Development and validation of an interpretable machine learning model for predicting acute kidney injury after pancreatic surgery

Abstract

BACKGROUND

Acute kidney injury (AKI) is a common and serious complication of major abdominal surgery. However, predictive models specific to pancreatic surgery remain scarce.

AIM

To develop and validate an interpretable machine learning model for early prediction of postoperative AKI following pancreatic surgery.

METHODS

Adults undergoing pancreaticoduodenectomy or distal or total pancreatectomy from 2014 to 2024 were retrospectively analyzed. AKI was defined by the kidney disease Improving global outcomes creatinine-based criteria. After matching, data from 2014-2021 trained seven models using Boruta/least absolute shrinkage and selection operator selection and five-fold cross-validation. Data from 2022-2024 were utilized for validation. Model performance was evaluated by the area under the receiver operating characteristic curve (AUROC); Shapley Additive Explanations were used for interpretation, and an online calculator was developed.

RESULTS

Among the 4216 eligible patients, 230 (5.5%) developed postoperative AKI. The categorical boosting model showed the best performance in the training cohort (AUROC = 0.803) and maintained a robust prediction in the validation cohort (AUROC = 0.751). Shapley Additive Explanations analysis highlighted operative time, postoperative serum creatinine level, intensive care unit admission after surgery, postoperative white blood cell count, and intraoperative red blood cell transfusion as key features for predicting postoperative AKI.

CONCLUSION

The developed models showed satisfactory performance for predicting postoperative AKI in patients undergoing major pancreatic surgery. They may facilitate early high-risk identification and inform perioperative management strategies.

Key Words: Acute kidney injury; Machine learning; Pancreatic surgery; Prediction model; Risk factors

Core Tip: Using a decade-long pancreatic surgery cohort, we developed and validated an interpretable machine learning model to identify patients at risk for postoperative acute kidney injury using routinely available perioperative data. The final model achieved good discrimination and highlighted modifiable drivers of acute kidney injury, including longer operative time, early postoperative serum creatinine elevation, admission to the intensive care unit, increased postoperative white blood cell count, and greater intraoperative blood loss. An online risk calculator is provided to support bedside individualized prediction and early renoprotective management.

INTRODUCTION

Acute kidney injury (AKI) is a common and serious postoperative complication characterized by an abrupt decline in renal function. The reported incidence of AKI after major abdominal surgery ranges from 3.1% to 35.3%, reflecting heterogeneity in both operative approaches and AKI diagnostic criteria[1]. Postoperative AKI is associated with prolonged intensive care unit (ICU) and hospital stays, a higher rate of complications, and significantly increased short-term mortality[2]. Moreover, postoperative AKI predisposes patients to chronic kidney disease[3,4] and long-term mortality[1,5,6]. Unlike other perioperative complications, AKI often presents insidiously, precluding timely renoprotective intervention. Thus, early identification of postoperative AKI is critical.

While extensive research has focused on AKI following cardiac surgery[7,8], fewer studies have addressed AKI in the context of major abdominal surgeries, especially pancreatic procedures[1,9,10]. Pancreaticoduodenectomy (PD) and other pancreatic surgeries have major complication rates of 30%-40% and a mortality rate of approximately 3% even at high-volume centers[11,12]. These operations often involve prolonged operative time, extensive visceral dissection, vascular reconstruction, substantial blood loss, and large fluid shifts, all of which increase the vulnerability of the kidney and underscore the necessity of procedure-specific risk prediction tools risk of AKI. This highlights the need for procedure-specific prediction tools.

Postoperative AKI is driven by a complex interplay of baseline patient susceptibility and perioperative stressors, and its causes are often complex and multifactorial rather than attributable to a single pathophysiologic process[2,13]. These stressors encompass surgery-related insults and early postoperative physiologic derangements, including renal hypoperfusion, oxido-inflammatory stress, nephrotoxic exposures, and downstream complications or care-related factors such as sepsis, mechanical ventilation, blood transfusion, and fluid imbalance[13].

While logistic regression remains a foundational machine learning (ML) method for establishing interpretable baseline prediction models, it primarily captures linear and additive effects unless nonlinear terms and higher order interactions are explicitly specified. To address these limitations advanced ML algorithms, particularly tree-based ensembles, can model nonlinear relationships and higher order interactions within high-dimensional perioperative data, and such methods have shown promising results in surgical cohorts[14,15]. Moreover, the development of techniques such as SHapley Additive exPlanations (SHAP) allows for transparent model interpretation, thereby enhancing clinical acceptability and utility[16].

In this study, we comprehensively analyzed perioperative data and constructed an interpretable ML-based model to predict postoperative AKI following pancreatic surgery.

MATERIALS AND METHODS

Study design

This retrospective cohort study included adult patients who underwent PD, distal pancreatectomy (DP), or total pancreatectomy (TP) at the Peking Union Medical College Hospital (PUMCH) between January 2014 and December 2024. Patients were excluded if they had chronic kidney disease ≥ G3a, preoperative AKI, prior pancreatic surgery, concurrent urinary tract resection, or incomplete laboratory data. A 1:1 matched control cohort was selected based on age, sex, body mass index, and surgery type. Ethical approval was obtained, and the requirement for informed consent was waived due to the retrospective nature of the study.

AKI definition

Postoperative AKI was defined by the kidney disease improving global outcomes (KDIGO) serum creatinine (SCr)-based criteria: An increase of ≥ 0.3 mg/dL within 48 hours or ≥ 1.5 × baseline within 7 days after surgery. The hourly urine output criteria were not applied because of incomplete records. AKI was further classified into three stages: SCr increase of 1.5 times to 2.0 times baseline (stage I); SCr increase of 2.0 times to 3.0 times baseline (stage II); And SCr increase of ≥ 3.0 times baseline (stage III). In this study AKI cases were retrospectively identified using the AKI virtual ward, a hospital-wide real-time monitoring system developed at PUMCH. This system continuously monitors SCr levels of all inpatients and automatically detects potential AKI events according to the KDIGO SCr-based criteria. Although the system was not in clinical use during the study period (2014-2024), it was retrospectively applied to identify eligible cases.

