Machine learning model-based prediction of postpancreatectomy acute pancreatitis following pancreaticoduodenectomy: A retrospective cohort study

Abstract

BACKGROUND

The International Study Group of Pancreatic Surgery has established the definition and grading system for postpancreatectomy acute pancreatitis (PPAP). There are no established machine learning models for predicting PPAP following pancreaticoduodenectomy (PD).

AIM

To explore the predictive model of PPAP, and test its predictive efficacy to guide the clinical work.

METHODS

Clinical data from consecutive patients who underwent PD between 2016 and 2024 were retrospectively collected. An analysis of PPAP risk factors was performed, various machine learning algorithms [logistic regression, random forest, gradient boosting decision tree, extreme gradient boosting, light gradient boosting machine, and category boosting (CatBoost)] were utilized to develop predictive models. Recursive feature elimination was employed to select several variables to achieve the optimal machine algorithm.

RESULTS

The study included 381 patients, of whom 88 (23.09%) developed PPAP. PPAP patients exhibited a significantly higher incidence of postoperative pancreatic fistula (55.68% vs 14.68%, P < 0.001), grade C postoperative pancreatic fistula (9.09% vs 1.37%, P = 0.001). The CatBoost algorithm outperformed other algorithms with a mean area under the receiver operating characteristic curve of 0.859 [95% confidence interval (CI): 0.814-0.905] in the training cohort and 0.822 (95%CI: 0.717-0.927) in the testing cohort. According to shapley additive explanations analysis, pancreatic texture, main pancreatic duct diameter, body mass index, estimated blood loss, and surgery time were the most important variables based on recursive feature elimination. The CatBoost algorithm based on selected variables demonstrated superior performance, with an area under the receiver operating characteristic curve of 0.837 (95%CI: 0.788-0.886) in the training cohort and 0.812 (95%CI: 0.697-0.927) in the testing cohort.

CONCLUSION

We developed the first machine learning-based predictive model for PPAP following PD. This predictive model can assist surgeons in anticipating and managing this complication proactively.

Key Words: Pancreaticoduodenectomy; Postoperative complication; Acute pancreatitis; Machine learning; Predictive value

Core Tip: We demonstrate a novel machine learning algorithm for predicting postpancreatectomy acute pancreatitis, achieving excellent performance. The model’s coherent interpretation provides clinical utility in the management and treatment of patients following pancreaticoduodenectomy. This study explains the serious harm caused by such complications and establishes a possible effective preventive measure, providing assistance for clinical work.

INTRODUCTION

While post-endoscopic retrograde cholangiopancreatography pancreatitis is widely recognized with established research and guidelines[1]. Postpancreatectomy acute pancreatitis (PPAP) remained unrecognized until Connor[2] introduced it in 2016 as a distinct postoperative complication separate from postoperative pancreatic fistula (POPF). His criteria is different from the revised Atlanta classification of acute pancreatitis[3]. The serum amylase (AMY) levels exceeding the upper limit within 48 hours postoperatively is considered as PPAP according to Connor[2]. Varying reports on its incidence and prognosis across centers reflect the absence of clear diagnostic standards for PPAP. The International Study Group of Pancreatic Surgery (ISGPS) defined PPAP, established diagnostic criteria, and developed a grading system in 2022[4]. Since then, recognition of PPAP has increased, highlighting its adverse impact on other complications and perioperative recovery, prompting systematic research efforts[5-10]. PPAP is defined as inflammation of the remnant pancreas following partial pancreatic resection[2]. According to ISGPS, postoperative hyperamylasemia (POH), on the other hand, is characterized by elevated serum AMY levels persisting above the upper limit of normal (> 100 U/L) for 48 hours postoperatively[4]. Diagnosis of PPAP requires meeting the criteria mentioned above along with computed tomography (CT) evidence of pancreatic changes (exudates, fluid collection, edema, necrosis) or relevant clinical manifestations[4]. Like the definition of clinically relevant POPF[11], abnormal laboratory findings alone define POH (grade A), whereas PPAP diagnosis (grade B) additionally necessitates radiological changes or clinical impacts indicative of mild or moderate complications; grade C indicates severe, life-threatening complications[4].

