The Ranson Criteria, first proposed by John H.C. Ranson and his team in 1974, was the first scale developed to classify the severity of acute pancreatitis [7]. This scoring system was developed based on clinical data from 100 patients and comprises 11 indicators, which are assessed at two time points: “on admission” and “48 hours after admission”. Five indicators are assessed at admission: age > 55 years, white blood cell count > 16,000/µL, blood glucose > 11.1 mmol/L, aspartate aminotransferase (AST) > 250 IU/L, and lactate dehydrogenase (LDH) > 350 IU/L. The assessment at 48 hours after admission included the following six criteria: a decrease in hematocrit > 10%, an increase in blood urea nitrogen (BUN) > 1.8 mmol/L, serum calcium < 2.0 mmol/L, arterial oxygen partial pressure (PaO₂) < 60 mmHg, a base deficit > 4 mmol/L, and an estimated fluid accumulation > 6 L. The risk of death increases as the Ranson score rises: the mortality rate is less than 3% for scores of 0 - 2, 10% - 20% for scores of 3 - 4, and as high as 40% or more for scores of 5 or higher [7][8].

Although the Ranson score is widely used, it has significant shortcomings. On one hand, it requires 48 hours of data, making early risk assessment difficult; on the other hand, it was developed for people with alcoholic pancreatitis, so it can’t be directly applied to patients with other causes [9]. Thirdly, the scores need to be collected in two separate periods for lab results, which is a complicated process and prone to missing data; its AUC for predicting severe acute pancreatitis within 48 hours is 0.81 - 0.95 [10]-[13].

Clement W. Imrie first proposed the Glasgow Scale in 1970, also known as the Imrie criteria. By 1984, Blamey et al. refined this framework [14]. The system includes eight factors: age > 55 years, white blood cell count > 15,000/uL, blood glucose > 10 mmol/L, serum urea > 16 mmol/L, arterial PO₂ < 60 mmHg, serum calcium < 2.0 mmol/L, serum albumin < 32 g/L, and LDH > 600 IU/L. Like Ranson, the Glasgow score requires 48 hours for full assessment. Its AUC for predicting SAP at 48 hours ranges from 0.80 to 0.85 15. Its advantage is a relatively concise indicator set and more extensive validation in European populations. However, its 48-hour window limits early utility; when only admission data are used, sensitivity drops to 68% - 71% [15].

The APACHE II score, introduced by Knaus et al. in 1985 for general ICU severity assessment [16], was adapted for AP by Larvin and McMahon in 1989 [17]. It includes 12 acute physiologic variables, age, and chronic health status, with scores from 0 to 71. Unlike Ranson and Glasgow, APACHE II can be assessed at any time, including at admission. A cutoff of 8 is typically used for predicting SAP. During the first 24 hours, sensitivity ranges from 63% to 74% and specificity from 73% to 81% [13][17]. Serial APACHE II monitoring can track disease progression and treatment response.

APACHE II has several practical limitations. Its calculation requires 12 physiologic parameters, many not readily available in the emergency setting. It was not designed specifically for AP, and some parameters have low relevance to pancreatitis. The calculation is cumbersome for routine clinical use. Mounzer et al. compared nine scoring systems and found APACHE II had an AUC of 0.78 for predicting persistent organ failure using 24-hour data, no better than simpler tools such as BISAP [15].

The Bedside Index for Severity in Acute Pancreatitis (BISAP) was developed by Wu et al. in 2008 using classification and regression tree (CART) analysis of approximately 18,000 patients from 212 hospitals [18]. The BISAP score incorporates five parameters assessed within the first 24 hours: Blood urea nitrogen > 25 mg/dL, Impaired mental status (Glasgow Coma Scale < 15), Systemic inflammatory response syndrome (SIRS), Age > 60 years, and Pleural effusion on imaging. Each parameter is assigned one point, yielding a total score from 0 to 5. A BISAP score of 3 or higher is associated with significantly increased mortality risk [18][19].