Data collection

Given the complex and multifactorial etiology of postoperative AKI, we sought to construct a comprehensive feature set to capture potential risk factors across the entire perioperative continuum. Variable selection was guided by clinical relevance based on prior literature and data availability within routine electronic health records. Consequently, a total of 53 variables were collected and categorized as preoperative, intraoperative, and postoperative. Preoperative data encompassed demographics, comorbidities (e.g., hypertension, diabetes mellitus, coronary artery disease, pancreatitis, and jaundice), medication history [e.g., nonsteroidal anti-inflammatory drugs (NSAIDs), angiotensin-converting enzyme inhibitors/angiotensin receptor blockers, neoadjuvant therapy], American Society of Anesthesiologists (ASA) status, and the most recent laboratory tests within 3 days before surgery.

Intraoperative variables included surgery type (PD, DP, or TP), surgical approach (open, laparoscopic, or robotic), operative time, estimated blood loss, transfusion, volume of administered fluids (crystalloid/colloid), and urine output. Cases that converted from laparoscopy to open surgery were classified as open. Postoperative variables included the first laboratory values on postoperative day (POD) 1 and daily fluid input/output over the first 3 days. Major complications, such as clinically relevant postoperative pancreatic fistula (CR-POPF), post-pancreatectomy hemorrhage (PPH), and delayed gastric emptying, were defined according to the International Study Group of Pancreatic Surgery criteria. Intra-abdominal infection was defined as postoperative fever with positive culture or radiological data. Bile leak and gastrointestinal fistula were diagnosed endoscopically or radiographically.

Model development and validation

Patients treated from 2014 to 2021 formed the training cohort while those treated from 2022 to 2024 comprised the validation cohort. To minimize the risk of data leakage and ensure the utility of the model for early prediction, only variables available within the first 24 hours postoperatively were retained for the subsequent feature selection process. Boruta and least absolute shrinkage and selection operator (LASSO) algorithms were used for feature selection, and selected features were used to train seven ML algorithms: Logistic regression; Random forest; K-nearest neighbors; Support vector machine; Light gradient boosting machine; Extreme gradient boosting with classification trees; And categorical boosting (CatBoost). Five-fold cross-validation and grid search were employed for tuning. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), and the model with the best performance was tested on the validation cohort. Calibration was assessed using calibration plots and the Brier score, precision-recall curves appraised discrimination, and decision curve analysis quantified clinical net benefit.

Model interpretation

To ensure transparent and robust model interpretation, the SHAP method was employed to evaluate the importance of the model variables and rank them based on their contribution to the outcome. Rooted in cooperative game theory[17], the SHAP framework offers significant advantages over conventional feature importance techniques. This approach not only ranks features but also reveals the nature of their influence (e.g., positive or negative correlation with the outcome) through rich visualizations such as SHAP summary plots.

Development of a web-based calculator

We developed an interactive web-based risk prediction calculator with the Streamlit framework (https://streamlit.io/) to estimate the probability of postoperative AKI in patients undergoing pancreatic surgery. This online tool allows clinicians to input patient-specific perioperative variables and incorporates a SHAP force plot to illustrate how each variable influences the patient’s predicted risk.

Statistical analysis

All analyses were conducted using Python version 3.6.5 and SPSS Statistical Software Version 23.0 (IBM Corp., Armonk, NY, United States). Continuous variables were reported as median (interquartile range) and compared using the Wilcoxon rank-sum test. Categorical variables were presented as n (%) and compared using Pearson’s χ² test or Fisher’s exact test as appropriate. Outliers were handled via the 3-sigma rule. Missing data were imputed using multiple imputations with predictive mean matching. Variance inflation factor analysis was used to screen for multicollinearity with variables exhibiting a variance inflation factor > 10 considered for exclusion. A two-sided P < 0.05 was considered statistically significant.

RESULTS

Patient characteristics

Between January 2014 and December 2024, 4216 consecutive patients underwent pancreatic surgery at PUMCH and met the inclusion criteria (Figure 1). Among them, 230 (5.5%) patients developed postoperative AKI and were matched with 234 controls without AKI, yielding an overall cohort of 464 patients. Among the overall cohort, the median age was 64 years and 285 (61.0%) patients were male. As summarized in Table 1, patients with AKI had a higher prevalence of coronary artery disease, chronic kidney disease, malignancy history, and long-term NSAIDs use. The AKI group also had higher blood urea nitrogen, prothrombin time (PT), and carbohydrate antigen 19-9. Imaging demonstrated a larger tumor diameter in the AKI cohort, and a greater proportion was classified as ASA ≥ IIIa.

Figure 1 Flow diagram of patient selection. ¹Pancreatic surgery encompassed pancreaticoduodenectomy, distal pancreatectomy, and total pancreatectomy. AKI: Acute kidney injury; CKD: Chronic kidney disease; BMI: Body mass index; KDIGO: Kidney disease improving global outcomes; SCr: Serum creatinine.

Table 1 Characteristics of the patient cohorts, n (%).

