Association of prealbumin with risk of abdominal surgery and endoscopic recurrence in Crohn’s disease: A machine learning study

Abstract

BACKGROUND

Prealbumin (PA) reflects inflammatory and nutritional status and may therefore serve as a pragmatic biomarker for risk stratification and personalized management in Crohn’s disease (CD). However, its associations with abdominal surgical risk and postoperative endoscopic recurrence remain unclear.

AIM

To explore its associations with abdominal surgical risk and postoperative endoscopic recurrence.

METHODS

This retrospective study enrolled 1060 newly diagnosed CD patients and 282 patients who underwent abdominal surgery at our institution. Cox regression, Kaplan-Meier, and restricted cubic spline analyses were used to assess the associations between PA and disease outcomes. Mediation analysis explored the inflammatory, coagulation, and nutritional pathways. Finally, seven machine learning (ML) models were developed to predict disease outcomes, and their predictive performance was compared.

RESULTS

Trend and restricted cubic spline analyses showed that higher PA levels were associated with lower risks of abdominal surgery and endoscopic recurrence [Q4 vs Q1: Surgery, hazard ratio (HR) = 0.45; recurrence, HR = 0.20; both P < 0.001]. Inflection points at 208.149 mg/L and 148.816 mg/L were used to define PA categories, and accordingly, high PA levels were associated with reduced risks of surgery (HR = 0.39) and recurrence (HR = 0.52). PA also showed good discrimination, with a 5-year area under the curve of 0.697 for surgery and a 2-year area under the curve of 0.756 for recurrence. Mediation analysis indicated that fibrinogen and body mass index mediated 3.8% and 8.1% of the associations with surgery and recurrence, respectively, whereas inflammatory markers accounted for about 7% in both settings. Moreover, among the ML models, the PA-based gradient boosting machine model performed best for surgery prediction, while the random survival forest model performed best for recurrence prediction, with C-index values of 0.784 and 0.764, respectively; both also provided clear prognostic stratification (both P < 0.001). Notably, biologic therapy was associated with improved outcomes mainly in patients with low PA levels or those classified as intermediate- or high-risk. Consistently, serum PA levels showed approximately 70% concordance with ML-derived risk stratification. Finally, an online clinical calculator was developed to support treatment decision-making in CD.

CONCLUSION

Serum PA is a strong predictor of abdominal surgery and postoperative endoscopic recurrence in CD. Incorporating PA into ML-based risk stratification may improve early risk prediction and support more individualized treatment decisions.

Key Words: Crohn’s disease; Prealbumin; Abdominal surgery; Postoperative endoscopic recurrence; Machine learning; Predictive and personal medicine

Core Tip: Serum prealbumin (PA), reflecting inflammatory and nutritional status, may serve as a pragmatic biomarker for risk stratification in Crohn’s disease. Higher PA levels were associated with lower risks of abdominal surgery and postoperative endoscopic recurrence (highest vs lowest quartile: Hazard ratio = 0.45 and 0.20, respectively). Incorporating PA into machine learning models further improved prognostic performance (C-index = 0.784 for surgery and 0.764 for recurrence). Moreover, serum PA showed approximately 70% concordance with machine learning-derived risk tiers and helped identify patients more likely to benefit from biologic therapy, thereby supporting earlier risk prediction and more individualized treatment decisions.

INTRODUCTION

Crohn’s disease (CD) is a chronic inflammatory bowel disease with a rising global incidence[1,2]. Despite the widespread use of biologic therapies to treat CD patients, their long-term outcomes remain poor, and many patients still progress to stricturing or penetrating disease and eventually require abdominal surgery[3-5]. However, such surgery is not curative, and postoperative endoscopic recurrence remains common. Importantly, endoscopic recurrence often occurs before clinical relapse and is strongly associated with a subsequent clinical recurrence and the need for repeat surgery[6,7]. Together, these findings highlight the marked heterogeneity of CD across both the natural and postoperative disease course and underscore the need to identify patients at high risk of poor outcomes at an early stage[8-10]. This has also driven CD management towards more individualized and risk-based treatments, albeit the current prognostic assessment still relies mainly on the consideration of limited clinical factors, which often provide only a modest discrimination of cases. As a result, low-risk patients may end up receiving unnecessary intensive treatment, whereas high-risk patients may not receive timely intervention[11,12]. Simple and accessible biomarkers are needed to improve the risk stratification of CD patients and support the more appropriate use of biologic therapies[13].

Prealbumin (PA) is a hepatic transport protein with a short half-life of approximately 2-3 days[14]. Because its synthesis is rapidly suppressed by proinflammatory cytokines, PA reflects the nutrition-inflammation axis and is a sensitive marker of both the nutritional status and systemic inflammation. Low PA levels have been linked to adverse outcomes across a range of medical and surgical settings[15-20]. In CD, chronic inflammation and malnutrition frequently coexist and jointly contribute to disease progression, and lower PA levels have been associated with higher disease activity and postoperative complications[21,22]. However, the prognostic value of PA for predicting the need for long-term abdominal surgery and of postoperative endoscopic recurrence in CD has not yet been systematically evaluated, and its potential nonlinear relationship with these outcomes remains unclear. Conventional models to assess these relationships, such as Cox regression, may not fully capture these complex relationships, whereas machine learning (ML) methods are considered a better option as they are more suited to modelling nonlinear associations and higher-order interactions in multidimensional clinical data[3,23-28].

This study aimed to evaluate the prognostic value of PA in CD, examine its associations with needing abdominal surgery and postoperative endoscopic recurrence, and assess whether incorporating PA into ML-based survival models could improve the risk stratification of CD patients and support better treatment decision-making.

MATERIALS AND METHODS

Study population

Patients with CD treated at our hospital between 2016 and 2022 were enrolled in the study as the initial diagnosis cohort, and patients who underwent their first abdominal surgery at our hospital between 2016 and 2024 were included in the initial surgical cohort. The surgical cohort included both patients diagnosed and operated on at our hospital and those initially diagnosed elsewhere but referred to our centre for their first surgery. CD was diagnosed according to current clinical guidelines or as confirmed by endoscopic, radiological, or histopathological findings, as appropriate.

For the initial diagnosis cohort, the patients’ baseline data were extracted from the hospital’s electronic medical record system and included their sex, smoking history, body mass index (BMI), white blood cell count (WBC), C-reactive protein (CRP), haemoglobin (Hb), erythrocyte sedimentation rate (ESR), PA, albumin (ALB), prothrombin time, fibrinogen (Fg), D-dimer, Montreal classification, medication history, and CD Activity Index (CDAI) at diagnosis. Biologic therapy was defined as treatment with infliximab, adalimumab, vedolizumab, ustekinumab, or upadacitinib. Patients were required to have had their first diagnosis at The Second Xiangya Hospital, Central South University, but were excluded if they had undergone previous abdominal or intestinal surgery other than appendectomy, their records had more than 20% missing key clinical data or lacked follow-up data, or if they had died from unrelated causes.