The BISAP score has gained considerable popularity due to its simplicity, use of readily available variables, and ability to be calculated entirely at the bedside within the first 24 hours. In the initial validation cohort, the BISAP score demonstrated an AUC of 0.82 for predicting mortality, comparable to APACHE II but with substantially greater ease of use [18][19]. Subsequent studies have confirmed these findings, with AUC values for predicting SAP ranging from 0.77 to 0.85 at 24 hours [13][18]. However, BISAP performs poorly in predicting persistent organ failure: at a cutoff of 2, sensitivity drops to 38% [15]. Additionally, the system assigns disproportionate weight to age, reducing its effectiveness in elderly patients.

Papachristou et al. compared scoring systems in 185 AP patients and found the Ranson score had the best predictive ability (AUC 0.94), outperforming BISAP (0.81) and APACHE II (0.78). However, Ranson requires 48 hours, limiting early clinical application [13]. The predictive performance of traditional scores varies across populations and settings, with limited external generalizability [20]. As a result, a lot of research has started focusing on cutting-edge machine learning models, new biomarkers and advanced imaging technologies.

C-reactive protein (CRP) is the most widely used single biomarker for predicting progression to SAP. A meta-analysis reported an AUC of 0.85, sensitivity of 0.76, and specificity of 0.79 at 48 hours after onset [21].CRP ≥ 150 mg/L is generally used to predict severe acute pancreatitis, but CRP only peaks 48 - 72 hours after inflammation starts, which isn’t great for early risk assessment. CRP is a non-specific acute phase reactant and can be affected by infections and other inflammatory diseases, which limits its accuracy in predicting severe acute pancreatitis [21][22].

Procalcitonin (PCT) is a popular alternative marker to CRP; it can rise within 6 to 12 hours of inflammation, signalling systemic inflammation earlier. A meta-analysis of 18 studies (1,764 patients) reported a pooled sensitivity of 0.80, specificity of 0.84, and AUC of 0.89 for SAP prediction within 24 hours of admission [23]. Many studies show that PCT predicts SAP is better than CRP, but this indicator is costly to test and not widely available, so it can’t be used routinely on a large scale. [23][24].

Blood urea nitrogen (BUN) has gained significant attention in recent years due to its simplicity, low cost, and wide availability. In a study by Wu et al. that included 3 prospective cohorts totaling 1,043 patients with AP, a BUN level of ≥20 mg/dL at admission was found to be significantly associated with an increased risk of death (OR = 4.6) [25]. Also, consistently high BUN levels within 24 hours of hospital admission are an independent risk factor for increased mortality. Monitoring BUN dynamically can help assess disease progression and prognosis. The BISAP score includes BUN ≥ 25 mg/dL as a key evaluation criterion, further showing its role in early risk stratification of acute pancreatitis [18].

Interleukin-6 (IL-6) is a key pro-inflammatory cytokine that induces CRP synthesis. Its levels rise rapidly in early AP, making it useful for early risk prediction. A meta-analysis of 13 studies reported an AUC of 0.85 within 24 hours of onset and a diagnostic odds ratio of 14 (95% CI: 7 - 27) [26]. Despite its promise, clinical application is limited by a lack of standardized testing methods, high costs, and the absence of a uniform cutoff value 26,27. Other interleukins such as IL-8 and IL-10 are also involved in the inflammatory response but show inferior predictive ability compared to IL-6 [27].

Immune cell ratio indices derived from complete blood counts have been applied to early AP risk assessment due to their simplicity and low cost. The neutrophil-to-lymphocyte ratio (NLR) reflects both inflammatory response and immune status. A meta-analysis by Kong et al. reported a pooled sensitivity of 0.79, specificity of 0.71, and AUC of 0.82 for SAP prediction [28]. The platelet-to-lymphocyte ratio (PLR) also has predictive value; combining NLR and PLR can further improve performance [29]. However, these markers lack sufficient accuracy to be used alone for clinical decision-making.

Trypsinogen activation peptide (TAP) directly reflects the premature activation of trypsinogen within the pancreas, and premature trypsinogen activation is one of the key pathophysiological mechanisms underlying the onset of AP.A multicenter study involving 246 patients showed that urinary TAP had a sensitivity of 58% and 83% for predicting severe acute pancreatitis at 24 and 48 hours after onset, respectively, with a specificity of approximately 72% to 73% [30]. However, due to the complexity of the testing process, the susceptibility of TAP in urine to degradation, and the fact that commercially available test kits are not yet fully developed, its clinical application remains somewhat limited.