Under the ISGPS standards, the incidence rates of POH and PPAP were approximately 50% and 25%, respectively[5-10]. Previous studies have focused on univariate and multivariate logistic analyses to identify independent risk factors and list predictive variables such as softer pancreatic texture, narrower pancreatic duct diameter, higher body mass index (BMI), longer surgery time, and no low-risk pathology type [pancreatic ductal adenocarcinoma (PDAC)/chronic pancreatitis (CP)] as recognized independent predictors of PPAP. What’s more, studies have identified PPAP as an independent risk factor for POPF onset[12], playing an inductive and accelerative role in its formation and progression[13,14]. The occurrence of PPAP exacerbates the severity of various postoperative complications and directly prolongs the patient’s recovery time[5,7,9,10]. Therefore, the ability to predict the occurrence of PPAP could guide clinicians in managing and treating patients perioperatively. Currently, no predictive models exist, so we try to develop a machine learning model to achieve this goal.

MATERIALS AND METHODS

Study design and participants

This retrospective cohort study was conducted at Beijing Tsinghua Changgung Hospital, affiliated with Tsinghua University, encompassing a time frame from January 2016 to June 2024. Consecutive patients who underwent pancreaticoduodenectomy (PD) during this period were included in the analysis. Exclusion criteria included the absence of postoperative AMY levels on postoperative days (PODs) 1 and 2, lack of CT scans within 10 days after operation, and incomplete perioperative data.

Procedures

All PDs were performed by the same surgical team comprising five senior surgeons with more than 15 years of experience. The procedure included both pylorus-preserving PD and classical PD. Pancreatic-enteric anastomoses were all performed as end-to-side double layer duct-to-mucosa anastomoses. An internal stenting was routinely placed, and external stenting was used optionally based on the surgeon’s preference, especially when there was a high estimated risk of POPF development. Three intra-abdominal drains were routinely placed: One posterior to the cholangio-enterostomy and the others posterior and anterior to the pancreatic-enteric anastomoses. Prophylactic somatostatin and antibiotics were administered to all patients for 3 to 5 days after surgery. A routine abdominal CT scan was performed within ten days after surgery. Drain fluid AMY (DFA) levels were regularly examined on PODs 1, 3, and 5. A conservative drainage management approach was applied, and the drain was usually started to be removed on POD 5, depending on the DFA level (< 5000 U/L on POD 1 and < 400 U/L on POD 5) and absence of signs of intra-abdominal infection.

Baseline characteristics were collected including: Age, sex, BMI, history of diabetes mellitus, hypertension, coronary artery disease, alcohol consumption, smoking, jaundice, chronic obstructive pulmonary disease, and American Society of Anesthesiologists physical status classification. Intraoperative details comprised: Main pancreatic duct diameter (MPDD), surgical modality, venous resection, pancreatic texture, pancreatic duct stent, surgery time, estimated blood loss (EBL), and pathology type. To assess postoperative laboratory changes in PPAP patients, data were collected including DFA, white blood cell count (WBC), serum AMY, and C-reactive protein levels on PODs 1 and 3. According to the definition proposed by ISGPS, POH was defined as sustained serum AMY levels exceeding the upper limit (> 100 U/L) within 48 hours after surgery. PPAP was diagnosed when clinical symptoms or radiological findings accompany POH4. Postoperative CT scans of patients with POH were independently reviewed by two experienced radiologists who were blinded to the clinical outcomes to assess for evidence of PPAP, including features such as exudates, fluid collections, edema, and necrosis[4]. Several typical CT changes of PPAP were shown on Figure 1. The complications including POPF, delayed gastric emptying, and postpancreatectomy hemorrhage were diagnosed according to ISGPS guidelines[11,15,16]. Other collected perioperative data included organ failure, 90-day mortality, interventional treatments, reoperations, drain removal time, postoperative hospital stay, and 30-day readmission.

Figure 1 Several typical computed tomography changes of postpancreatectomy acute pancreatitis. A: Peripancreatic fluid extravasation: Indicated by arrows as exudation belt; B: Residual peripancreatic fluid collection: Indicated by arrows as low-density fluid accumulation; C: Edema: Noticeable increase in the diameter of the pancreatic body and tail compared to preoperative dimensions; D: Necrotizing pancreatitis: Low-density shadow near the splenic hilum in the pancreatic tail.

Ethical considerations

Ethical approval for this study was obtained from the Ethics Committee of Beijing Tsinghua Changgung Hospital (Ethics Approval Code: 24017-7-01). Patient confidentiality and anonymity were rigorously maintained throughout the study, with all data anonymized prior to analysis. Informed consent was waived due to the retrospective nature of the study; however, measures to protect patient data confidentiality were strictly upheld.