Variables	Overall cohort			P value	Training cohort (n = 370)	Validation cohort (n = 94)	P value
	Total (n = 464)	AKI (n = 230)	Non-AKI (n = 234)
Demographic data
Age (years), median (IQR)	64 (56 to 69)	65 (57 to 70)	63 (55 to 68)	0.118	63.5 (56 to 69)	64 (57 to 70)	0.402
Male	285 (61)	143 (62)	142 (61)	0.742	229 (63)	56 (57)	0.327
BMI (kg/m2), median (IQR)	23.05 (20.81 to 25.30)	23.02 (20.89 to 25.80)	23.10 (20.77 to 24.65)	0.210	23.04 (20.76 to 25.35)	23.38 (20.94 to 25.00)	0.991
Smoking	150 (32)	69 (30)	81 (35)	0.288	118 (32)	32 (33)	0.938
Drinking	133 (29)	63 (27)	70 (30)	0.548	107 (29)	26 (27)	0.599
Medical history
HTN	173 (37)	90 (39)	83 (35)	0.415	130 (36)	43 (44)	0.129
DM	139 (30)	76 (33)	63 (27)	0.150	107 (29)	32 (33)	0.512
CHD	50 (11)	36 (16)	14 (6)	< 0.001	34 (9)	16 (16)	0.046
Stroke	30 (6)	11 (5)	19 (8)	0.144	21 (6)	9 (9)	0.218
Pancreatitis	36 (8)	17 (7)	19 (8)	0.769	21 (6)	15 (15)	0.002
Jaundice	155 (33)	82 (36)	73 (31)	0.309	119 (33)	36 (37)	0.431
Malignancy history	38 (8)	28 (12)	10 (4)	0.002	24 (7)	14 (14)	0.013
CKD of G1/G2	6 (1)	6 (3)	0 (0)	0.013	4 (1)	2 (2)	0.461
Abdominal surgery history	139 (30)	72 (31)	67 (29)	0.530	100 (27)	39 (40)	0.017
Preoperative medication
NSAIDs	19 (4)	17 (7)	2 (1)	< 0.001	13 (4)	6 (6)	0.254
ACEI/ARB	67 (14)	35 (15)	32 (14)	0.637	45 (12)	22 (22)	0.011
Neoadjuvant	18 (4)	9 (4)	9 (4)	0.970	16 (4)	2 (2)	0.289
Preoperative laboratory findings
HGB (g/L), median (IQR)	130 (119 to 141)	129 (117 to 141)	131 (120 to 141)	0.169	130 (119 to 141)	131 (117 to 142)	0.685
WBC (× 109/L), median (IQR)	5.74 (4.79 to 6.83)	5.75 (4.86 to 7.11)	5.74 (4.72 to 6.64)	0.600	5.80 (4.86 to 6.82)	5.55 (4.53 to 6.91)	0.210
PLT (× 109/L), median (IQR)	211 (172 to 252)	210 (163 to 249)	212.5 (178 to 259)	0.181	214 (173 to 251)	200 (166 to 257)	0.240
ALT (U/L), median (IQR)	31 (16 to 82)	32.5 (16 to 79)	28.5 (15 to 84)	0.524	31.5 (16 to 82)	27.5 (14 to 76)	0.891
Alb (g/L), median (IQR)	40 (37 to 43)	40 (37 to 43)	41 (37 to 43)	0.416	40 (37 to 43)	40 (37 to 44)	0.904
TBil (μmol/L), median (IQR)	15.15 (10.00 to 50.90)	17.75 (9.80 to 70.30)	14.05 (10.20 to 39.00)	0.067	15.15 (9.60 to 54.90)	15.15 (11.00 to 42.40)	0.659
SCr (μmol/L), median (IQR)	64 (54 to 73)	64 (54 to 73)	64 (56 to 72)	0.806	64 (55 to 73)	64 (53 to 72)	0.393
Urea (μmol/L), median (IQR)	4.59 (3.78 to 5.81)	4.77 (3.94 to 6.01)	4.43 (3.57 to 5.56)	0.014	4.57 (3.78 to 5.79)	4.62 (3.79 to 5.84)	0.600
hsCRP (mg/L), median (IQR)	1.97 (0.84 to 5.82)	1.97 (0.85 to 5.49)	1.97 (0.75 to 6.16)	0.847	2.10 (0.84 to 7.32)	1.72 (0.87 to 4.32)	0.305
PT (seconds), median (IQR)	11.8 (11.4 to 12.5)	11.9 (11.4 to 12.6)	11.7 (11.3 to 12.3)	0.024	11.9 (11.4 to 12.5)	11.7 (11.3 to 12.3)	0.362
APTT (seconds), median (IQR)	26.8 (25.2 to 28.9)	27.0 (25.2 to 29.0)	26.5 (25.1 to 28.7)	0.133	26.85 (25.2 to 28.9)	26.55 (25.1 to 28.5)	0.633
CA19-9 (U/mL), median (IQR)	50.5 (14.6 to 189.0)	66.7 (18.8 to 284.6)	39.5 (11.8 to 145.0)	0.004	49.9 (14.2 to 181.5)	50.8 (17.2 to 199.0)	0.613
CEA (U/mL), median (IQR)	2.53 (1.67 to 4.30)	2.67 (1.70 to 4.61)	2.39 (1.61 to 3.90)	0.066	2.50 (1.67 to 4.30)	2.62 (1.66 to 4.37)	0.856
Preoperative imaging findings
Tumor length (cm), median (IQR)	2.7 (1.8 to 3.8)	3.0 (2.0 to 4.0)	2.5 (1.7 to 3.5)	0.002	2.7 (1.8 to 3.8)	2.9 (1.8 to 4.0)	0.979
ASA physical status ≥ III	107 (24)	63 (29)	44 (19)	0.017	85 (23)	24 (24)	0.793
Type of surgery
PD	361 (78)	180 (78)	181 (77)	0.813	285 (78)	76 (78)	0.946
DP/TP	103 (22)	50 (22)	53 (23)		81 (22)	22 (22)
Open surgery	259 (56)	143 (62)	116 (50)	0.006	209 (57)	50 (51)	0.281
Laparoscope/robotic	205 (44)	87 (38)	118 (50)		157 (43)	48 (49)
Intraoperative variables
Duration of surgery (hour), median (IQR)	4.50 (4.13 to 6.75)	5.29 (4.50 to 7.33)	4.50 (4.00 to 5.42)	< 0.001	4.50 (4.08 to 6.75)	4.94 (4.33 to 7.00)	0.332
Vein reconstruction	39 (8)	27 (12)	12 (5)	0.010	32 (9)	7 (7)	0.612
Artery reconstruction	9 (2)	7 (3)	2 (1)	0.087	3 (1)	6 (6)	< 0.001
Multi-organ resection	40 (9)	27 (12)	13 (6)	0.018	29 (8)	11 (11)	0.301
Blood loss (mL), median (IQR)	400 (200 to 800)	500 (300 to 800)	400 (200 to 650)	< 0.001	400 (200 to 800)	400 (200 to 800)	0.877
Crystalloid (mL), median (IQR)	3100 (2300 to 3700)	3200 (2600 to 3900)	2800 (2200 to 3600)	< 0.001	3100 (2300 to 3700)	3100 (2300 to 3700)	0.820
Colloidal (mL), median (IQR)	1000 (500 to 1000)	1000 (500 to 1000)	500 (500 to 1000)	0.152	500 (500 to 1000)	1000 (500 to 1000)	0.145