For the initial surgical cohort, the patients’ baseline data were obtained from perioperative hospital records and included their demographic characteristics, smoking history, BMI, Montreal classification, preoperative laboratory parameters, preoperative medication exposure, lesion location, and postoperative biologic use. Preoperative laboratory parameters included WBC, CRP, Hb, ESR, ALB, PA, and coagulation indices. Eligible patients were those aged > 18 years old, who underwent their first CD-related abdominal intestinal resection at our hospital, and who had at least one postoperative follow-up colonoscopy with complete endoscopic records for outcome assessment. Patients with concomitant malignancy, moderate-to-severe renal insufficiency, or more than 20% missing key clinical data were excluded.

Baseline data and follow-up information

All the data were independently collected by two investigators, with interobserver agreement assessed using the kappa statistic, with κ = 0.85 indicating good consistency between the two investigators. Any discrepancies were resolved by a third investigator. Information on surgical treatment was obtained from the patients’ electronic medical records and by telephone follow-up.

The primary endpoint was first abdominal surgery. Indications for surgery included recurrent intestinal bleeding, acute perforation, chronic intestinal fistula, recurrent bowel obstruction, malignancy, treatment failure, an uncertain diagnosis, and growth retardation in children. Follow-up was conducted every 3 months, calculated from the time of first diagnosis to the date of surgery or the last follow-up visit.

The secondary endpoint was postoperative endoscopic recurrence. According to routine follow-up practice, patients were generally advised to undergo colonoscopy at around 6 months after surgery; if this was not performed, the most recent colonoscopy within 1 year after surgery was used for assessment. Endoscopic recurrence was evaluated using the Rutgeerts score, which grades mucosal lesions at the neoterminal ileum and anastomosis. In this study, recurrence was defined as a Rutgeerts score of i2 or higher. Scores of i0-i1, including ≤ 5 isolated aphthous ulcers, were classified as no recurrence, while i2-i4 indicated endoscopic recurrence. The end of follow-up was defined as the date of the last colonoscopy or the last clinical visit. This study was approved by the Ethics Committee of The Second Xiangya Hospital, Central South University, No. LYEC2025-0156 and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived as this was a retrospective observational study based on anonymized data and involved no intervention. Clinical trial registration was therefore not applicable either.

Statistical analysis

Baseline data description: Missing clinical data were handled using multiple imputation with the mice package in R, generating five imputed datasets by predictive mean matching. The pooled dataset was then used for all the subsequent analyses. Normality of the continuous variables was assessed using the Shapiro-Wilk test. Normally distributed variables are presented herein as the mean ± SD and were compared using the independent-samples t test, while non-normally distributed variables are presented as the median (interquartile range) and were compared using the Mann-Whitney U test. Categorical variables are expressed as a n (%) and were compared using the χ² test or Fisher’s exact test, as appropriate.

Survival analysis: Kaplan-Meier curves were used to estimate surgery-free survival and recurrence-free survival, and differences between groups were compared using the log-rank test. Associations between candidate predictors and clinical outcomes were evaluated using Cox proportional hazards regression, with the results reported as hazard ratios (HRs) and 95% confidence intervals (CIs). The Cox model is defined as: h(t|X) = h₀(t)exp(βX), where h(t|X) is the hazard function conditioned on the covariates, h₀(t) is the baseline hazard function representing the hazard without covariates, β is the regression coefficient vector, and X is the vector of the covariates. To explore potential nonlinear associations between prognostic variables and the surgical risk, restricted cubic splines with four knots placed at the 5^th, 35^th, 65^th, and 95^th percentiles were incorporated into the multivariable Cox model. Nonlinearity was assessed using the Wald test.

ML survival model construction

The overall study cohort was randomly divided into a training set and a validation set in a 7:3 ratio using a fixed random seed. Variables with P < 0.10 in the univariable Cox regression were entered into the multivariable analysis, and the final predictors were selected using backward stepwise regression.

In addition to the Cox proportional hazards regression model, to better capture nonlinear associations and higher-order interactions, we also developed seven ML survival models: Random survival forest (RSF), gradient boosting machine (GBM), CoxBoost, survival support vector machine, XGBoost, SuperPC, and PLSR-Cox. Before model development, continuous variables were standardized using Z-score normalization. Hyperparameters for each model were optimized in a training set through grid search with 5-fold cross-validation, and the final hyperparameter settings were determined according to the mean Harrell’s concordance index (C-index) across the five validation folds.

Model performance was primarily evaluated using Harrell’s C-index, while the time-dependent area under the receiver operating characteristic curve (AUC) was used to assess discrimination at different time points. The net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were further calculated to quantify the incremental predictive value of considering the PA level when added to the model. Calibration curves were used to assess agreement between the predicted and observed outcomes, while decision curve analysis was performed to evaluate clinical utility. Differences in C-index between the models were further compared. SHapley Additive exPlanations (SHAP) was applied to interpret the contribution of individual predictors in the ML models. Finally, to facilitate clinical application, a web-based prediction platform was developed and is available at: https://crohndisease.shinyapps.io/cdpasurgeryandrecurrence/. All the statistical analyses were performed using R, SPSS, Python, and GraphPad Prism. A two-sided P < 0.05 was considered statistically significant.

RESULTS

Baseline characteristics

In total, 1060 newly diagnosed CD patients and 282 patients undergoing their first CD-related surgery were enrolled in this study (Figure 1). In the newly diagnosed CD cohort, the majority of patients were male (786, 74.2%), with 274 females (25.8%). A history of smoking was reported in 216 patients (20.4%), and 373 (35.2%) had concomitant perianal disease. According to the Montreal classification, the patients’ ages at diagnosis were distributed as follows: A1 in 133 patients (12.5%), A2 in 788 (74.3%), and A3 in 139 (13.1%); disease behaviour was classified as B1 in 473 patients (44.6%), B2 in 346 (32.6%), and B3 in 241 (22.7%); disease location included L1 in 268 patients (25.3%), L2 in 67 (6.3%), L3 in 771 (67.1%), and L4 in 14 (1.3%). Regarding initial therapy, 95 patients (9.0%) received 5-aminosalicylic acid agents, 152 (14.3%) corticosteroids, 261 (24.6%) immunosuppressants, and 667 (62.9%) biologics (Table 1).

