Intraductal papillary mucinous neoplasms (IPMNs) are precursor lesions of pancreatic adenocarcinoma (1). Classification of low grade dysplasia (LGD), high grade dysplasia (HGD), and invasive intraductal papillary mucinous carcinoma (I-IPMC) is based on the degree of cyto-architectural dysplasia in accordance with the World Health Organization (WHO) classification system published in 2019 (2). Several reports have shown that I-IPMC has characteristics similar to pancreatic ductal adenocarcinoma (PDAC), including potential lymph node (LN) or distant metastasis, postoperative recurrence, and poor survival (3–9). Depending on the grade of IPMN during the progression from noninvasive IPMN to invasive IPMC, the choice of treatment varies from a conservative approach to radical pancreatectomy with LN dissection (10). Therefore, it is important to determine the grade of IPMN accurately to optimize the therapy and follow-up strategy. Preoperative distinction of IPMN grade is not easy, even when multiple modalities are used (10–17). Several guidelines propose radiological and clinical criteria to assess the risk of HGD and cancer in patients affected by IPMN (18–21), such as the International Consensus Guidelines (ICG) published in 2012, in which predictors of malignancy were categorized as high risk stigmata and worrisome features (18). However, the predicting model still does not identify the degree of malignancy in patients with IPMN.

Although HGD is not malignant, HGD-IPMN has been shown to exhibit an increased risk of subsequent development of pancreatic cancer in a few studies. Therefore, HGD and I-IPMC were studied together in various guidelines (9, 18–24). However, the median overall survival for patients who underwent pancreatic resection for IPMN with HGD was similar to the survival of patients with LGD, and superior to that of patients with IPMN-associated PDAC (22). Thus, early recognition of HGD is important to improve prognosis of the disease.

The recent consensus is that the malignant potential of IPMN is dependent on the presence and extent of invasive cancer (24–30). However, I-IPMC categorized by the Armed Forces Institute of Pathology (AFIP) and WHO classifications covers cancer with variable biological behaviors, with the 5-year survival rate varying substantially between 36% and 90% (2, 24–30). This may be due to the heterogeneity of I-IPMC, including biological behavior and an invasive component of various sizes. Invasion of ≤ 0.5 cm presents with excellent prognosis whereas the prognosis for IPMN at an advanced stage of invasion is as poor as that of conventional PDAC (2). In addition, no LN metastasis was observed in patients with early stage invasion ≤ 5 mm (31, 32). Thus, it would be more appropriate to stage invasive carcinomas with conventional staging protocols according to WHO (2019) and then further substage the pT1 category (early invasion ≤ 2 cm) into pT1a for those with an invasive component ≤ 5 mm (2). This is important for the early detection of pancreatic cancer.

HGD and pT1a are considered as the early stage of malignant IPMN. The ability to distinguish HGD or pT1a IPMN from LGD IPMN will help physicians treat IPMN patients in which it is difficult to identify by clinical and radiological characteristics, allowing for clear surgical indications and the ability to cure pancreatic cancer at an embryonic stage. Many reports have attempted to identify predictive factors for HGD and I-IPMC that might influence the management of IPMN patients by logistic regression analysis and nomogram (10, 16, 17, 31, 32). However, there have been no reports of a subdivided model to predict HGD/pT1a in patients with IPMN that could be widely applied in clinical practice. Therefore, it is necessary to find a simple and objective method to discriminate LGD or HGD/pT1a IPMN and advanced IPMC accurately based on clinical and radiological data.

Decision tree classification algorithms are an inductive learning method that can produce a tree-type classification model from given training samples. Each non-leaf node in the tree shows which feature is used to judge the category, and each leaf node represents the final category. The root node to each leaf node forms a path for classification.

Model interpretation is often essential to illustrate a machine learning model, especially in medical research. The factors which have strong significance in a classification or regression model can help medical workers attach importance to certain factors. An intuitive interpretation of a model can be easily understood. Scott ML et al. developed the method of calculating the SHAP value for interpretation of these models (33). The significance is illustrated by the absolute SHAP value; larger values have stronger significance.