Contrast-enhanced computed tomography (CECT) remains the primary imaging modality for morphologic assessment in AP. In 1985, Balthazar et al. described a CT grading system based on pancreatic morphology [31], later combined with a necrosis scoring component in 1990 to form the CT Severity Index (CTSI) [32]. The CTSI uses a 10-point system: the Balthazar index (grades A - E, 0 - 4 points) plus a necrosis score (0 points for no necrosis, 2 for <30%, 4 for 30% - 50%, 6 for >50%). Assessed after 72 hours, a score greater than or equal to 7 is associated with high risk of purulent necrosis, prolonged hospitalization, and increased mortality [33].

The Modified CT Severity Index (MCTSI), introduced by Mortele et al. in 2004, simplified the original CTSI by using a 10-point scale that incorporates inflammation (0 - 4 points), necrosis (0 - 4 points), and extrapancreatic complications (2 points) [34]. The MCTSI demonstrated improved correlation with patient outcomes, including length of hospital stay, need for intervention, and organ failure. In a comparative analysis by Bollen et al., various CT scoring systems, including the CTSI and MCTSI, demonstrated comparable predictive accuracy, with no statistically significant differences among them [35].

CT-based scoring systems have fundamental limitations for early prediction. CECT is not recommended within the first 48 - 72 hours of admission because early imaging may underestimate necrosis and expose patients to contrast during a period of potential renal compromise [6][36]. CT scores cannot reliably predict early mortality or systemic organ failure before necrosis develops. The 2026 iLATAM-AP guidelines recommend optimal CECT timing for severity assessment at least 72 hours after pain onset [37]. MRI is an alternative but is limited by prolonged scanning times, high cost, and limited availability [5].

Traditional scoring tools and individual biomarkers have inherent analytical limitations. Machine learning (ML) and artificial intelligence (AI) can capture complex, non-linear interactions across multiple data types. Unlike standard regression, which requires manual specification of cross-variable relationships, ML algorithms learn data hierarchies autonomously and can identify predictive patterns not apparent through conventional analysis [38][39].

Extreme Gradient Boosting (XGBoost) performs steadily in SAP prognosis assessment. Thapa and others built an XGBoost model based on 61,894 electronic medical records with 24-hour clinical data, achieving an AUC of 0.921, outperforming HAPS and BISAP scores [40]. Luo and others compared five machine learning models, and the results showed that the random forest (RF) performed best at admission, with internal and external validation AUCs of 0.969 and 0.961, clearly better than the Ranson score (0.847) and BISAP score (0.796) [41].

Hong et al. applied a random forest model with 16 variables to 648 AP cases, achieving a cross-validated AUC of 0.89 and a test-phase AUC of 0.96 [42]. Feature importance analysis picked out blood sugar, blood calcium, HDL, LDL, albumin, creatinine, and BUN as key predictors; researchers used the LIME method to visually explain individual risks [42].

Litvin et al. developed an artificial neural network (ANN) that outperformed logistic regression for admission-day SAP prediction (AUC 0.92 vs. 0.84) [43]. The multicenter PANCREATIA study tested models using only emergency room vital signs and demographic data, without laboratory or imaging results, and achieved an AUC of 0.849 for mortality prediction [44], demonstrating that AI can extract clinically useful signals from minimal data.

Current research tends to integrate multiple data sources to build a unified prediction model. Yin and others developed the PrismSAP model framework, combining CT deep learning features, radiomics, laboratory results, and clinical indicators, achieving a 72-hour prediction AUC of 0.916 [45]. Li and colleagues used a voting ensemble to sort AP into three severity levels (MAP, MSAP, SAP), getting a C-statistic of 0.902 and using SHAP to make it easier to understand [46]. These findings suggest that integrating multidimensional clinical data enables more accurate risk stratification than any single data type alone.

Qian et al. synthesized 33 studies of ML models for SAP prediction and reported pooled C-indices of 0.88 (95% CI: 0.86 - 0.90) in external validation and 0.87 (95% CI: 0.84 - 0.89) internally [47]. However, due to differences in study populations, predictor variables, study endpoints and model frameworks, there is considerable heterogeneity in current research methodologies, so the results should be interpreted with caution. Further comparisons of machine learning models and traditional scores are needed within the same prospective cohort to better verify the clinical superiority of the former. Critelli and others summarised that machine learning models have an average AUC of 0.91, but most studies are single-centre retrospective analyses, lacking external validation and with limited generalizability [48]. The author emphasised that the relevant models need to be validated through prospective, multicentre studies.