Figure 1 Flowchart for patient selection and data analysis. CD: Crohn’s disease; ROC: Receiver operating characteristic; GBM: Gradient boosting machine; RSF: Random survival forest; SVM: Support vector machine; AUC: Area under the receiver operating characteristic curve; SHAP: SHapley Additive exPlanations; PLSR-Cox: Partial least squares regression for Cox models.

Table 1 Baseline characteristics of the cohort of newly diagnosed Crohn’s disease patients, n (%)/median (interquartile range).

Characteristics	Total cohort	Q1	Q2	Q3	Q4	P value
	1060	265	265	265	265
Demographics						< 0.001
Gender						0.007
Female	274 (25.8)	86 (32.5)	74 (27.9)	59 (22.3)	55 (20.8)
Male	786 (74.2)	179 (67.5)	191 (72.1)	206 (77.7)	210 (79.2)
Smoking						0.674
Yes	216 (20.4)	56 (21.1)	47 (17.7)	56 (21.1)	57 (21.5)
No	844 (79.6)	209 (78.9)	218 (82.3)	209 (78.9)	208 (78.5)
BMI (kg/m2)	18.5 (16.7, 20.9)	18.4 (16.6, 20.8)	18.4 (16.8, 20.9)	18.4 (17.0, 20.8)	18.7 (16.7, 21.1)	0.982
Laboratory tests
WBC (109)	6.44 (5.04, 8.21)	7.07 (5.39, 9.43)	7.35 (5.57, 9.3)	5.91 (4.7, 7.23)	5.92 (4.82, 7.41)	< 0.001
NLR	1.51 (0.8, 3.02)	2.53 (1.32, 5.59)	1.56 (0.80, 3.24)	1.18 (0.64, 2.33)	1.10 (0.68, 2.35)	< 0.001
LMR	4.16 (2.77, 6.55)	4.13 (2.84, 6.59)	4.27 (2.91, 6.35)	4.16 (2.99, 6.18)	4.12 (2.50, 6.66)	0.96
CRP (mg/dL)	1.44 (0.46, 3.94)	5.90 (2.15, 8.91)	2.11 (0.83, 4.32)	0.82 (0.30, 1.68)	0.59 (0.24, 1.32)	< 0.001
Hb (g/dL)	12.1 (10.3, 13.6)	10.3 (8.6, 12.1)	11.8 (10.4, 13.2)	12.4 (10.8, 13.9)	13.5 (11.8, 14.7)	< 0.001
ESR (mm/hour)	21 (9, 43)	44 (26, 59)	25 (13, 44)	14 (7, 30)	11 (6, 21)	< 0.001
PA (mg/L)	194.4 (139.6, 243.2)	101.4 (77.5, 126.2)	168.9 (152.8, 180.5)	215.5 (209.1, 227.1)	264.6 (252.4, 279.9)	< 0.001
ALB (g/dL)	3.7 (3.22, 4.13)	2.53 (1.32, 5.59)	1.56 (0.80, 3.24)	1.18 (0.64, 2.33)	1.10 (0.68, 2.35)	< 0.001
PT (seconds)	12 (11.1, 13)	12.7 (11.7, 13.7)	12.1 (11.4, 13)	11.6 (10.9, 12.3)	11.6 (11, 12.3)	< 0.001
Fg (g/L)	3.7 (2.9, 4.6)	4.31 (3.4, 5.03)	3.8 (3.2, 4.8)	3.31 (2.6, 4.05)	3.2 (2.4, 3.9)	< 0.001
D-dimer (mg/L)	0.33 (0.19, 0.64)	0.57 (0.34, 1.00)	0.34 (0.20, 0.64)	0.28 (0.17, 0.54)	0.26 (0.15, 0.45)	< 0.001
Crohn’s disease related
Montreal classification A						0.86
A1	133 (12.5)	35 (13.2)	34 (12.8)	34 (12.8)	30 (11.3)
A2	788 (74.3)	189 (71.3)	200 (75.5)	197 (74.3)	202 (76.2)
A3	139 (13.1)	41 (15.5)	31 (11.7)	34 (12.8)	33 (12.5)
Montreal classification B						0.074
B1	473 (44.6)	124 (46.8)	123 (46.4)	129 (48.7)	97 (36.6)
B2	346 (32.6)	84 (31.7)	90 (34.0)	78 (29.4)	94 (35.5)
B3	241 (22.7)	57 (21.5)	52 (19.6)	58 (21.9)	74 (27.9)
Montreal classification L						< 0.001
L1	268 (25.3)	55 (20.8)	57 (21.5)	78 (29.4)	78 (29.4)
L2	67 (6.3)	23 (8.7)	28 (10.6)	2 (0.8)	14 (5.3)
L3	711 (67.1)	183 (69.1)	176 (66.4)	182 (68.7)	170 (64.2)
L4	14 (1.3)	4 (1.5)	4 (1.5)	3 (1.1)	3 (1.1)
CDAI						0.205
Remission	133 (12.5)	32 (12.1)	34 (12.8)	39 (14.7)	28 (10.6)
Mild activity	482 (45.5)	113 (42.6)	120 (45.3)	122 (46.0)	127 (47.9)
Moderate activity	392 (37.0)	101 (38.1)	104 (39.2)	95 (35.8)	92 (34.7)
Severe activity	53 (5.0)	19 (7.2)	7 (2.6)	9 (3.4)	18 (6.8)
Perianal abscess						0.336
No	687 (64.8)	184 (69.4)	168 (63.4)	169 (63.8)	166 (62.6)
Yes	373 (35.2)	81 (30.6)	97 (36.6)	96 (36.2)	99 (37.4)
Use of drugs
5-aminosalicylic acid	90 (9.0)	23 (8.7)	18 (6.8)	28 (10.6)	21 (7.9)	0.241
Steroid	152 (14.3)	34 (12.8)	44 (16.6)	34 (12.8)	40 (15.1)	0.365
Immunosuppressant	261 (24.6)	55 (20.8)	70 (26.4)	64 (24.2)	72 (27.2)	0.115
Biological agent	667 (62.9)	153 (57.7)	159 (60.0)	179 (67.5)	176 (66.4)	0.049

BMI: Body mass index; WBC: White blood cells; NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; CRP: C-reactive protein; Hb: Haemoglobin; ESR: Erythrocyte sedimentation rate; PA: Prealbumin; ALB: Albumin; PT: Prothrombin time; Fg: Fibrinogen; CDAI: Crohn’s Disease Activity Index.