RBC transfusion (mL), median (IQR)	0 (0 to 800)	400 (0 to 800)	0 (0 to 400)	< 0.001	0 (0 to 550)	0 (0 to 800)	0.171
Plasma transfusion (mL), median (IQR)	0 (0 to 400)	0 (0 to 400)	0 (0 to 400)	< 0.001	0 (0 to 400)	0 (0 to 400)	0.069
Urine output (mL), median (IQR)	600 (400 to 1100)	700 (400 to 1100)	600 (400 to 1000)	0.370	600 (400 to 1000)	800 (400 to 1300)	0.126
Postoperative laboratory findings
HGB (g/L), median (IQR)	116 (105 to 129)	116 (105 to 126)	117 (106 to 131)	0.112	116 (105 to 129)	117 (106 to 129)	0.400
WBC (× 109/L), median (IQR)	12.18 (9.85 to 15.05)	12.53 (10.06 to 15.81)	11.85 (9.68 to 14.04)	0.036	12.29 (10.08 to 15.24)	11.52 (9.05 to 13.87)	0.022
PLT (× 109/L), median (IQR)	180 (143 to 215)	177 (132 to 209)	182 (151 to 219)	0.064	180 (143 to 219)	180 (137 to 209)	0.371
Alb (g/L), median (IQR)	32 (29 to 35)	31 (28 to 35)	33 (29 to 36)	0.002	32 (29 to 36)	32 (29 to 34)	0.360
TBil (μmol/L), median (IQR)	20.9 (14.1 to 47.1)	25.4 (15.1 to 61.7)	18.6 (13.2 to 35.9)	< 0.001	21.0 (14.3 to 48.6)	20.5 (13.5 to 41.2)	0.520
SCr (μmol/L), median (IQR)	63 (52 to 77)	67 (52 to 83)	61 (52 to 72)	0.005	63 (52 to 76)	64 (50 to 77)	0.607
Urea (μmol/L), median (IQR)	4.12 (3.14 to 5.42)	4.44 (3.42 to 5.57)	3.87 (2.99 to 5.08)	< 0.001	4.13 (3.14 to 5.42)	4.08 (3.11 to 5.48)	0.718
PT (seconds), median (IQR)	12.8 (12.3 to 13.6)	13.0 (12.3 to 13.8)	12.8 (12.2 to 13.5)	0.016	12.8 (12.2 to 13.5)	13.0 (12.3 to 13.8)	0.245
APTT (seconds), median (IQR)	27.5 (25.2 to 30.9)	27.6 (25.3 to 31.1)	27.4 (24.8 to 30.7)	0.315	27.4 (25.2 to 30.6)	27.9 (25.3 to 32.0)	0.366
Postoperative fluid management
Total input on POD0 (mL/kg), median (IQR)	89.34 (72.85 to 115.45)	93.79 (74.42 to 126.07)	86.12 (71.65 to 106.30)	0.004	88.52 (72.85 to 111.50)	96.79 (73.03 to 128.95)	0.044
Input on POD1 (mL/kg), median (IQR)	44.58 (37.72 to 51.69)	44.54 (36.51 to 50.87)	44.59 (38.68 to 53.22)	0.464	44.37 (36.92 to 51.60)	45.72 (40.37 to 52.08)	0.150
Output on POD1 (mL/kg), median (IQR)	33.96 (26.91 to 43.87)	34.12 (27.55 to 44.73)	33.41 (26.54 to 42.57)	0.446	33.30 (26.35 to 42.57)	35.89 (28.72 to 48.27)	0.040
Net fluid of POD1 (mL/kg), median (IQR)	8.94 (0.65 to 17.38)	8.34 (-0.19 to 17.75)	9.65 (1.96 to 16.86)	0.281	9.00 (1.54 to 17.51)	8.45 (-2.6 to 16.38)	0.185
Input on POD2 (mL/kg), median (IQR)	43.08 (36.34 to 49.25)	43.60 (35.34 to 49.70)	42.90 (37.07 to 48.53)	0.765	43.07 (36.25 to 49.19)	43.21 (36.54 to 49.26)	0.613
Output on POD2 (mL/kg), median (IQR)	32.07 (24.25 to 42.38)	34.98 (24.92 to 44.00)	29.18 (23.54 to 39.51)	0.002	31.82 (24.06 to 42.08)	32.69 (24.34 to 47.55)	0.336
Net fluid of POD2 (mL/kg), median (IQR)	9.49 (0.23 to 18.26)	8.23 (-1.37 to 16.28)	10.64 (3.01 to 19.32)	0.002	9.68 (0.33 to 18.49)	9.23 (-0.09 to 17.42)	0.329
Input on POD3 (mL/kg), median (IQR)	40.99 (34.59 to 47.51)	41.81 (34.01 to 49.09)	40.50 (35.22 to 46.51)	0.396	40.69 (34.60 to 47.22)	41.65 (34.16 to 49.21)	0.475
Output on POD3 (mL/kg), median (IQR)	30.95 (23.98 to 40.10)	33.11 (24.68 to 44.13)	29.30 (22.57 to 38.05)	< 0.001	30.63 (24.14 to 39.65)	31.99 (23.86 to 43.75)	0.426
Net fluid of POD3 (mL/kg), median (IQR)	9.55 (2.35 to 17.76)	7.17 (0.48 to 16.51)	11.92 (4.32 to 18.85)	< 0.001	9.51 (2.28 to 17.60)	10.27 (2.59 to 17.85)	0.865
Postoperative complications
CR-POPF	77 (17)	47 (20)	30 (13)	0.028	56 (15)	21 (21)	0.148
PPH grades B and C	75 (16)	59 (26)	16 (7)	< 0.001	58 (16)	17 (17)	0.720
Abdominal infection	96 (21)	77 (33)	19 (8)	< 0.001	73 (20)	23 (23)	0.444
Shock	57 (12)	54 (23)	3 (1)	< 0.001	42 (11)	15 (15)	0.305
DGE grades B and C	89 (19)	47 (21)	42 (18)	0.454	72 (20)	18 (18)	0.772
Undesirable healing	23 (5)	12 (5)	11 (5)	0.798	17 (5)	6 (6)	0.549
BL	27 (6)	17 (7)	10 (4)	0.148	24 (7)	3 (3)	0.189
Gas/intestinal fistula	7 (2)	6 (3)	1 (0)	0.054	5 (1)	2 (2)	0.627
Clinical outcome
AKI stage							0.556
I		190 (83)			152 (84)	38 (78)
II		23 (10)			17 (9)	6 (12)
III		17 (7)			11 (6)	6 (12)
AKI onset time (POD) (days)		4 (2 to 10)			3 (1 to 8)	3 (2 to 8)
Length of ICU stay (days), median (IQR)	1.74 (4.05)	2.93 (5.43)	0.56 (0.89)	< 0.001	1.76 (4.21)	1.66 (3.37)	0.745
Length of hospital stay (days), median (IQR)	23 (17 to 34)	29 (18 to 44)	20 (15 to 28)	< 0.001	22 (17 to 34)	24 (17 to 33)	0.637
Reoperation	40 (9)	35 (15)	5 (2)	< 0.001	29 (8)	11 (11)	0.301
In-hospital mortality	12 (3)	12 (5)	0 (0)	< 0.001	11 (3)	1 (1)	0.272