Statistical analysis

Numerical variables were expressed as median [interquartile range (IQR)] or as median ± SD if normally distributed. Categorical variables are expressed as numbers and proportions. Student’s t-test or Mann-Whitney U test was used to assess numerical variables, while Pearson’s χ² test was employed for categorical variables. A significance level of P value < 0.05 indicated statistical differences. Machine learning models utilized 19 features encompassing both preoperative and intraoperative data to predict the risk of PPAP. Machine learning algorithms, including logistic regression (LR), random forest (RF), gradient-boosted decision trees (GBDT), extreme gradient boosting (XGBoost), category boosting (CatBoost), and light gradient boosting machine (LGBM) were employed to develop predictive models. The cut-off value of six models for the prediction of PPAP was determined using a receiver operating characteristic curve. The performance of these models was evaluated using a mean area under the receiver operating characteristic curve (AUC). Patients were randomly allocated to training and testing cohorts in an 8:2 ratio. Additionally, recursive feature elimination (RFE) was used to select several variables to optimize the machine learning algorithm. Feature importance was analyzed using the mean absolute Shapley value to gain a deeper understanding of the machine learning model. Statistical analyses were performed using Python (version 3.12.4) and the machine learning procedure was performed using Python Scikit-learn package.

Machine learning algorithms

In this study, a conventional LR approach was utilized alongside five widely used machine learning classification algorithms: RF, GBDT, XGBoost, LGBM, and CatBoost. These algorithms represent the most popular and state-of-the-art supervised machine learning methods for classification tasks.

LR is primarily used for binary classification tasks, LR applies a sigmoid function to predict probabilities and performs a logistic transformation on the result. This is a linear model known for its simplicity and interpretability[17]. RF is an ensemble method based on multiple decision trees, using majority voting to determine the final classification. RF is robust against noisy data and generally achieves high prediction accuracy[18,19]. GBDT is an ensemble of decision trees trained sequentially to minimize prediction error by fitting residuals at each step, recognized for its accuracy, efficiency, and interpretability[20]. XGBoost is an optimized gradient boosting algorithm that leverages second-order gradients for improved approximation, parameter tuning, and parallel learning, making it highly efficient and suitable for distributed computing[21]. LGBM is an improved GBDT variant using a histogram-based algorithm to aggregate gradient information, significantly enhancing training efficiency, especially on large datasets[22]. CatBoost is a gradient boosting algorithm with specific optimizations for categorical features, addressing gradient bias and prediction shift effectively, making it ideal for datasets with numerous categorical variables[23].

RESULTS

Clinical baseline characteristics and perioperative data

During the study period, 431 PDs were performed in our center. Among them, 50 patients were excluded: 40 due to missing serum AMY tests on PODs 1 and 2, 8 patients due to lacking postoperative CT scans within 10 days, and 2 patients due to other missing data. Finally, a total of 381 patients were included in the study (Figure 2). In the final statistical analysis, 39 features were included, comprising 19 preoperative and intraoperative variables incorporated into the ultimate machine learning models.

Figure 2 Flowchart: Method for sample selection, model training, validation, feature selection, and determination of the final model. The overall process of data extraction, training, and testing. AMY: Amylase; CT: Computed tomography; LR: Logistic regression; RF: Random forest; GBDT: Gradient boosting decision tree; LGBM: Light gradient boosting machine; XGBoost: Extreme gradient boosting; CatBoost: Category boosting; RFE: Recursive feature elimination; SHAP: Shapley additive explanation.

The clinical baseline characteristics and perioperative data are summarized in Table 1. Among these patients, 243 were male (63.8%). The median age was 65 years. One hundred fifty-four patients (40.4%) underwent PD for PDAC and CP. The median MPDD for all patients was 2.9 (IQR: 1.7, 4.7) mm. Pancreatic texture was soft in 187 patients (49.1%). Vascular reconstruction was performed in 83 patients (21.8%). The median intraoperative blood loss was 300 (IQR: 200, 400) L, and the median operation time was 370 (IQR: 306, 507) minutes.

Table 1 Clinical baseline data and surgical information, n (%).