The surgical cohort demonstrated notably different clinical characteristics to the newly diagnosed CD cohort. Of the 282 patients in this cohort, 220 (78.0%) were male and 62 (22.0%) were female. Ninety-three (33.0%) had a smoking history, and 42 (14.9%) presented with perianal disease. Montreal classification by age was A1 in 20 patients (7.1%), A2 in 197 (69.9%), and A3 in 65 (23.0%); disease behaviour was B1 in 16 patients (5.7%), B2 in 113 (40.1%), and B3 in 153 (54.3%); Disease distribution was L1 in 87 patients (30.9%), L2 in 16 (5.7%), L3 in 159 (56.4%), and L4 in 20 (7.1%). Preoperative medical treatment included 5-aminosalicylic acid agents in 83 patients (29.4%), corticosteroids in 78 (27.7%), immunosuppressants in 105 (37.2%), and biologics in 103 (36.5%) (Table 2).

Table 2 Baseline characteristics of the primary Crohn’s disease surgical cohort, n (%)/median (interquartile range).

Characteristics	Total cohort	Q1	Q2	Q3	Q4	P value
	282	71	71	70	70
Demographics
Gender						0.069
Female	62 (22.0)	19 (26.8)	20 (28.2)	15 (21.4)	8 (11.4)
Male	220 (78.0)	52 (73.2)	51 (71.8)	55 (78.6)	62 (88.6)
Smoking						0.977
Yes	93 (33.0)	24 (33.8)	22 (31.0)	24 (34.3)	23 (32.9)
No	189 (67.0)	47 (66.2)	49 (69.0)	46 (65.7)	47 (67.1)
BMI (kg/m2)	17.9 (15.9, 20.1)	17.0 (15.1, 19.1)	17.8 (15.9, 20.4)	18.0 (16.9, 20.4)	17.8 (16.0, 20.1)	0.155
Laboratory tests
WBC (109)	5.86 (4.27, 7.78)	6.40 (4.46, 9.02)	5.29 (4.14, 8.38)	6.19 (4.77, 7.50)	5.40 (4.21, 7.06)	0.173
NLR	3.73 (2.41, 6.10)	5.26 (3.54, 10.2)	3.88 (2.17, 6.60)	3.18 (2.45, 4.68)	3.34 (2.20, 5.07)	< 0.001
LMR	2.65 (1.72, 4.20)	1.81 (1.35, 3.27)	2.36 (1.58, 3.81)	3.11 (2.41, 4.72)	3.31 (2.18, 4.53)	< 0.001
CRP (mg/dL)	2.02 (0.77, 6.76)	3.58 (1.93, 9.13)	2.21 (1.10, 7.52)	1.20 (0.48, 2.68)	1.08 (0.43, 3.05)	< 0.001
Hb (g/dL)	115 (98, 132)	99 (84, 120)	110 (93, 121.5)	126 (113, 135.8)	127.5 (113, 139)	< 0.001
ESR (mm/hour)	21.5 (11, 41)	38 (20.5, 62)	31 (17, 49)	16 (9, 26.5)	12 (7, 27.75)	< 0.001
PA (mg/L)	164.7 (108.7, 217.7)	76.5 (58.8, 96.8)	137.7 (128, 152.8)	194.2 (175.27, 206.8)	247.4 (222.9, 262.3)	< 0.001
ALB (g/dL)	35 (30.1, 39.4)	30.1 (26.5, 36.5)	33.9 (30.4, 37.5)	36.6 (33.5, 40.1)	37.3 (32.8, 41.5)	< 0.001
Fg (g/L)	2.95 (2.79, 3.52)	3.24 (2.86, 3.83)	3.01 (2.81, 3.76)	2.88 (2.74, 3.20)	2.91 (2.71, 3.18)	< 0.001
D-dimer (mg/L)	0.51 (0.26, 1.08)	0.75 (0.35, 1.98)	0.81 (0.41, 1.39)	0.34 (0.20, 0.64)	0.33 (0.21, 0.62)	< 0.001
Crohn’s disease related
Montreal A classification						0.140
1	20 (7.1)	3 (4.2)	3 (4.2)	9 (12.9)	5 (7.1)
2	197 (69.9)	55 (77.5)	49 (69.0)	49 (70.0)	44 (62.9)
3	65 (23.0)	13 (18.3)	19 (26.8)	12 (17.1)	21 (30.0)
Montreal B classification						0.053
1	16 (5.7)	4 (5.6)	7 (9.9)	1 (1.4)	4 (5.7)
2	113 (40.1)	24 (33.8)	21 (29.6)	37 (52.9)	31 (44.3)
3	153 (54.3)	43 (60.6)	43 (60.6)	32 (45.7)	35 (50.0)
Montreal L classification						0.668
1	87 (30.9)	18 (25.4)	20 (28.2)	19 (27.1)	30 (42.9)
2	16 (5.7)	5 (7.0)	4 (5.6)	4 (5.7)	3 (4.3)
3	159 (56.4)	42 (59.2)	42 (59.2)	42 (60.0)	33 (47.1)
4	20 (7.1)	6 (8.5)	5 (7.0)	5 (7.1)	4 (5.7)
CDAI						0.011
Remission	74 (26.2)	9 (12.7)	20 (28.2)	19 (27.1)	26 (37.1)
Active	208 (73.8)	62 (87.3)	51 (71.8)	51 (72.9)	44 (62.9)
Perianal abscess						0.085
No	240 (85.1)	54 (76.1)	61 (85.9)	62 (88.6)	63 (90.0)
Yes	42 (14.9)	17 (23.9)	10 (14.1)	8 (11.4)	7 (10.0)
Preoperative use of drugs
5-aminosalicylic acid	83 (29.4)	20 (28.2)	17 (23.9)	26 (37.1)	20 (28.6)	0.282
Steroid	78 (27.7)	18 (25.4)	16 (22.5)	24 (34.3)	20 (28.6)	0.425
Immunosuppressant	105 (37.2)	24 (33.8)	30 (42.3)	28 (40.0)	23 (32.9)	0.624
Biological agent	103 (36.5)	26 (36.6)	23 (32.4)	25 (35.7)	29 (41.4)	0.737

BMI: Body mass index; WBC: White blood cells; NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; CRP: C-reactive protein; Hb: Haemoglobin; ESR: Erythrocyte sedimentation rate; PA: Prealbumin; ALB: Albumin; PT: Prothrombin time; Fg: Fibrinogen; CDAI: Crohn’s Disease Activity Index.

The patients were stratified into quartiles based on their PA levels. In the newly diagnosed CD cohort, the higher PA level group exhibited the following characteristics: A greater proportion of males, higher rate of biological therapy, more frequent L1 disease localization and less L2 involvement, lower inflammatory markers (including reduced WBC, neutrophil-to-lymphocyte ratio, CRP, and ESR levels), better nutritional status (reflected by higher ALB and Hb levels), and improved coagulation profiles (evidenced by lower prothrombin time, Fg, and D-dimer levels). In the surgical CD cohort, the higher PA level group was associated with lower inflammatory markers, better nutritional status, and improved coagulation parameters, as well as higher CDAI scores (Tables 1 and 2).