Ensemble learning is a common method in machine learning. While learning, multiple models or classifiers are produced independently and ensembled into a final prediction model. The ensemble model usually performs better than any independent prediction model. The random forest algorithm is a combination of bagging and decision tree. Each tree is produced by part of the training sample and features. The final random forest model ensembles the decision trees by voting. Therefore, the random forest model has a higher accuracy.

The current consensus regarding the malignant potential of IPMN is that its aggressiveness is dependent on the presence of invasive cancer, the extent of cancer invasion, and the biological characteristics of cancer cells (24–30). These findings suggest that the therapeutic strategy for IPMN should differ according to the grade. Therefore, it is becoming more important to accurately categorize IPMN preoperatively into its corresponding pathological subtypes. Previous studies have focused on I-IPMC, but did not focus on the early stage of malignant IPMN. However, in the clinic, such patients are the most difficult to identify in time, and if identified earlier, the prognosis of patients will be greatly improved. In our study, HGD IPMN shows similar optimistic postoperative outcome to pT1a IPMC, whereas I-IPMC beyond pT1a has a poorer outcome compared with HGD/pT1a IPMN. The postoperative survival and 5-year survival rate of IPMN were nearly consistent with previous reports (22, 45). Moreover, there were no significant differences between most variables between HGD IPMN and pT1a IPMC in the present study. Therefore, we believe that pT1a and HGD behave similarly and both have excellent prognoses; thus, we integrated both pT1a and HGD into HGD/pT1a IPMN in the present study. However, performing grade classification preoperatively is generally difficult, and preoperative discrimination between IPMN with HGD/pT1a and LGD is nearly impossible. Several predictive factors for malignant IPMN have been reported, including dilated branch (30 mm), MPD dilatation, the presence of mural nodules, elevated CEA, and elevated CA19-9. These factors, however, were not able to differentiate HGD/pT1a from LGD lesions. Similarly, Kang et al. found that adding diffusion-weighted images to MRCP improved the ability to detect invasive malignancy in IPMN (46). In a recent study, the authors found that the use of contrast-enhanced harmonic EUS (CH-EUS) provided an increased diagnostic yield in the identification of malignant features of mural nodules within pancreatic cystic neoplasms (PCNs) (47). However, defining the degree of dysplasia in IPMN was not possible with imaging and classical tumor markers or EUS characteristics alone. In this study, we focused on establishing a pioneering system that combines the features into a patient-specific risk using unique clinical and standardized preoperative examination modalities and standardized pathological diagnostic criteria by tree models to perform accurate grade classification preoperatively.

The present study found 16 variables to be independently associated with the progression from LGD to HGD and pT1a to invasive carcinoma ( Table 2 ). Among them, accompanied by an unclear focus boundary, DM (especially new-onset DM), elevated serum CA19-9, and obstructive jaundice are obvious in the advanced stage of invasive IPMC, but are not common in HGD or pT1a IPMC. Many previous studies and guidelines have suggested that elevated serum CA19-9, DM, and obstructive jaundice are risk predictors of malignant IPMN (18, 20, 28, 45, 48, 49). However, HGD and pT1a IPMN have not been previously analyzed separately from malignant IPMN. Thus, in the present study, we found the three factors could not be used as a sign of early malignancy, and their appearance often indicates that the disease is at an advanced stage of invasion. Moreover, our study suggests the presence of mural nodules should be regarded as highly suspicious of invasive carcinoma or HGD in all morphological types, which is consistent with a malignant predictor in several reports (32, 43, 48, 50, 51). However, we also found that nearly 12–23% of cases of IPMN in the HGD and I-IPMC groups had no mural nodules; thus, other factors need to be considered. CBD diameter ≥ 15 mm and MPD dilatation are two specific signs of HGD and invasive disease (LGD vs. HGD vs. pT1a vs. I-IPMC beyond pT1a, 1 [0.7%] vs. 1 [7.7%] vs. 9 [37.5%] vs. 17[50.0%], p = 0.000; 0.4 (2.9, 0.1) vs. 1.1 (4.3, 0.4) vs. 1.3 (3.9, 0.5) vs. 0.84 (3.0, 0.2), p = 0.000). Although MT-IPMN was also associated with malignant disease (P = 0.000), the pathology of 35.3% of BD-IPMN was I-IPMC beyond pT1a. Assessing the malignant risk of BD-IPMN will require further investigation.