AKI: Acute kidney injury; BMI: Body mass index; IQR: Interquartile range; HTN: Hypertension; CHD: Coronary heart disease; DM: Diabetes mellitus; ACEI: Angiotensin-converting enzyme inhibitors; CKD: Chronic kidney disease; NSAID: Nonsteroidal anti-inflammatory drugs; ARB: Angiotensin receptor blockers; HGB: Hemoglobin; ALT: Alanine aminotransferase; PLT: Platelet count; WBC: White blood cell; Alb: Albumin; TBil: Total bilirubin; hsCRP: High-sensitivity C-reactive protein; SCr: Serum creatinine; PT: Prothrombin time; APTT: Activated partial thromboplastin time; CA19-9: Carbohydrate antigen 19-9; CEA: Carcinoembryonic antigen; ASA: American Society of Anesthesiologists; PD: Pancreaticoduodenectomy; DP: Distal pancreatectomy; TP: Total pancreatectomy; POD: Postoperative day; CR-POPF: Clinically relevant postoperative pancreatic fistula; RBC: Red blood cell; DGE: Delayed gastric emptying; PPH: Post-pancreatectomy hemorrhage; BL: Bile leak; ICU: Intensive care unit.

Regarding intraoperative factors, patients with AKI more frequently underwent open surgery, had longer operative times, and more frequently needed venous reconstruction and concomitant resection of adjacent organs. They also experienced greater blood loss, received more transfusions, and received a higher volume of fluids. Postoperatively, patients with AKI exhibited higher white blood cell (WBC) count, total bilirubin, SCr, blood urea nitrogen, and PT and had lower albumin (all P < 0.05). For fluid management patients with AKI had higher input (crystalloid + colloid + blood transfusion) on the day of surgery and higher urine output on POD2-3 without a significant difference in fluid input on POD1-3.

AKI and clinical outcomes

According to the KDIGO criteria, the majority of AKI cases (82.6%) were classified as stage I with a typical onset on POD4. Notably, patients who developed AKI had significantly higher odds of experiencing postoperative complications, such as CR-POPF, PPH, abdominal infection, and shock. Consequently, the development of AKI was also associated with prolonged hospitalization, extended stays in the ICU, higher rates of reoperation, and increased in-hospital mortality.

Subgroup analysis

Subgroup analysis (Table 2) further revealed that patients with stage II and stage III AKI had a higher prevalence of hypertension and alcohol consumption compared with those with stage I AKI. These patients also had higher preoperative and postoperative PT and activated partial thromboplastin time (APTT) although only postoperative APTT reached statistical significance. Furthermore, patients with stage II and stage III AKI experienced more postoperative complications including CR-POPF, PPH, delayed gastric emptying, intra-abdominal infection, and shock and had significantly longer hospital stays and ICU admissions as well as higher rates of in-hospital mortality.

Table 2 Subgroup analysis of stage I and stage II/III acute kidney injury, n (%).

Variables	All AKI (n = 230)	Stage I AKI (n = 190)	Stage II and III AKI (n = 40)	P value
Age (years), median (IQR)	65 (57 to 70)	65 (57 to 69)	65 (56 to 72)	0.721
Male	143 (62)	117 (62)	26 (65)	0.690
BMI (kg/m2), median (IQR)	23.02 (20.89 to 25.80)	23.18 (21.08 to 25.83)	22.94 (20.71 to 25.23)	0.644
HTN	90 (39)	68 (36)	22 (55)	0.024
Drinking	63 (27)	47 (25)	16 (40)	0.049
Preoperative variables
PT (seconds), median (IQR)	11.9 (11.4 to 12.6)	11.9 (11.4 to 12.5)	12.2 (11.7 to 12.7)	0.081
APTT (seconds), median (IQR)	27.0 (25.2 to 29.0)	26.8 (25.1 to 28.7)	27.6 (26.3 to 30.2)	0.051
Postoperative variables
PT (seconds), median (IQR)	13.0 (12.3 to 13.8)	12.9 (12.3 to 13.7)	13.1 (12.2 to 14.3)	0.316
APTT (seconds), median (IQR)	27.7 (25.3 to 31.1)	27.5 (25.3 to 30.5)	30.5 (26.3 to 35.6)	0.013
Postoperative complications
CR-POPF	47 (20)	32 (17)	15 (38)	0.003
PPH, grade B and C	59 (26)	43 (23)	16 (40)	0.022
DGE, grade B and C	47 (20)	34 (18)	13 (33)	0.037
Abdominal infection	77 (33)	56 (29)	21 (53)	0.005
Shock	54 (23)	39 (21)	15 (38)	0.021
Clinical outcomes
Length of hospital stay (days), median (IQR)	29 (18 to 44)	27 (18 to 39)	43 (28 to 53)	< 0.001
Length of ICU stay (days), median (IQR)	1 (0 to 3)	1 (0 to 3)	3 (1 to 5)	0.017
In-hospital mortality	12 (5)	6 (3)	6 (15)	< 0.001

AKI: Acute kidney injury; BMI: Body mass index; IQR: Interquartile range; HTN: Hypertension; PT: Prothrombin time; APTT: Activated partial thromboplastin time; CR-POPF: Clinically relevant postoperative pancreatic fistula; DGE: Delayed gastric emptying; PPH: Post-pancreatectomy hemorrhage; ICU: Intensive care unit.