Variables	Total (n = 381)
Age, median (IQR), years	65 (57, 70)
Sex: Female	138 (36.2)
BMI, mean ± SD, kg/m2	22.8 ± 3.3
Diabetes mellitus	91 (23.9)
Hypertension	105 (27.6)
CAD	28 (7.4)
Alcohol consumption	66 (17.3)
Smoking	74 (19.4)
Jaundice	241 (63.3)
COPD	13 (3.4)
ASA
I/II	243 (63.8)
III/IV	138 (36.2)
Pathology: PDAC/CP	154 (40.4)
MPDD, median (IQR), mm	2.9 (1.7, 4.7)
Surgical modality
PD	229 (60.1)
PPPD	152 (39.9)
Venous resection	83 (21.8)
Pancreatic texture
Firm	194 (50.9)
Soft	187 (49.1)
Pancreatic duct stent
No	6 (1.6)
Internal stent	344 (90.3)
External stent	31 (8.1)
surgery time, median (IQR), minute	370 (306, 507)
EBL, median (IQR), mL	300 (200, 400)

IQR: Interquartile range; BMI: Body mass index; CAD: Coronary artery disease; COPD: Chronic obstructive pulmonary disease; ASA: American Society of Anesthesiologists; PDAC: Pancreatic ductal adenocarcinoma; CP: Chronic pancreatitis; MPDD: Main pancreatic duct diameter; PD: Pancreaticoduodenectomy; PPPD: Pylorus-preserving pancreatoduodenectomy; EBL: Estimated blood loss.

Clinical outcomes

Among the 381 patients, 175 (45.9%) developed POH, and 88 of these patients (23.1%) exhibited postoperative CT changes (exudates, fluid collections, edema, and necrosis), meeting the criteria for a diagnosis of PPAP. Tables 2 and 3 compares and presents the clinical outcomes of patients with and without PPAP. Patients with PPAP had a lower incidence of PDAC/CP pathology type (22.7% vs 45.7%, P < 0.001), a smaller MPDD (median: 1.9 mm vs 3.1 mm, P < 0.001), a higher prevalence of a soft pancreas (71.6% vs 42.3%, P < 0.001), and greater intraoperative EBL (median: 400 mL vs 300 mL, P = 0.040). Additionally, PPAP patients had a higher BMI (mean: 23.4 kg/m²vs 22.7 kg/m²), though this difference was not statistically significant (P = 0.064). On POD 3, laboratory parameters including DFA, white blood cell count, C-reactive protein, and serum AMY were significantly elevated in PPAP patients (median: 2349 U/L vs 145 U/L, 11.2 × 10⁹/L vs 9.4 × 10⁹/L, 153.3 mg/L vs 101.3 mg/L, 95.0 U/L vs 30.0 U/L, respectively; all P < 0.001).

Table 2 Comparison of baseline preoperative and intraoperative characteristics in relation to postpancreatectomy acute pancreatitis, n (%).

Variables	Without PPAP (n = 293)	PPAP (n = 88)	P value
Age, median (IQR), years	65(58, 70)	64 (55, 69)	0.37
Sex, male	187 (63.8)	56 (63.6)	1
BMI, mean ± SD, kg/m2	22.7 ± 3.3	23.4 ± 3.2	0.064
Diabetes mellitus	73 (24.9)	18 (20.5)	0.473
Hypertension	84 (28.7)	21 (23.9)	0.454
CAD	23 (7.9)	5 (5.7)	0.652
Alcohol consumption	49 (16.7)	17 (19.3)	0.687
Smoking	53 (18.1)	21 (23.9)	0.295
Jaundice	185 (63.1)	56 (63.6)	1
COPD (%)	11 (3.8)	2 (2.3)	0.736
ASA, III/IV	101 (34.5)	37 (42.1)	0.242
Pathology, PDAC/CP	134 (45.7)	20 (22.7)	< 0.001
MPDD, median (IQR), mm	3.1 (1.9, 5.2)	1.9 (1.3, 3.0)	< 0.001
Surgical modality, PPPD	112 (38.2)	40 (45.5)	0.276
Venous resection	62 (21.2)	21 (23.9)	0.695
Pancreatic texture, soft	124 (42.3)	63 (71.6)	< 0.001
Pancreatic duct stent, external stent	19 (6.5)	12 (13.6)	0.094
Surgery time, median (IQR), minute	360 (305, 508)	393 (314, 491)	0.3
EBL, median (IQR), mL	300 (200, 400)	400 (200, 500)	0.04

PPAP: Postpancreatectomy acute pancreatitis; IQR: Interquartile range; BMI: Body mass index; CAD: Coronary artery disease; COPD: Chronic obstructive pulmonary disease; ASA: American Society of Anesthesiologists; PDAC: Pancreatic ductal adenocarcinoma; CP: Chronic pancreatitis; MPDD: Main pancreatic duct diameter; PPPD: Pylorus-preserving pancreatoduodenectomy; EBL: Estimated blood loss.

Table 3 Comparison of postoperative events in relation to postpancreatectomy acute pancreatitis, n (%).