Even if the AUC values are high in retrospective studies, they often can’t be directly applied to real clinical work. Most predictive models are built using data from a single centre, making it hard to reliably use them in other population cohorts [47]. Predictive models that are optimised for specific groups (like East Asians or Caucasians) usually work poorly when applied to other groups. This is because the genetic risk factors, disease causes, and medical resource conditions differ between populations. On top of that, variations in how hospitals record medical histories, imaging protocols, and testing methods can further make these models perform worse during external validation [48][49].

One big hurdle for applying machine learning models in clinics is their lack of interpretability. Traditional scoring systems have clear logic and the derivation process is easy to follow. In contrast, gradient boosting models and deep neural networks are ‘black box models’, making it hard to explain the prediction for each patient from a biological perspective [50]. These black box issues make clinicians responsible for diagnostic decisions a bit wary. Currently, the academic world has proposed quite a few methods, like SHAP, LIME and attention mechanisms, which can explain model predictions after the fact [40][42], but these methods have their own limitations, including instability and vulnerability to manipulation.

Most machine learning-based severe acute pancreatitis (SAP) prediction models haven’t been prospectively validated or tested in randomised controlled trials. Most of the existing evidence comes from retrospective analyses, which have inherent biases like selection bias, information bias, and confounding bias [48]. So far, no model has met the strict standards for clinical application: none have been proven to be practically useful in clinics, nor shown to improve patient outcomes, and there’s a lack of evidence on cost-effectiveness. Regulators still haven’t put forward a clear set of rules to assess AI clinical decision support tools [51].

A very common methodological flaw in machine learning research is data leakage: that is, using data when training a model that wouldn’t be available at the moment of prediction, such as later trends in test indicators, CT scan results 72 hours after onset, or treatment-related data [52]. This could lead to inflated performance metrics in the model reports. Any further research must ensure that the variables included in the model are limited to data available at the initial triage assessment [53].

Prospective studies should clearly define their clinical targets. Models must distinguish among persistent organ failure, overall SAP, local complications, and all-cause mortality, as optimal predictors and risk thresholds differ across these outcomes. Composite endpoints are insufficient for guiding targeted interventions [54].

Although most machine learning studies show ideal AUC results, just having good discriminatory ability is not enough to prove that the model can safely guide clinical treatment [55]. The model also needs to be well calibrated, meaning its predicted probabilities should match the actual event occurrence rates [56]. If the model calibration is poor, it can lead to unreasonable clinical interventions, unnecessary treatments, or missed necessary interventions. Therefore, decision curve analysis (DCA) is needed to verify the model’s actual clinical value at different risk threshold levels [57]. The best cutoff value can’t be generalised. When screening in general wards for SAP or ruling out severe cases, try not to miss any diagnoses (high sensitivity); when deciding whether a patient needs to be transferred to the ICU, try not to get it wrong (high specificity). Before the model is officially used in clinics, it must pass both calibration and DCA clinical net benefit validation [58].

It’s still really tricky to use machine learning models at the bedside. A lot of hospital IT systems don’t support real-time calculations; in the busy pace of the emergency department, extra manual data entry is hard to make happen [59][60]. There’s still no clear rule on who’s responsible when AI-assisted diagnosis goes wrong. If the model underestimates the risk for high-risk patients and their condition gets worse, it’s still unclear what kind of legal responsibility everyone should bear [61]. For a model to really take off in clinics, it needs to tick three boxes: smoothly connect with existing hospital electronic medical records, have a simple and user-friendly interface, and clearly define the division of labour and decision-making rules between doctors and AI.

There are currently several cutting-edge directions to further optimise SAP prediction models. By combining bedside indicators, imaging, and multi-omics data, we can find new prognostic markers and explain the molecular mechanisms behind necrosis progression and organ failure. Federated learning can enable multi-centre joint modelling without compromising patient privacy, while also improving model generalisability. Natural language processing can extract useful information from medical records, enriching the model’s input features. Finally, the model needs to evolve from a single static assessment to a dynamic risk prediction system that can be updated in real time.