Machine learning-based radiomic nomogram from unenhanced computed tomography and clinical data predicts bowel resection in incarcerated inguinal hernia

Abstract

BACKGROUND

Early identification of bowel resection risks is crucial for patients with incarcerated inguinal hernia (IIH). However, the prompt detection of these risks remains a significant challenge. Advancements in radiomic feature extraction and machine learning algorithms have paved the way for innovative diagnostic approaches to assess IIH more effectively.

AIM

To devise a sophisticated radiomic-clinical model to evaluate bowel resection risks in IIH patients, thereby enhancing clinical decision-making processes.

METHODS

This single-center retrospective study analyzed 214 IIH patients randomized into training (n = 161) and test (n = 53) sets (3:1). Radiologists segmented hernia sac-trapped bowel volumes of interest (VOIs) on computed tomography images. Radiomic features extracted from VOIs generated Rad-scores, which were combined with clinical data to construct a nomogram. The nomogram’s performance was evaluated against standalone clinical and radiomic models in both cohorts.

RESULTS

A total of 1561 radiomic features were extracted from the VOIs. After dimensionality reduction, 13 radiomic features were used with eight machine learning algorithms to develop the radiomic model. The logistic regression algorithm was ultimately selected for its effectiveness, showing an area under the curve (AUC) of 0.828 [95% confidence interval (CI): 0.753-0.902] in the training set and 0.791 (95%CI: 0.668-0.915) in the test set. The comprehensive nomogram, incorporating clinical indicators showcased strong predictive capabilities for assessing bowel resection risks in IIH patients, with AUCs of 0.864 (95%CI: 0.800-0.929) and 0.800 (95%CI: 0.669-0.931) for the training and test sets, respectively. Decision curve analysis revealed the integrated model’s superior performance over standalone clinical and radiomic approaches.

CONCLUSION

This innovative radiomic-clinical nomogram has proven to be effective in predicting bowel resection risks in IIH patients and has substantially aided clinical decision-making.

Key Words: Incarcerated inguinal hernia; Radiomics; Bowel resection; Unenhanced computed tomography; Texture analysis; Machine learning

Core Tip: This study developed an innovative radiomic-clinical nomogram to predict bowel resection risks in patients with incarcerated inguinal hernia (IIH). By extracting 13 radiomic features from unenhanced computed tomography scans and combining them with clinical data, a predictive model was created. The nomogram showed strong performance with area under the curves of 0.864 in the training set and 0.800 in the test set. Decision curve analysis demonstrated that the integrated model outperformed standalone clinical and radiomic approaches, offering a valuable tool for improving clinical decision-making in IIH patient management.

INTRODUCTION

Inguinal hernia is a prevalent condition in general surgery, with over 20 million people worldwide undergoing groin hernia repairs each year[1]. This condition can become life-threatening, particularly in elderly individuals, when it progresses to incarcerated or strangulated inguinal hernias, with an incidence rate of 3.25-7.17 per 100000 population annually[2]. The intestine, often a predominant component in hernias, can become acutely incarcerated within the hernia ring, potentially leading to bowel necrosis, perforation, and subsequent intra-abdominal infection[3]. Patients suffering from bowel necrosis due to incarcerated inguinal hernia (IIH) typically experience prolonged hospitalization and face a higher rate of postoperative complications, ranging from 6% to 43%, with a mortality rate up to as high as 7.5%[4]. Surgical resection is often necessary, and approximately 15% of IIH cases necessitate bowel resection[5,6]. Delays in surgical intervention can exacerbate bowel obstruction and intra-abdominal sepsis, significantly increasing morbidity risk[7]. Thus, early identification of the risk for bowel resection is crucial for patients with IIH.

In the preoperative context, predicting the likelihood of bowel ischemia and necrosis is challenging, although several risk factors have been identified, including older age, hernia type, severe coexisting disease, delayed hospitalization, and the presence of peritonitis[4,8-10]. However, these clinical findings are often individually identified, sometimes controversial, and lack external validation from large perspective sets, making it impractical to assess bowel resection risk based on a single factor. While some potential tools derived from laboratory tests exist, these biomarkers are mostly validated for intestinal obstruction and ischemia[11]. Given the different pathophysiological processes between intestinal obstruction and IIH, these tools may have limited utility in routine clinical practice for IIH patients. Radiological evaluations, particularly computed tomography (CT), are commonly used to assess IIH[12]. Recent studies have shown that features on preoperative enhanced CT imaging, such as the absence of bowel wall enhancement, could serve as a potential indicator of irreversible bowel ischemic changes in incarcerated hernias[13,14]. Whereas it is important to note that, in the context of acute abdominal pain, enhanced CT is often not the first choice for IIH patients due to its cost, potential contrast-induced side effects, and time-consuming nature of the examination[15]. Unenhanced CT has been often regarded as a rapid, effective, and accurate imaging technique in daily practice, although the correlations between imaging findings and incarcerated bowel status remain largely unexplored.