Variables	Without PPAP (n = 293)	PPAP (n = 88)	P value
DFA on POD 1, median (IQR), U/L	553 (82, 2726)	4135 (1379, 8481)	< 0.001
DFA on POD 3, median (IQR), U/L	145 (42, 791)	2349 (627, 6153)	< 0.001
WBC on POD 1, median (IQR), 109/L	12.4 (10.2, 15.0)	13.6 (10.7, 16.9)	0.058
WBC on POD 3, median (IQR), 109/L	9.4 (7.0, 12.5)	11.2 (8.7, 15.1)	< 0.001
CRP on POD 1, median (IQR), mg/L	62.7 (43.2, 90.7)	76.7 (49.8, 116.1)	0.005
CRP on POD 3, median (IQR), mg/L	101.3 (67.0, 142.3)	153.3 (81.0, 201.1)	< 0.001
Serum AMY on POD 1, median (IQR), U/L	110.5 (53.0, 249.3)	312.0 (189.4, 508.3)	< 0.001
Serum AMY on POD 3, median (IQR), U/L	30.0 (18.2, 59.0)	95.0 (62.6, 149.0)	< 0.001
Drain removal time, median (IQR), days	12 (9, 20)	25 (14, 33)	< 0.001
Postoperative hospital stay, median (IQR), days	15.0 (11, 21)	23.5 (15, 33)	< 0.001
POPF	43 (14.7)	49 (55.7)	< 0.001
Grade C POPF	4 (1.4)	8 (9.1)	0.001
DGE	78 (26.6)	34 (38.6)	0.042
PPH	21(7.2)	18 (20.5)	0.001
Interventional treatment	26 (8.9)	28 (31.8)	< 0.001
Biliary fistula	7 (2.4)	5 (5.7)	0.229
Re-operation	8 (2.7)	11 (12.5)	0.001
Organ failure	1 (0.3)	6 (6.8)	< 0.001
90-day mortality	7 (2.4)	7 (8.0)	0.035
30-day readmission	9 (3.1)	6 (6.8)	0.203

PPAP: Postpancreatectomy acute pancreatitis; DFA: Drain fluid amylase; POD: Postoperative day; IQR: Interquartile range; WBC: White blood cell count; CRP: C-reactive protein; AMY: Amylase; POPF: Postoperative pancreatic fistula; DGE: Delayed gastric emptying; PPH: Postpancreatectomy hemorrhage.

In terms of complications, PPAP patients exhibited a significantly higher incidence of POPF (55.7% vs 14.7%, P < 0.001), grade C POPF (9.1% vs 1.4%, P = 0.001), delayed gastric emptying (38.6% vs 26.6%, P = 0.042), postpancreatectomy hemorrhage (20.5% vs 7.2%, P = 0.001), and 90-day mortality (8.0% vs 2.4%, P = 0.035). The higher incidence of major complications in PPAP patients contributed to significantly longer drain removal time (median: 25 days vs 12 days, P < 0.001) and postoperative hospital stay (median: 23.5 days vs 15.0 days, P < 0.001), as well as higher rates of reoperation (12.5% vs 2.7%, P = 0.001) and interventional treatment (31.8% vs 8.9%, P < 0.001).

Develop and evaluate the performance of the machine learning prediction models

Using a total of 19 preoperative and intraoperative variables, samples were randomly divided into training and validation sets in an 8:2 ratio to develop machine learning models (LR, RF, GBDT, XGBoost, LGBM, and CatBoost) for predicting the occurrence of PPAP after PD. Among the six models, CatBoost demonstrated superior performance with an AUC of 0.859 [95% confidence interval (CI): 0.814-0.905] in the training set and 0.822 (95%CI: 0.717-0.927) in the validation set. It also exhibited a specificity of 0.667, sensitivity of 0.857, negative predictive value (NPV) of 0.955, and positive predictive value (PPV) of 0.364. The performance evaluation of all six models is detailed in Figure 3 and Table 4. To obtain the optimal CatBoost model, RFE was employed, ultimately selecting five variables (pancreatic texture, MPDD, EBL, BMI, surgery time) to achieve the optimal CatBoost algorithm, which yielded an AUC of 0.812 (95%CI: 0.697-0.927). The performance evaluation of the final model is detailed in Figure 4 and Table 5.

Figure 3 Evaluating the performance of six different machine learning algorithms using preoperative and postoperative variables in both training and testing datasets to predict postpancreatectomy acute pancreatitis. A: Receiver operating characteristic curves of six machine learning algorithms in the training dataset; B: Receiver operating characteristic curves of six machine learning algorithms in the testing dataset. ROC: Receiver operating characteristic curves; AUC: Area under the receiver operating characteristic curve; LR: Logistic regression; RF: Random forest; GBDT: Gradient boosting decision tree; LGBM: Light gradient boosting machine; XGB: Extreme gradient boosting; CB: Category boosting.