Feature selection

As Boruta and LASSO were applied separately to the training data (Figure 2), eight predictors were identified using both methods: Operative time; Postoperative ICU admission; Surgical approach; Intraoperative red blood cell (RBC) transfusion; Postoperative SCr; Postoperative bilirubin; History of malignancy; And history of stroke. Boruta alone selected 10 additional variables (tumor size, estimated blood loss, age, sex, postoperative PT, postoperative APTT, postoperative albumin, preoperative PT, total input on POD0, and angiotensin-converting enzyme inhibitor/angiotensin receptor blocker), whereas LASSO uniquely identified three more predictors (NSAIDs, postoperative WBC, and preoperative WBC). In total the combination of Boruta and LASSO yielded 21 features for subsequent model development.

Figure 2 Feature selection by Boruta algorithm and least absolute shrinkage and selection operator. A: Through Boruta the 18 filtered variables were as follows: Surgery duration; Intensive care unit admission after surgery; Surgical approach; Intraoperative red blood cell (RBC) transfusion; Tumor length; Postoperative serum creatinine; Postoperative albumin; Intraoperative blood loss; Sex; Postoperative prothrombin time; Age; Postoperative bilirubin; Postoperative activated partial thromboplastin time; Preoperative prothrombin time; Malignancy history; History of stroke; Total input on postoperative day 0; And angiotensin-converting enzyme inhibitor/angiotensin receptor blocker medication; B: Through least absolute shrinkage and selection operator, the 11 filtered variables were as follows: Intensive care unit admission after surgery; Nonsteroidal anti-inflammatory drugs; Malignancy history; History of stroke; Surgery duration; Postoperative white blood cell; Surgical approach; Postoperative serum creatinine; Postoperative bilirubin; Preoperative bilirubin; And intraoperative RBC transfusion. A footnote of 0 (such as variable 0) indicates the preoperative value while a footprint of 1 (such as variable 1) indicates the postoperative value. CVM: Cross-validation mean; Alb: Albumin; APTT: Activated partial thromboplastin time; TBil: Total bilirubin; PT: Prothrombin time; RBC: Red blood cell; ICU: Intensive care unit; SCr: Serum creatinine.

Model development and validation

Seven ML algorithms were developed to predict postoperative AKI. The CatBoost algorithm achieved the highest discrimination [AUROC = 0.803, 95% confidence interval (CI): 0.759-0.848], followed by light gradient boosting machine (AUROC = 0.786, 95%CI: 0.740-0.832), random forest (AUROC = 0.782, 95%CI: 0.735-0.828), and extreme gradient boosting with classification trees (AUROC = 0.781, 95%CI: 0.734-0.827) in the training cohort (Figure 3). The CatBoost model exhibited good calibration as evidenced by a well-aligned calibration curve (Brier score = 0.180, Figure 4), and favorable clinical utility based on the decision curve analysis curve (Figure 5). In the validation cohort CatBoost maintained a strong AUROC of 0.751 (95%CI: 0.652-0.850, Figure 3) and a Brier score of 0.203 (Figure 4). The detailed parameters for model evaluation are summarized in Table 3.

Figure 3 Receiver operating characteristic curves of different models. A: Developing cohort; B: Validation cohort. AUC: Area under the curve; CatBoost: Categorical boosting; KNN: K-nearest neighbor; LightGBM: Light gradient boosting machine; SVM: Support vector machine; XGBoost: Extreme gradient boosting with classification trees.

Figure 4 Calibration curves of different models. A: Developing cohort; B: Validation cohort. CatBoost: Categorical boosting; KNN: K-nearest neighbor; LightGBM: Light gradient boosting machine; SVM: Support vector machine; XGBoost: Extreme gradient boosting with classification trees.

Figure 5 Decision curves of different models. A: Developing cohort; B: Validation cohort. CatBoost: Categorical boosting; KNN: K-nearest neighbor; LightGBM: Light gradient boosting machine; SVM: Support vector machine; XGBoost: Extreme gradient boosting with classification trees.

Table 3 Machine learning evaluation metrics for the training and validation cohorts.

Models	AUC (95%CI)	Accuracy	Sensitivity	Specificity	Brier score
Developing set
CatBoost	0.803 (0.759, 0.848)	0.738	0.780	0.710	0.180
KNN	0.772 (0.725, 0.819)	0.703	0.791	0.609	0.195
LightGBM	0.786 (0.740, 0.832)	0.722	0.764	0.676	0.192
LR	0.776 (0.728, 0.824)	0.724	0.754	0.693	0.189
RF	0.782 (0.735, 0.828)	0.719	0.780	0.654	0.191
SVM	0.765 (0.716, 0.813)	0.700	0.733	0.665	0.197
XGBoost	0.781 (0.734, 0.827)	0.730	0.780	0.676	0.191
Validation set
CatBoost	0.751 (0.652, 0.850)	0.702	0.767	0.647	0.203
KNN	0.655 (0.544, 0.766)	0.596	0.674	0.529	0.232
LightGBM	0.756 (0.657, 0.854)	0.681	0.651	0.706	0.203
LR	0.709 (0.605, 0.812)	0.617	0.651	0.589	0.218
RF	0.712 (0.607, 0.817)	0.628	0.651	0.608	0.216
SVM	0.714 (0.611, 0.817)	0.638	0.674	0.608	0.212
XGBoost	0.730 (0.627, 0.833)	0.681	0.674	0.686	0.210

AUC: Area under the curve; CI: Confidence interval; CatBoost: Categorical boosting; KNN: K-nearest neighbor; LightGBM: Light gradient boosting machine; LR: Logistic regression; RF: Random forest; SVM: Support vector machine; XGBoost: Extreme gradient boosting with classification trees.