Radiomics, an emerging field that has garnered considerable interest, involves extracting quantitative parameters from CT images and converting this data into a high-dimensional format amenable to mining and subsequent analysis, enhancing decision-making processes[16]. This method has been thoroughly researched within clinical environments, notably in the diagnosis and prognosis of gastrointestinal and oncological conditions[17,18]. Radiomics extends the scope of imaging quantification beyond the capabilities of the naked eye, substantially increasing the amount of information that can be derived from CT images, which are otherwise assessable visually. Previously, we demonstrated the association between radiomic features and acute bowel ischemia, establishing that a model based on radiomics could effectively predict intestinal ischemia in cases of acute intestinal obstruction[19]. This led us to propose that radiomics could generate a relatively basic variable based solely on unenhanced CT, revealing significant differences in IIH patients with a higher or lower risk of bowel resection, due to variations in phenotypic expressions during the physiological processes of bowel compression and ischemia.

The objective of this study was to examine the relationship between the clinical-radiomic features and the bowel resection probability in patients with IIH. Range of machine learning (ML) algorithms were assessed, which were based on significant clinical-radiomic characteristics, and the most suitable ML algorithm was selected to develop predictive models for IIH patients.

MATERIALS AND METHODS

Study patients

The ethics committee of our hospital approved this study (No. QYFYWZLL29121), and the researchers adhered to all the rules set forth in the Declaration of Helsinki. For the two illustrative cases included in the discussion section, written informed consent was obtained from the patients.

This retrospective study was based on the clinical data of IIH patients treated at our center from January 2018 to December 2023. Inclusion criteria were: Inclusion criteria: (1) Age > 18 years; (2) IIH diagnosis; (3) Unenhanced CT within 4 hours post-admission; and (4) Surgical intervention during the hospital stay. Exclusion criteria included: (1) Absence of complete imaging or clinical records (n = 273); (2) Intraoperative identification of femoral hernia (n = 31); (3) Incarceration of non-bowel tissues (n = 53); and (4) Bowel resection attributed to factors unrelated to incarceration (n = 16). Ultimately, 214 patients were enrolled and randomly distributed into a training set (n = 161) and a test set (n = 53) in a 3:1 ratio using a randomization algorithm previously reported[20] (Supplementary Figure 1).

Patient demographics [age, gender, body mass index, American Society of Anesthesiologists (ASA) score], medical history (prior conditions, length of incarceration, the occurrence of peritonitis and bowel obstruction), as well as laboratory markers [white blood cell (WBC) and platelet counts, hemoglobin, albumin, D-dimer, and C-reactive protein (CRP) levels] were all part of the clinical data collected. Surgery-related data encompassed surgical approaches, complications, and duration of hospitalization. Peritonitis and bowel obstruction were identified from electronic records using specific International Classification of Diseases (ICD), tenth revision codes, which were secondary diagnoses of IIH and conformed to diagnostic criteria established by the World Society of Emergency Surgery[21].

Surgical treatment

Surgeons assessed the incarcerated bowel intraoperatively for resection indications after reduction. Clear indications for bowel resection included bowel perforation with fluid exudate, exposed bowel mucosa, black or purple necrosis, absence of bowel peristalsis, and lack of improvement after observation[22]. In cases where only minimal damage was noted on the intestinal serosa, suture repair was preferred to resection. Surgical approaches including the open and laparoscopic were determined based on surgeon preference and patient general conditions. Techniques employed ranged from mesh placements such as Lichtenstein and transabdominal preperitoneal patch plasty repairs to non-mesh methods (e.g., Bassini repair and high ligation). The meshes used were typically lightweight monofilament polypropylene with large pores.

Clinical model construction

Univariate analysis was performed to compare baseline characteristics between patients with and without bowel resection in the training set. Afterwards, a model based on significant clinical characteristics was developed by multivariable logistic regression (LR). The relative risks of independent variables were estimated using odds ratios (ORs) and 95% confidence interval (CIs).