Figure 4 Using the recursive feature elimination method to identify the optimal variables. A: Five variables were employed for the optimal category boosting algorithm; B: Receiver operating characteristic curves of the category boosting model based on selected variables. ROC: Receiver operating characteristic curves; AUC: Area under the receiver operating characteristic curve.

Table 4 Performance of different models for predicting postpancreatectomy acute pancreatitis.

Model	Training AUC	Testing AUC	Specificity	Sensitivity	MCC	Kappa	NPV	PPV
LR	0.605	0.69	0.524	0.857	0.295	0.214	0.943	0.286
RF	0.824	0.815	0.508	1	0.398	0.273	1	0.311
GBDT	0.875	0.735	0.81	0.643	0.392	0.379	0.911	0.429
XGBoost	0.871	0.706	0.762	0.714	0.392	0.365	0.923	0.4
LGBM	0.87	0.73	0.746	0.714	0.375	0.345	0.922	0.385
CatBoost	0.859	0.822	0.667	0.857	0.408	0.343	0.955	0.364

AUC: Area under the receiver operating characteristic curve; MCC: Matthews correlation coefficient; NPV: Negative predictive value; PPV: Positive predictive value; LR: Logistic regression; RF: Random forest; GBDT: Gradient boosting decision tree; XGBoost: Extreme gradient boosting; LGBM: Light gradient boosting machine; CatBoost: Category boosting.

Table 5 Performance of category boosting model for predicting postpancreatectomy acute pancreatitis based on selected variables.

Model	CatBoost
Training AUC	0.837
Testing AUC	0.812
Specificity	0.873
Sensitivity	0.714
MCC	0.535
Kappa	0.529
NPV	0.932
PPV	0.556

CatBoost: Category boosting; AUC: Area under the receiver operating characteristic curve; MCC: Matthews correlation coefficient; NPV: Negative predictive value; PPV: Positive predictive value.

Model interpretation

The shapley additive explanations method was also used to interpret the predictions achieved by the CatBoost model. According to the shapley additive explanations results, the two variables with the highest contribution to the model were pancreatic texture and MPDD. The other three significant variables were EBL, BMI, and surgery time, which may be associated with adverse outcomes resulting from the surgical procedures (Figure 5). Furthermore, the local interpretable model-agnostic explanations algorithm was utilized to elucidate the impact of different variables in the CatBoost model on prediction outcomes. We randomly selected four cases (true positive, true negative, false negative, and false positive) from the testing dataset to interpret the visualized prediction results (Figure 6).

Figure 5 Five most important variables and their impact on the category boosting model output by shapley additive explanations analysis. A: Summary of shapley additive explanations analysis on the dataset: Each dot represents a case, with the color indicating the feature’s value - blue for the lowest range and red for the highest range; B: The ranking of the five variables’ importance is determined by shapley additive explanations analysis. MPDD: Main pancreatic duct diameter; EBL: Estimated blood loss; BMI: Body mass; SHAP: Shapley additive explanations.

Figure 6 Results of local interpretable model-agnostic explanations applied to four randomly selected patients using category boosting model. MPDD: Main pancreatic duct diameter; BMI: Body mass index; EBL: Estimated blood loss.

DISCUSSION

As for diagnosing postoperative pancreatitis previously relied on the revised Atlanta criteria[3], referencing clinical presentations and imaging findings, which were subjective and influenced by clinicians’ experience, resulting in widely varying reported incidence rates across centers[12,24,25]. In 2016, Connor[2] established diagnostic criteria for PPAP, which required serum AMY levels to exceed the upper limit within 48 hours after surgery. However, relying exclusively on laboratory markers for diagnosis lead lower the diagnostic threshold, potentially resulting in a high incidence of PPAP, with some studies reporting rates exceeding fifty percent[25]. Our own research similarly found that relying solely on biochemical markers resulted in 175 out of 381 patients (45.93%) developing POH. Consequently, some scholars have questioned whether PPAP and POH represent distinct entities[25].