Associations between PA and surgery risk and recurrence risk

Predictive ability of the prognostic outcomes based on the PA level: Over a median follow-up of 43.8 months, 218 patients (20.6%) in the newly diagnosed cohort underwent abdominal surgery. Kaplan-Meier analysis revealed surgery-free survival rates of 89.3% at 1 year, 82.5% at 3 years, and 75.1% at 5 years (Figure 2A). In the surgical cohort, following a median follow-up time of 17 months after surgery, 142 patients (50.4%) developed endoscopic recurrence. The recurrence-free survival rates were 77.1% at 0.5 years, 61.2% at 1 year, and 28.3% at 2 years (Figure 2B). The patients in the highest PA quartile had significantly better surgery-free and recurrence-free survival rates than those in the lowest quartile (both P < 0.05; Figure 2C and D).

Figure 2 Kaplan-Meier and time-dependent receiver operating characteristic analyses of prealbumin for predicting abdominal surgery and postoperative endoscopic recurrence. A: Kaplan-Meier curve for surgery-free survival in the newly diagnosed cohort; B: Kaplan-Meier curve for recurrence-free survival in the surgical cohort; C and D: Kaplan-Meier curves for surgery-free survival (C) and recurrence-free survival (D) stratified by baseline prealbumin quartiles; E and F: Time-dependent receiver operating characteristic curves of prealbumin for predicting abdominal surgery (E) and postoperative endoscopic recurrence (F). P values were calculated using the log-rank test. AUC: Area under the receiver operating characteristic curve; PA: Prealbumin.

Time-dependent receiver operating characteristic analyses further supported the predictive value of PA for assessing the risk of needing abdominal surgery and of postoperative endoscopic recurrence. Regarding the need for abdominal surgery, the AUCs were 0.701 (95%CI: 0.649-0.752) at 1 year, 0.735 (0.692-0.779) at 3 years, and 0.697 (0.639-0.755) at 5 years (Figure 2E). Regarding postoperative endoscopic recurrence, the corresponding AUCs were 0.580 (0.491-0.669) at 0.5 years, 0.667 (0.583-0.751) at 1 year, and 0.756 (0.665-0.848) at 2 years (Figure 2F).

To assess whether PA was independently associated with long-term outcomes, we constructed six Cox proportional hazards models. Models 1 and 4 were unadjusted; while models 2 and 5 were adjusted for sex, age, BMI, and smoking history; and models 3 and 6 were further adjusted for variables with P < 0.1 in the univariable analyses (Supplementary Tables 1 and 2). In the fully adjusted models, compared with the lowest PA quartile, the highest quartile was associated with a markedly lower risk of needing abdominal surgery (HR = 0.20, 95%CI: 0.13-0.33; P < 0.001; P for trend < 0.001) and of postoperative endoscopic recurrence (HR = 0.45, 95%CI: 0.26-0.80; P = 0.007; P for trend < 0.001; Table 3).

Table 3 Quantile regression analysis of the association between prealbumin and the risks of needing surgery and endoscopic recurrence.

	Q1	Q2, HR (95%CI) P value	Q3, HR (95%CI) P value	Q4, HR (95%CI) P value	P for trend
Surgery
Model 11	Recurrence	0.45 (0.32-0.64), < 0.001	0.33 (0.23-0.48), < 0.001	0.25 (0.17-0.37), < 0.001	< 0.001
Model 22	Recurrence	0.45 (0.32-0.63), < 0.001	0.32 (0.22-0.46), < 0.001	0.24 (0.16-0.35), < 0.001	< 0.001
Model 33	Recurrence	0.50 (0.35-0.73), < 0.001	0.31 (0.20-0.47), < 0.001	0.20 (0.13-0.33), < 0.001	< 0.001
Recurrence
Model 44	Recurrence	0.93 (0.60-1.45), 0.747	0.55 (0.34-0.89), 0.014	0.38 (0.23-0.61), < 0.001	< 0.001
Model 55	Recurrence	0.94 (0.60-1.48), 0.789	0.49 (0.30-0.81), 0.005	0.34 (0.21-0.56), < 0.001	< 0.001
Model 66	Recurrence	1.25 (0.75-2.07), 0.389	0.59 (0.33-1.05), 0.074	0.45 (0.26-0.80), 0.007	< 0.001

¹Model 1 was the crude model.

²Model 2 was adjusted for gender, body mass index, and smoking history.

³Model 3 was further adjusted for gender, body mass index, smoking history, neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, C-reactive protein, haemoglobin, erythrocyte sedimentation rate, albumin, prothrombin time, fibrinogen, D-dimer, the Montreal A and B classifications, Crohn’s Disease Activity Index, perianal abscess, and the biological therapy.

⁴Model 4 was the crude model.

⁵Model 5 was adjusted for gender, age, body mass index, and smoking history.

⁶Model 6 was further adjusted for gender, age, body mass index, smoking history, white blood cells, neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, albumin, fibrinogen, Crohn’s Disease Activity Index, perianal abscess, and preoperative-biologics.

HR: Hazard ratio; CI: Confidence interval.

Dose-response relationship between PA and prognostic outcomes: Restricted cubic spline analysis was performed to examine the dose-response associations of PA with abdominal surgery and postoperative endoscopic recurrence. After adjustment for the demographic covariates and variables with P < 0.1 in the univariable analyses, PA showed an apparent J-shaped association with abdominal surgery, with an inflection point at 208.149 mg/L (P for nonlinearity = 0.164; Figure 3A). By contrast, PA showed a significant nonlinear association with postoperative endoscopic recurrence, with an inflection point at 148.816 mg/L (P for nonlinearity < 0.001; Figure 3B). Below 208.149 mg/L, the surgical risk decreased with an increasing PA level and then plateaued, whereas the recurrence risk declined more clearly above 148.816 mg/L. Segmented regression analyses based on these thresholds yielded consistent results (Table 4).

Figure 3 Restricted cubic spline analyses of the associations between prealbumin and clinical outcomes. A: Association between prealbumin and the risk of abdominal surgery; B: Association between prealbumin and the risk of postoperative endoscopic recurrence. The solid lines represent the estimated associations, and the dashed lines indicate the 95% confidence intervals. HR: Hazard ratio; CI: Confidence interval; PA: Prealbumin.

Table 4 Association between prealbumin levels and the risks of needing surgery and endoscopic recurrence based on segmented regression analysis.