To observe the influence of high-risk factors, we used the interpretability method for tree models, which can help improve clinician understanding of IPMNs and important characteristics for diagnosis. As shown in Figure 4 , the nearest nodes to the root of the decision tree have the greatest influence on the classification. For a simple analysis, we can see the key to distinguishing benign from malignant IPMN is mural nodules. If the mural nodule size is less than 0.575 cm, LGD IPMN is more likely. If the mural nodule size is greater than 0.575 cm, and CA19-9 is greater than 26.85 U/ml, IPMC is more likely. If the mural nodule size is greater than 0.575 cm, but CA19-9 is less than 26.85 U/ml, no history of diabetes, and CEA is greater than 5.245 ng/ml, HGD/pT1a is more likely; if CEA is less than 5.245 ng/ml, LGD is more likely. From this, we can see the appearance of mural nodules( >0.575 cm) suggests early or advanced stage malignant IPMC and elevated serum CA19-9 reflects an advanced stage of invasive IPMC, which is similar to the recently published paper (52). In addition, the mural nodules of our decision tree algorithm was 5.75mm, and it can also be considered as a cutoff value, which was closed to previous studies (4, 43, 53),and the high accuracy of our model showed unique practical significance in our classification of LGD, HGD/pT1a and I-IPMC beyond pT1a. In the random forest model, which had a higher accuracy, the SHAP value was used for interpretation. As in Figure 5 , the mural nodule size, MPD diameter, serum CA19-9 levels, whether the lesion edge was clear, and the presence of CBD dilatation all influenced the random forest classification model. The decision tree provided an outstanding prediction capability for each stage for IPMN, with an accuracy of 86.9% on the testing set. The random forest model also showed high diagnostic ability with an accuracy of 96.0% on the testing set.

The random forest model uses the ensemble learning method, which assembles multiple decision trees, and as a result, is more accurate for this 3-class classification task. Comprehensive classification, including benign, early malignant, and advanced invasive, is convenient for clinical decisions. Additionally, we have made a convenient software for this random forest model for clinicians. By inputting relative captured features, the predicted class of IPMN of a single patient can be shown.

There are some limitations to the present study. First, this model was developed in a population of patients who were subsequently treated surgically. Hence, the cohort may not accurately represent patients with the whole IPMN population, and the cases could have selection bias. Second, prospective investigations conducted at multiple institutions are necessary for validating the results obtained in the current study. Finally, HGD and pT1a IPMN patients were relatively few. Therefore, internal validation (10-fold cross-validation) was used to evaluate the diagnostic performance for the invasion potency of IPMN because we could not collect enough patients to perform AI when patients were separated into groups for training and test data, and the results of the cross-validation have also shown high accuracy in classifying LGD, HGD/pT1a, and I-IPMC beyond pT1a, which better proves the efficiency of our AI models.

Overall, a simple and objective predictive model for IPMN of each stage that can be used for any lesion type was constructed. This predictive model provides important information for clinicians and patients in assessing an individual’s risk for early stage malignant IPMN and may help identify masses that appear benign but require surgery.

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 Ruijin hospital Ethics Committee Shanghai Jiaotong University School of Medicine. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

XH, DZ and YY had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design : XH, RF, DZ. Acquisition, analysis, or interpretation of data: XH, RF, JS, YR. Drafting of the manuscript: XH. Critical revision of the manuscript for important intellectual content: DZ and YY. Statistical analysis: XH. Technical, or material support: XZ, WR. All authors contributed to the article and approved the submitted version.