In order to standardize diagnostic criteria, the ISGPS proposed a consensus in 2022[4]. According to ISGPS guidelines, the diagnosis of POH (grade A) is made when serum AMY levels exceed the upper limit within 48 hours postoperatively. The diagnosis of PPAP (grade B or C) requires abdominal CT changes in the residual pancreas within 7 days postoperatively (such as exudates, fluid collections, edema, and necrosis) and clinically relevant features, graded according to the severity of associated harm (mild complications altering clinical management for grade B; severe complications including organ failure or death for grade C). Subsequently, studies on PPAP have trended towards normalization and standardization, with reported incidence rates of POH and PPAP across most centers showing less disparity, approximately 50% and 25%, respectively[5-10].

While PPAP generally represents a self-limiting inflammatory process, it can occasionally lead to necrosis or other serious postoperative complications. In cases of necrotizing pancreatitis, total pancreatectomy is the most effective approach. Loos et al[26] reported in a large clinical study that necrotizing pancreatitis is a primary indication for total pancreatectomy following PD, directly increasing the 90-day mortality rate by more than tenfold. PPAP significantly contributes to poor perioperative outcomes, with its occurrence leading to higher rates of POPF and more severe consequences of POPF[5]. Additionally, it is associated with increased rates of complications and perioperative mortality[7,8], resulting in longer hospital stays and higher medical costs, thus exacerbating both disease and economic burdens[9,10]. These hazards underscore the critical importance of early recognition of PPAP and intervention to facilitate recovery in post-PD patients, potentially reducing rates of other complications and mortality. Interestingly, Quero et al[6] reported that PPAP does not impact long-term prognosis (5-year overall survival and disease-free survival). Unfortunately, there is currently no effective model to predict the occurrence of PPAP.

To our knowledge, although previous studies have used univariate and multivariate LR methods to identify independent risk factors for PPAP[5-10], the application of novel methodologies such as machine learning algorithms for prediction has not been previously reported, making our study the first to employ this approach. After the ultimate model establishment and optimization, the final Catboost model, selected after variable screening using RFE, demonstrated commendable performance with an AUC of 0.837 (95%CI: 0.788-0.886) in the training cohort and 0.812 (95%CI: 0.697-0.927) in the testing cohort. CatBoost was selected as the final model due to its strong performance across accuracy, interpretability, and efficiency, which are critical for clinical applications. Recent studies highlight CatBoost’s utility in medical prediction. For instance, it has been used effectively in intensive care unit mortality and obesity risk prediction[27,28]. These studies showcase CatBoost’s suitability for clinical settings, especially in handling imbalanced datasets and categorical variables, common in healthcare. Our results also reflect this trend. CatBoost achieved a balanced performance in predicting PPAP compared to models like LR and RF. This balance between interpretability and predictive power supports CatBoost as a reliable choice in complex medical prediction tasks, justifying its selection in our study.

In our study, the final CatBoost model achieved a PPV of 0.556 compared to other models, which, while modest, is still meaningful in a clinical context where predictive accuracy for high-risk cases is challenging due to low prevalence. A higher PPV would indicate a greater likelihood that a positive prediction truly corresponds to PPAP, aiding in early identification of high-risk patients. NPV measures the probability that patients predicted as low-risk genuinely do not develop PPAP. With an NPV of 0.932, the CatBoost model demonstrates a high confidence level for ruling out PPAP in patients classified as low-risk. This high NPV is particularly valuable in the postoperative context, where unnecessary treatments can be avoided, improving patient outcomes and resource allocation. Kappa and Matthews correlation coefficient (MCC) further contextualize the model’s performance by providing metrics on overall agreement and the balance between positive and negative predictions, respectively. These values reflect how well the model performs relative to random chance (Kappa) and the correlation between true and predicted values (MCC), ensuring a more nuanced understanding of its prediction capability. CatBoost’s balanced Kappa and MCC values demonstrate that it consistently predicts PPAP cases more accurately than random chance, reinforcing its suitability for clinical deployment.

Machine learning models have numerous applications beyond predicting postoperative complications, as demonstrated in this article. Guo et al[29] analyzed advancements in adjuvant and neoadjuvant immunotherapy related to oncology, uncovering critical features and correlations in the field that underscore its importance in cancer treatment. Such qualitative insights support clinical decision-making, reducing potential wastage of medical resources[29]. Additionally, the AlphaFold artificial intelligence model has been applied to research in molecular biology and drug discovery, transforming biological research with a disruptive impact in structural biology[30]. With continuous innovation and iteration in artificial intelligence technology, it is anticipated that even greater achievements will emerge across diverse fields in the future.