Model	Surgery, HR (95%CI), P value		Model	Recurrence, HR (95%CI), P value
	PA ≤ 208.149 mg/L	PA > 208.149 mg/L		PA ≤ 148.816 mg/L	PA > 148.816 mg/L
Model 11	0.989 (0.986-0.993), < 0.001	0.997 (0.989-1.004), 0.380	Model 44	0.999 (0.993-1.006), 0.849	0.993 (0.987-0.999), 0.035
Model 22	0.989 (0.986-0.993), < 0.001	0.997 (0.989-1.004), 0.383	Model 55	0.999 (0.993-1.005), 0.773	0.992 (0.996-0.999), 0.016
Model 33	0.991 (0.987-0.995), < 0.001	0.997 (0.989-1.005), 0.483	Model 66	1.002 (0.994-1.010), 0.593	0.988 (0.981-0.996), 0.001

¹Model 1 was the crude model.

²Model 2 was adjusted for gender, body mass index, and smoking history.

³Model 3 was further adjusted for gender, body mass index, smoking history, neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, C-reactive protein, haemoglobin, erythrocyte sedimentation rate, albumin, prothrombin time, fibrinogen, D-dimer, the Montreal A and B classifications, Crohn’s Disease Activity Index, perianal abscess, and the biological therapy.

⁴Model 4 was the crude model.

⁵Model 5 was adjusted for gender, age, body mass index, and smoking history.

⁶Model 6 was further adjusted for gender, age, body mass index, smoking history, white blood cells, neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, albumin, fibrinogen, Crohn’s Disease Activity Index, perianal abscess, and preoperative-biologics.

PA: Prealbumin; HR: Hazard ratio; CI: Confidence interval.

When PA was dichotomized at these inflection points, patients with PA > 208.149 mg/L had a significantly lower risk of abdominal surgery than those with PA ≤ 208.149 mg/L (HR = 0.39, 95%CI: 0.27-0.55; P < 0.001; Figure 4A). Similarly, patients with PA > 148.816 mg/L had a lower risk of postoperative endoscopic recurrence than those with PA ≤ 148.816 mg/L (HR = 0.52, 95%CI: 0.33-0.83; P < 0.001; Figure 4B). Significant interactions were also observed between the PA level and biologic therapy. Among patients with PA ≤ 208.149 mg/L, biologic therapy was associated with a lower risk of abdominal surgery (HR = 0.66, 95%CI: 0.48-0.91; P = 0.011; Figure 4C), whereas no significant association was observed in those with PA > 208.149 mg/L (HR = 0.77, 95%CI: 0.46-1.27; P = 0.303; Figure 4D). Likewise, intensified biologic therapy was associated with a lower risk of postoperative endoscopic recurrence in patients with PA ≤ 148.816 mg/L (HR = 0.40, 95%CI: 0.24-0.65; P < 0.001; Figure 4E), but not in those with PA > 148.816 mg/L (HR = 0.64, 95%CI: 0.40-1.03; P = 0.066; Figure 4F). For postoperative endoscopic recurrence, intensified biologic therapy significantly reduced recurrence risk in patients with PA ≤ 148.816 mg/L (HR = 0.40, 95%CI: 0.24-0.65, P < 0.001; Figure 4E), but did not achieve statistical significance in patients with PA > 148.816 mg/L (HR = 0.64, 95%CI: 0.40-1.03, P = 0.066; Figure 4F).

Figure 4 Clinical outcomes stratified by prealbumin thresholds. A and B: Kaplan-Meier analyses of surgery-free survival according to the prealbumin (PA) threshold for abdominal surgery (A) and recurrence-free survival according to the PA threshold for postoperative endoscopic recurrence (B); C-F: Subgroup analyses of biologic therapy showing surgery-free survival in patients with PA ≤ 208.149 mg/L (C) and > 208.149 mg/L (D), and recurrence-free survival in patients with PA ≤ 148.816 mg/L (E) and > 148.816 mg/L (F). HR: Hazard ratio; PA: Prealbumin.

Analysis of the mediating effects

To further explore potential mediators in the relationship between PA and the clinical outcomes in CD, we constructed a composite inflammation score by summing the Z-scores of CRP, ESR, and WBC to provide an integrated measure of systemic inflammation (Supplementary Figure 1). Mediation analysis showed that both the composite inflammation score and Fg significantly mediated the association between PA and the long-term surgical risk. The inflammation score accounted for 7.4% of the total effect (P = 0.040, Supplementary Figure 1A), while Fg mediated 3.8% (P = 0.038, Supplementary Figure 1F). In the pathway from PA to the endoscopic recurrence risk, the inflammation score and BMI were identified as significant mediators, explaining 7.13% (P < 0.001, Supplementary Figure 1G) and 8.08% (P = 0.040, Supplementary Figure 1I) of the total effect, respectively.

Predictive models for assessing the surgery and recurrence risks

Selection of the prognostic variables using backward stepwise Cox regression: The cohort was randomly divided into training and validation sets, with no significant baseline differences between them (Supplementary Tables 3 and 4). Univariable Cox regression was first performed, and variables with P < 0.1 were entered into multivariable Cox models followed by backward stepwise selection (Supplementary Tables 5 and 6). Regarding the risk of needing abdominal surgery, the final model included CRP (HR = 1.05, 95%CI: 1.01-1.09; P = 0.010), lymphocyte-to-monocyte ratio (LMR, HR = 0.94, 95%CI: 0.89-0.99; P = 0.038), Fg (HR = 0.70, 95%CI: 0.61-0.81; P < 0.001), PA (HR = 0.991, 95%CI: 0.988-0.994; P < 0.001), Montreal B classification (HR = 1.26, 95%CI: 1.03-1.54; P = 0.022), and CDAI (HR = 1.30, 95%CI: 1.06-1.60; P = 0.014) (Supplementary Table 5). Regarding postoperative endoscopic recurrence, the final model included Fg (HR = 0.62, 95%CI: 0.46-0.82; P < 0.001), PA (HR = 0.996, 95%CI: 0.992-0.999; P = 0.016), ALB (HR = 0.966, 95%CI: 0.934-0.999; P = 0.041), Montreal B classification (HR = 1.35, 95%CI: 0.96-1.91; P = 0.088), anal fistula (HR = 1.72, 95%CI: 1.08-2.76; P = 0.023), and preoperative biologic therapy (HR = 0.44, 95%CI: 0.27-0.70; P < 0.001) (Supplementary Table 6).