Model interpretation

Figure 6A summarizes the SHAP values of the key variables within the CatBoost model. Operative time was identified as the strongest predictor, followed by postoperative SCr, ICU admission after surgery, postoperative WBC, and intraoperative RBC transfusion. Furthermore, the SHAP Beeswarm plot (Figure 6B) quantified the directional impact of variables on AKI risk. Longer operative time, higher postoperative SCr, ICU admission after surgery, higher postoperative WBC, more RBC transfusions, higher postoperative bilirubin, older age, higher PT before and after surgery, lower APTT after surgery, use of NSAIDs, malignancy history, open surgery, higher intravenous fluid intake on POD0, larger tumor size, higher blood loss, and lower albumin were all associated with a higher predicted probability of postoperative AKI. The same trend was observed in the SHAP-dependence plots (Figure 7).

Figure 6 SHapley Additive exPlanations summary plot for the categorical boosting model. A: Average absolute impact of variables on the final model output magnitude ordered by decreasing feature importance; B: Beeswarm plot of the final model. A footnote of 0 (such as variable 0) means the preoperative value while a footprint of 1 (such as variable 1) means the postoperative value. SCr: Serum creatinine; ICU: Intensive care unit; WBC: White blood cell; RBC: Red blood cell; TBil: Total bilirubin; PT: Prothrombin time; APTT: Activated partial thromboplastin time; Alb: Albumin; NSAIDs: Nonsteroidal anti-inflammatory drugs; ACEI: Angiotensin-converting enzyme inhibitors; ARB: Angiotensin receptor blockers; POD: Postoperative day; SHAP: SHapley Additive exPlanations.

Figure 7 SHapley Additive exPlanations dependence plot of the categorical boosting model. Each panel shows that each feature affects the output of the final model. A: Operative time; B: Postoperative serum creatinine; C: Postoperative white blood cell count; D: Intraoperative red blood cell transfusion; E: Postoperative bilirubin; F: Postoperative prothrombin time; G: Postoperative activated partial thromboplastin time; H: Total input on the day of surgery; I: Postoperative albumin. The X-axis represents the raw values of each feature, and the Y-axis indicates the SHapley Additive exPlanations (SHAP) values of the features. When the SHAP value of a specific feature exceeds zero, it indicates an increased risk of acute kidney injury. A footnote of 0 (such as variable 0) means the preoperative value while a footprint of 1 (such as variable 1) means the postoperative value. SHAP: SHapley Additive exPlanations; SCr: Serum creatinine; WBC: White blood cell; RBC: Red blood cell; TBil: Total bilirubin; PT: Prothrombin time; APTT: Activated partial thromboplastin time; Alb: Albumin; POD: Postoperative day.

Web-based calculator

Using the developed CatBoost model, we constructed a web-based risk calculator to estimate the risk of postoperative AKI after pancreatic surgery (https://akiafterpancreaticsurgery.streamlit.app/, Figure 8). We suggest that providers use the first postoperative lab results, which may allow for the early identification of patients at high risk for AKI. This could help guide timely interventions and personalized management strategies to improve postoperative outcomes.

Figure 8 Online acute kidney injury risk calculator. AKI: Acute kidney injury; Alb: Albumin; ICU: Intensive care unit; TBil: Total bilirubin; PT: Prothrombin time; SCr: Serum creatinine; CHD: Coronary heart disease.

DISCUSSION

In this retrospective cohort study involving 4216 patients, we developed and validated an interpretable ML prediction model for AKI after pancreatic surgery. By employing advanced feature selection techniques (Boruta and LASSO), 21 key predictors were identified. Our final CatBoost model demonstrated superior performance in terms of discrimination (AUROC = 0.803 and 0.751 in the training and validation cohorts, respectively), calibration (Brier scores = 0.180 and 0.203, respectively), and clinical utility. The SHAP analysis revealed that operative time, postoperative SCr, ICU admission after surgery, postoperative WBC, and intraoperative RBC transfusions were the most important predictors of AKI. These findings may facilitate the accurate and timely identification of patients at high risk of postoperative AKI following pancreatic surgery.

While previous studies have investigated postoperative AKI following abdominal surgery[18-20], this is the first study to apply ML algorithms for AKI prediction specifically in the context of pancreatic surgery. This is important as pancreatic resections pose unique challenges due to extensive tissue dissection, frequent vascular reconstruction, and significant perioperative fluid shifts that increase renal vulnerability[1,21]. Our observed AKI incidence rate of 5.5% aligns with previous reports and underscores its clinical significance[1,10]. Furthermore, while prior studies have identified risk factors for postoperative AKI after pancreatic surgery[9,10], we applied advanced feature selection and interpretable ML techniques to identify several novel predictors, including prolonged operative time, postoperative ICU admission, postoperative WBC, perioperative coagulopathy, fluid overload, hypoalbuminemia, and chronic NSAIDs use. These findings augment existing literature by providing a more granular, multidimensional risk profile for AKI after pancreatic surgery. Importantly, our model also confirmed some established risk factors, such as older age, cardiovascular disease, increased blood loss, elevated perioperative total bilirubin, increased perioperative SCr, and the specific type of surgery (PD, DP or TP)[10,22].

