interpretable RF model exhibited the highest discriminatory performance to predict SAP. Interpretation with LIME plots could be useful for individualized prediction in the clinical setting. A nomogram consisting of albumin, serum creatinine, glucose, and pleural effusion is also useful for the prediction of SAP.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Ethics Committee of the First Affiliated Hospital of Wenzhou Medical University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

WH conceived the study and carried out the majority of the work. WH participated in data collection and conducted data analysis. WH, YL, XZ, SJ, JP, QL, and SY drafted the manuscript. ZB, MZ, and HG helped to finalize the manuscript. All the authors read and approved the manuscript.

This work was supported by Zhejiang Medical and Health Science and Technology Plan Project (Number: 2022KY886), Wenzhou Science and Technology Bureau (Number: Y2020010).

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.

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.