Model development and performance evaluation: Using variables selected from the Cox models, we developed seven ML survival models for predicting the risk of needing abdominal surgery and of postoperative endoscopic recurrence. Regarding the risk of needing abdominal surgery, the GBM model showed the best overall performance, with C-indices of 0.820 (95%CI: 0.788-0.853) in the training cohort and 0.784 (95%CI: 0.728-0.839) in the validation cohort (Figure 5A). It also showed the best discrimination over time (Supplementary Figure 2), the lowest 5-year Brier scores (training 0.111; validation 0.151) (Supplementary Figure 3), and the greatest net benefit across threshold probabilities of 10%-90% (Supplementary Figure 4). Regarding postoperative endoscopic recurrence, the RSF model performed best, with C-indices of 0.796 (95%CI: 0.744-0.847) in the training cohort and 0.764 (95%CI: 0.728-0.839) in the validation cohort (Figure 5B). It also achieved the best time-dependent AUCs (Supplementary Figure 5), most favourable calibration at 2 years (Brier score 0.072 in training and 0.116 in validation) (Supplementary Figure 6), and the highest net benefit across threshold probabilities of 20%-80% (Supplementary Figure 7). These findings identified GBM and RSF as the optimal models for predicting the need for abdominal surgery and postoperative endoscopic recurrence, respectively.

Figure 5 Comparison of the predictive performance of the eight models for abdominal surgery and postoperative endoscopic recurrence. A: Harrell’s concordance index (C-index) of the eight models for predicting abdominal surgery in the training and validation cohorts; B: C-index of the eight models for predicting postoperative endoscopic recurrence in the training and validation cohorts. The evaluated models were Cox proportional hazards regression, random survival forest, gradient boosting machine, CoxBoost, survival support vector machine, XGBoost, SuperPC, and partial least squares regression for Cox models. RSF: Random survival forest; GBM: Gradient boosting machine; PLSR-Cox: Partial least squares regression for Cox models; SVM: Support vector machine.

SHAP analyses: In the surgical prediction model, PA showed the greatest contribution to the outcomes, followed by Fg and the Montreal B classification, whereas CRP, LMR, and CDAI had smaller effects (Figure 6A). In the endoscopic recurrence model, PA again ranked highest, followed by preoperative biologic therapy, Fg, and ALB, while perianal abscess and the Montreal B classification contributed less (Figure 6B). Single-sample SHAP plots illustrated the contribution of individual variables to the model output (Supplementary Figure 8). Dependence analyses further showed that s higher CRP, Montreal B classification, and CDAI were associated with an increased predicted surgical risk, whereas a higher PA, Fg, and LMR were associated with a lower predicted risk (Supplementary Figure 9A). For postoperative endoscopic recurrence, the Montreal B classification and perianal abscess were positively associated with the predicted risk, whereas PA, Fg, ALB, and preoperative biologic therapy were negatively associated with the predicted risk (Supplementary Figure 9B).

Figure 6 SHapley Additive exPlanations beeswarm plots. A: The relative contributions of individual features to the overall prediction of abdominal surgery gradient boosting machine model; B: Postoperative endoscopic recurrence (random survival forest model). Fg: Fibrinogen; PA: Prealbumin; CDAI: Crohn’s Disease Activity Index; LMR: Lymphocyte-to-monocyte ratio; CRP: C-reactive protein; ALB: Albumin; SHAP: SHapley Additive exPlanations.

Role of PA in improving the risk-prediction models

To assess the incremental value of considering the PA level, we developed ML and Cox models excluding PA and compared their performance with the corresponding full models in the validation cohort using NRI and IDI. For abdominal surgery, the predictive performance improved progressively from the Cox model to the GBM model without PA and then to the full GBM model at 1, 3, and 5 years (Figure 7A). The full GBM model achieved a 5-year NRI of 0.463 and an IDI of 0.182. Similarly, for postoperative endoscopic recurrence, the prediction performance improved stepwise from the Cox model to the RSF model without PA and then to the full RSF model at 0.5, 1, and 2 years (Figure 7B). The full RSF model achieved a 2-year NRI of 0.629 and an IDI of 0.265. These findings indicate that PA contributed materially to the risk discrimination and reclassification under both prediction settings.

Figure 7 Incremental contribution of prealbumin to the predictive performance of the models in the validation cohort. Net reclassification improvement and integrated discrimination improvement were used to evaluate the added predictive value of prealbumin in the models with and without prealbumin. A: Abdominal surgery; B: Postoperative endoscopic recurrence. NRI: Net reclassification improvement; IDI: Integrated discrimination improvement; PA: Prealbumin; GBM: Gradient boosting machine; CI: Confidence interval.

Risk stratification and clinical decision-making applications

Based on the risk scores generated by the GBM model for the risk of needing abdominal surgery and the RSF model for the risk of postoperative endoscopic recurrence, patients were stratified into quartiles. Surgery-free and recurrence-free survival differed significantly across risk groups in both the training and validation cohorts, with progressively worse outcomes at a higher predicted risk (Figure 8A-D).

Figure 8 Kaplan-Meier analyses of model-based risk stratification and the association of biologic therapy with the disease outcomes. A and B: Kaplan-Meier curves for abdominal surgery according to quartiles of the gradient boosting machine-derived risk score in the training (A) and validation (B) cohorts; C and D: Kaplan-Meier curves for postoperative endoscopic recurrence according to quartiles of the random survival forest-derived risk score in the training (C) and validation (D) cohorts. Q1 to Q4 represent increasing model-predicted risk; E-H: Kaplan-Meier curves showing the association between biologic therapy and abdominal surgery within each gradient boosting machine-defined risk quartile: Q1 (E), Q2 (F), Q3 (G), and Q4 (H); I-L: Kaplan-Meier curves showing the association between biologic therapy and postoperative endoscopic recurrence within each random survival forest-defined risk quartile: Q1 (I), Q2 (J), Q3 (K), and Q4 (L). P values were calculated using the log-rank test.

Subgroup analyses further suggested there was a differential benefit from early intensified biologic therapy across the risk strata. For abdominal surgery, no significant reduction in surgical risk was observed in the low-risk groups (Q1-Q2; Figure 8E and F), although not statistically significant in the intermediate-risk groups (Q3), a trend toward improvement was observed with extended follow-up (Figure 8G). A clearer benefit emerged in the high-risk group (Q4; Figure 8H). For postoperative endoscopic recurrence, early intensified biologic therapy was not significantly associated with a lower recurrence risk in the low-risk groups (Q1-Q2; Figure 8I and J), but was associated with a reduced recurrence in the intermediate- and high-risk groups (Q3-Q4; Figure 8K and L). These findings support the clinical utility of model-based risk stratification for identifying patients most likely to benefit from early biologic intervention.