The association of AKI with longer operative time, open surgery, greater intraoperative blood loss, and higher transfusion requirements suggests that surgical trauma and hemodynamic perturbations are central to its pathophysiology. Hemodynamic instability is a well-established driver of AKI, and even minutes of hypotension can precipitate renal injury[23]. Conversely, reperfusion following periods of ischemia, often exacerbated by RBC transfusions, can trigger a cascade of inflammatory and oxidative processes, inflicting a secondary insult on the renal parenchyma[24-26]. Similarly, surgical trauma can elicit the systemic inflammatory response syndrome characterized by the release of proinflammatory cytokines and activation of immune cells; this is a well-recognized contributor to AKI[27]. The observation that a significantly elevated postoperative WBC count (greater than 15 × 10⁹/L, Figure 7) emerged as an important predictor of AKI in our model further strengthens this point. As a common clinical marker of systemic inflammation and infection, a significantly elevated WBC count can reflect a severe inflammatory state, contributing to renal dysfunction. These findings suggest that strategies aimed at modulating or mitigating the systemic inflammatory response in the early postoperative period, such as prevention and prompt management of infection, glycemic control, and maintenance of normothermia, could potentially confer renoprotective effects.

In our final model early elevations in postoperative SCr emerged as the second most important risk factor for AKI. This finding is consistent with existing literature in which even a modest immediate postoperative rise in SCr has been shown to be a reliable early predictor of AKI. Mechanistically, an early increase in SCr likely indicates that substantial renal insult has already occurred intraoperatively, heralding the onset of clinically apparent AKI.

Unsurprisingly, postoperative ICU admission was a powerful composite predictor in our model. The decision for ICU admission likely serves as a surrogate for underlying physiological frailty and magnitude of surgical stress that may not be fully captured by individual clinical parameters. Furthermore, patients requiring ICU care are often exposed to high-risk interventions, including prolonged mechanical ventilation, administration of high-risk antibiotics (such as vancomycin and aminoglycosides), vasopressors, and frequent contrast-enhanced computed tomography. These are all well-documented contributors to acute tubular necrosis and renal hypoperfusion[28-30].

Our findings also demonstrated that perioperative coagulopathy is a significant contributor to postoperative AKI. Specifically, prolonged PT and shortened APTT were both independent risk factors. Pancreatic and biliary tumors commonly cause obstructive jaundice, which impairs hepatic synthesis of coagulation factors and reduces vitamin K absorption. These disruptions lead to prolonged PT, increasing the risk of perioperative hemorrhage and subsequent AKI[31]. This mechanism is further supported by our observation that elevated postoperative bilirubin and decreased albumin levels were also key predictors of AKI.

Moreover, massive intraoperative bleeding and blood product resuscitation can exacerbate coagulopathy by diluting or depleting clotting factors[32]. In contrast, a shortened postoperative APTT, reflecting a hypercoagulable state, was also associated with an increased risk of AKI. This may be attributed to a tumor-related prothrombotic state and the postoperative inflammatory response that together can promote renal micro-thrombosis through a thrombo-inflammatory pathway[33]. Collectively, our results underscore the dual threat posed by bleeding and hypercoagulability in the perioperative period. This highlights the importance of managing coagulopathy and consideration of preoperative biliary drainage (if indicated) to mitigate the risk of postoperative AKI.

Perioperative fluid management during abdominal surgery remains controversial. Myles et al[34] showed that a restrictive regimen increased postoperative AKI compared with a liberal approach, whereas other studies suggest fluid restriction can decrease complications[35]. In this study SHAP interpretation revealed a moderately positive association between volume administered on the day of surgery and the risk of AKI. This observed association may be confounded by several intraoperative factors, including operative duration, blood loss, and intraoperative hypotension, all of which independently influence fluid requirements. Therefore, volume administered on the day of surgery may act as a proxy variable, reflecting the overall magnitude of surgical trauma and physiological stress. Alternatively, our findings raise the possibility that in patients undergoing shorter and hemodynamically stable procedures, excessive fluid administration may itself be a contributing factor to AKI development. Notably, patients with AKI had higher urine output than patients without AKI over the first 3 PODs, suggesting that the KDIGO oliguria criterion may fail to detect AKI in patients who are volume overloaded after pancreatic surgery.

Lower perioperative serum albumin levels were also independently associated with postoperative AKI in our model. Hypoalbuminemia may not only reflect impaired hepatic synthetic capacity but may also indicate significant intraoperative hemodilution[36]. In addition, exposure to NSAIDs represents a modifiable risk factor. By inhibiting cyclooxygenase-dependent prostaglandin synthesis, NSAIDs attenuate afferent arteriolar vasodilation, reduce renal blood flow, and in susceptible patients provoke direct tubular toxicity or acute interstitial nephritis[37].

Collectively, these findings suggest that targeted postoperative interventions in individuals at high risk after pancreatic surgery may reduce the risk of AKI. Such interventions include prevention and prompt management of infection, correction of coagulopathy, adequate biliary drainage if indicated, early initiation of enteral nutrition, enhanced liver function monitoring, and diligent renal function monitoring.

Several limitations of this study merit careful consideration. First, the single-center retrospective design inherently restricts the distinct generalizability of our findings. As the cohort was derived exclusively from a tertiary center in China, the population is likely ethnically homogeneous; thus, the applicability of our model to diverse genetic or demographic backgrounds remains unverified and warrants cautious interpretation. Second, we acknowledge that the lack of independent external validation poses a potential risk of overfitting, potentially affecting the reproducibility of the identified predictor rankings in other clinical settings. While we aimed to validate our findings externally, the comprehensive nature of our dataset, which incorporated 53 granular perioperative variables to capture the multifactorial etiology of AKI, presented significant challenges in identifying an external cohort with matching data density and definition consistency. To partially mitigate this and rigorously assess model stability, we employed a temporal validation strategy. This approach, which evaluates performance on future unseen data, serves as a rigorous internal proxy for assessing predictive utility. These findings therefore establish a substantial evidential basis for risk identification although future multicenter prospective studies with standardized data collection protocols remain necessary to calibrate the model for broader implementation.

CONCLUSION

This study developed and validated an interpretable ML model that accurately predicts AKI following pancreatic surgery. We identified both established and novel risk factors to create a multidimensional, comprehensive tool for early risk stratification. Ultimately, these predictive insights can facilitate the development of personalized perioperative strategies aimed at preserving renal function and improving outcomes in this high-risk surgical population.