PPAP and POPF are interrelated complications following PD[31,32], potentially linked through acinar cell density[33,34]. The acinar component of the remnant pancreas is recognized as an intrinsic risk factor, while operative stress, remnant hypoperfusion, and inflammation represent common mechanisms of acinar cell injury[14]. Studies have identified PPAP as an independent risk factor for POPF onset[12]. PPAP plays an important role in formation and progression of POPF[13,14]. On one hand, local inflammation from PPAP affects the pancreatoenteric anastomosis, while on the other hand, PPAP and POPF mutually exacerbate each other’s risks. Some studies differentiated subgroups to further validate mutual interactions exacerbating perioperative recovery in patients with concurrent PPAP and POPF[5,35]. Considering the inherent leakage and peripancreatic fluid collection in pancreatitis[36], distinguishing between peripancreatic fluid collections due to POPF on CT imaging remains challenging. Due to these factors, substantial variability exists among centers reporting PPAP combined with POPF, intricately linked to the progression of these complications, thereby increasing research complexity. In our study, employing model interpretability techniques, variables contributing most to machine learning prediction of PPAP included pancreatic texture, MPDD, and BMI, aligning with previous findings using LR methods for identifying PPAP risk factors[5,7,8]. This further validates the consistency and reliability of our machine learning model with real outcomes. Additionally, these three preoperative variables overlap with risk factors in POPF prediction models[37-41], suggesting a complementary and mutually reinforcing relationship between these two post-PD complications. Future research directions should involve larger sample sizes and multi-center collaborations to elucidate the mutual influence and preventive mechanisms of PPAP and POPF. Moreover, the development of a classification system for PPAP risk factors, akin to the clinically relevant POPF classification system for pancreas-associated risk factors[42], is urgently needed to enhance result comparability and guide management and prevention strategies for high-risk populations.

Efforts to prevent PPAP have involved various strategies. Bellotti et al[43] found that preoperative use of low molecular-weight heparin effectively reduces the incidence of PPAP, potentially by altering local microcirculation in the residual pancreas. Anti-inflammatory therapy targeting P-suPAR decrease has similarly shown protective effects[44]. Additionally, hydrocortisone and somatostatin can be utilized to suppress local inflammation and reduce pancreatic secretion for prevention[45-47]. Furthermore, transection plane of the pancreas on the left of the superior mesenteric vein, or strict control of intraoperative fluid intake (less than 3 mL/kg) is associated with a lower incidence of PPAP and POPF[48-50]. Protease inhibitors do not confer protective benefits for high-risk anastomoses[51], and several studies have reported a close correlation between PPAP and the inflammatory marker C-reactive protein levels ≥ 18.0 mg/dL[12]. Patients with malignancy and those undergoing neoadjuvant therapy exhibit pancreatic atrophy postoperatively with diminished exocrine function, seemingly advantageous for reducing PPAP and POPF incidence[13,47]. It is noteworthy that acute pancreatitis following endoscopic retrograde cholangiopancreatography can effectively be prevented with non-steroidal anti-inflammatory drugs[52], although there is currently a lack of studies or reports regarding non-steroidal anti-inflammatory drugs for PPAP prevention. Further exploration in this area is needed.

On top of that, the primary utility of this predictive model lies in its ability to guide clinical practice. We input preoperative data and certain intraoperative information into the model to perform risk assessments. For patients identified as high-risk, we can implement more proactive postoperative monitoring to timely detect and address the occurrence of complications. However, there are currently no widely recognized effective preventive strategies specifically targeting PPAP. Previous sections have highlighted several potentially useful approaches. Conversely, for patients categorized as low-risk, we can consider early removal of drainage tubes to facilitate recovery. This strategy aims to enhance the rehabilitation efficiency of patients following PD and to minimize the potential harms caused by complications.

However, it is important to acknowledge several limitations in our study. First, as a retrospective study, it is inherently subject to selection bias, and all data were sourced from a single medical center. Second, the retrospective nature of data collection led to instances of missing information, which adversely affected the performance of our model. Additionally, the absence of external validation for the dataset has somewhat undermined the reliability of our findings. Lastly, despite not accounting for a limited range of clinical features, there is potential for this model to be developed into software in the future, employing a standardized and simplified input format for risk assessment, thereby guiding clinical practice strategies and facilitating future large-scale prospective studies. We plan to collaborate with multiple centers for a joint research validation and hope to further optimize the model to achieve more accurate results. We will also continue to report our findings.

CONCLUSION

Our study demonstrated a novel machine learning algorithm for predicting PPAP, achieving excellent performance. The model’s coherent interpretation provided clinical utility in the management and treatment of patients following PD. However, future research necessitates larger cohorts and prospective studies to explore effective pharmacological preventive strategies and long-term prognostic implications, thereby advancing our understanding of this emerging complication concept.