Finally, to facilitate clinical translation, we developed an online tool for the individualized prediction of the need for abdominal surgery and postoperative endoscopic recurrence, with automated risk stratification and treatment guidance. The tool is available at: https://crohndisease.shinyapps.io/cdpasurgeryandrecurrence/.

Association between PA levels and the ML model-derived risk scores

Serum PA levels were closely associated with the ML-based risk stratification. For abdominal surgery, most patients with PA ≤ 208.149 mg/L were classified into the intermediate- and high-risk groups (Q3-Q4; 67.1%, 397/691), while most patients with PA > 208.149 mg/L were classified into the low-risk groups (Q1-Q2; 70.7%, 332/469) (Figure 9A). A similar pattern was observed for postoperative endoscopic recurrence, wherein 71.8% (87/121) of the patients with PA ≤ 148.816 mg/L were assigned to the intermediate- and high-risk groups (Q3-Q4), while 64.5% (104/164) of those with PA > 148.816 mg/L were classified into the low-risk groups (Q1-Q2) (Figure 9B).

Figure 9 Association between prealbumin and the risk scores based on the machine learning models. A: Surgery; B: Endoscopic recurrence.

DISCUSSION

Prognostic and clinical implications of PA in CD

Our study identified PA as a simple and actionable prognostic marker in CD. We found that a lower baseline PA level independently predicted both the need for long-term abdominal surgery and postoperative endoscopic recurrence, with stable performance over time. Patients in the highest PA quartile had an approximately 80% lower risk of needing surgery and a 55% lower risk of postoperative recurrence than those in the lowest quartile.

Spline analyses identified PA threshold levels of 208.1 mg/L for abdominal surgery and 148.8 mg/L for postoperative recurrence, below which the risks increased sharply. PA therefore appears useful not only as a continuous marker but also as a practical tool for risk stratification. Importantly, PA may also have value as a complementary biomarker in relation to other established indices in CD. CRP is widely used but mainly reflects systemic inflammation[29], whereas ALB may be less responsive to short-term changes in the inflammatory–nutritional status because of its longer half-life[30,31]. Faecal calprotectin is a well-validated marker of intestinal inflammation and postoperative recurrence[32,33], but stool-based testing is not always readily accepted in routine practice. In this context, PA may offer additional clinical value by integrating information on both the inflammation and nutritional status in a simple blood-based measurement[31]. This association is biologically plausible, as PA reflects the combined burden of inflammation and nutritional impairment. As a negative acute-phase protein with a short half-life, PA is rapidly suppressed during cytokine-driven inflammatory responses, particularly those mediated by interleukin-6 and tumor necrosis factor-α[34], while low PA levels may also reflect inadequate protein and micronutrient reserves required for epithelial restitution and mucosal healing[35]. Reduced PA levels may therefore indicate impaired hepatic protein synthesis, defective intestinal repair, and disrupted immune homeostasis, all of which may favour persistent transmural inflammation and a progression to stricturing or penetrating disease[36,37]. Low PA levels also identified patients more likely to benefit from earlier or intensified biologic therapy, supporting the potential role of considering the PA level in the treatment stratification and precision management of CD patients.

Mediation by inflammation score, Fg, and BMI

Mediation analyses suggested that the prognostic effect of PA in CD is mediated through multiple interconnected pathways, including inflammation, coagulation, and nutritional status, rather than a single biological axis[21,38,39]. Specifically, the inflammation score and Fg partly explained the association between PA and abdominal surgery, whereas the inflammation score and BMI partly explained the association with postoperative endoscopic recurrence. These findings are biologically plausible, as PA is suppressed during the acute-phase response, Fg reflects a prothrombotic and injury-prone intestinal milieu[40-42], and BMI captures information on the nutritional reserve relevant to mucosal healing[38,39]. However, we found that these mediators accounted for only a modest proportion of the total effect, suggesting that broader mechanisms, including metabolic, immune, and host–microbiome pathways, may also contribute[43-45].

Predictive modelling and clinical application

Our modelling analyses further support the prognostic relevance of PA in CD. Among the seven ML-based survival models we developed and investigated, GBM showed the best performance for abdominal surgery and RSF for postoperative recurrence, with both outperforming the conventional Cox model. This suggests that ML approaches may better capture the nonlinear and complex relationships underlying disease prognosis. Consistent with this, PA was the strongest predictor in both models upon SHAP analysis, and lower PA levels were consistently found to be associated with a higher predicted risk. This also has clinical implications, as risk stratification could identify patients more likely to benefit from early intensified biologic therapy. To support clinical translation, we also developed a web-based tool incorporating a calculator integrating PA, Fg, and other key variables for individualized risk estimation. The close concordance we found between the PA thresholds and model-derived risk categories further suggests that PA may serve as a practical surrogate for risk stratification when ML tools are not readily available.

Limitations and future directions

Several limitations should be acknowledged when interpreting our findings. First, this was a single-centre retrospective observational analysis, and several sources of bias cannot be excluded, including selection bias, centre-specific practice patterns, and residual confounding. In particular, treatment decisions, surgical timing, and postoperative surveillance may have been influenced by clinical judgment and local management strategies, which could have affected the observed associations. Second, although the models showed good performance in the internal validation, they were developed and validated within the same cohort, and their generalizability to other populations, healthcare settings, and treatment contexts remains uncertain. External validation in independent multicentre cohorts will therefore be essential before broader clinical application. Third, PA was measured at a single time point, which may not fully capture its dynamic changes during the disease course or around treatment escalation and surgery. Fourth, although we observed that PA-based and ML-based risk stratification may help identify patients more likely to benefit from biologic therapy, this finding remains observational and should be interpreted cautiously. The timing of biologic initiation, discontinuation, switching, and treatment intensification was not standardized, and confounding by indication cannot be excluded. Therefore, the potential value of PA-guided or ML-guided treatment stratification requires further validation using more rigorous causal inference approaches, including a better control of treatment-related confounding, target trial emulation, and ideally prospective randomized controlled trials. Finally, although we observed a prognostic relevance of PA, the underlying mechanisms could not be fully explored, as deeper biological data, including microbiome and multi-omics profiling, were not available. Future studies should therefore focus on multicentre external validation, a longitudinal assessment of dynamic PA changes, the integration of multi-omics data, and a prospective evaluation of PA-guided risk stratification and treatment strategies.

CONCLUSION

Serum PA is a simple and actionable prognostic biomarker in CD. Its integration into threshold-based and ML-based risk assessment may support individualized treatment and precision management of CD. The study’s findings highlight the clinical value of nutrition-inflammation biomarkers in inflammatory bowel disease.

ACKNOWLEDGEMENTS

We highly appreciate all the patients who were involved in the study and the guidance and dedication of Li Yue.