CT image acquisition

All patients underwent unenhanced CT scans utilizing a multidetector row scanner (iCT 256, Philips Healthcare; or SOMATOM Definition Flash, Siemens Medical Systems) in preoperative settings. The imaging settings were performed as follows: Tube voltage 120 kV, detector collimation 64 mm × 0.6 mm, matrix 512 × 512, slice thickness 5 mm, and slice interval 5 mm. During the scans, patients were positioned supine, and scans included the entire abdomen.

Lesion segmentation and radiomic feature extraction

Two highly trained radiologists, each with a minimum of 10 years of experience, worked in complete anonymity and used the three-dimensional (3D) Slicer software (version 5.1.0) to segment CT images. Radiologist 1 conducted 3D segmentation of the volume of interest (VOI) and delineated the incarcerated bowel within the hernia sac while avoiding adjacent mesentery and other structures. In cases of discrepancies, the radiologists 2 collaborated to review the segmentation criteria and reached a consensus.

Prior to radiomic feature extraction, comprehensive image preprocessing was performed, including Laplacian of Gaussian filtering, wavelet transformation, intensity discretization, and radiomics matrix symmetry processing. A total of 30 CT images were chosen at random for the purpose of assessing the repeatability of feature extraction. Intra-group and inter-group correlation coefficients (ICCs) were then computed. Radiologist 1 repeated the segmentation process twice within a month, and radiologist 2 independently performed the same procedure on the selected 30 cases.

Radiomic features were extracted from the incarcerated bowel VOI using the 3D Slicer Radiomics Extension Pack (version 5.1.0). These features include various morphological and textural characteristics, encompassing four categories: Shape features, first-order features, texture features (e.g., gray level co-occurrence matrix; gray level run-length matrix; gray level size zone matrix; gray level difference matrix; neighboring gray tone difference matrix), and transform features.

Radiomic feature selection and model development

A structured four-step approach was utilized to reduce the high-dimensional data’s complexity and isolate robust radiomic features. Initially, only features that had ICCs greater than 0.75 were kept for further procedure. Secondly, each characteristic was subjected to a statistical t-test or Mann-Whitney U test; features that showed a two-tailed P value less than 0.1 were kept. Then, to evaluate and choose features that were significant (P < 0.05), Spearman correlation analysis was used. To further distinguish between resection and non-resection outcomes, principal component analysis was employed to depict case distributions from the effective feature subsets, though the results were underwhelming[23]. Finally, radiomic features with high relevance and minimal redundancy were retained after analyzing the remaining features using least absolute shrinkage and selection operator (LASSO) regression to identify the most impacting features.

To establish a radiomic signature that exhibits enhanced predictive accuracy, a total of eight ML algorithms were employed, specifically LR, support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF), extra trees (ET), eXtreme gradient boosting (XGBoost), light gradient boosting machine (lightGBM), and multilayer perceptron (MLP). The diagnostic performances of the eight models were compared by the area under the curve (AUC) of the receiver operating characteristic curve (ROC), accuracy, sensitivity, specificity, positive prediction value (PPV), and negative prediction value (NPV).

Radiomic-clinical nomogram construction and validation

To incorporate important radiomic signatures with clinical parameters in the training set, a radiomic-clinical nomogram was built. This nomogram, along with two other models (clinical and radiomic model), were further validated in a test set. ROC curves were used to assess their performance, with measures including AUC, sensitivity, specificity, and 95%CI being highlighted. Calibration of the predictive nomogram was performed using calibration curves, and the clinical utility of the models was ascertained through decision curve analysis (DCA). Figure 1 illustrates the workflow of the radiomic analysis, encompassing CT image acquisition, lesion segmentation, radiomic feature extraction, dimensionality reduction via LASSO regression, and model validation using ROC and DCA.

Figure 1 Workflow of the radiomic analysis. CT: Computed tomography; 3D: Three-dimensional; LASSO: Least absolute shrinkage and selection operator; ROC: Receiver operating characteristic; AUC: Area under the curve; DCA: Decision curve analysis.

Statistical analysis

Continuous variables were expressed as mean ± SD for data that followed a normal distribution, while for data that did not conform to normality, the median along with the interquartile range was utilized. To facilitate comparisons, the t-test was utilized for normally distributed continuous variables, while the Mann-Whitney U test was employed for those that were non-normally distributed. Counts and percentages were utilized to summarize categorical data, which were then analyzed through Pearson’s χ² test or Fisher’s exact test. The effectiveness of multivariable models was assessed using ROC analysis, which included metrics such as AUC, sensitivity, specificity, accuracy, PPV, and NPV. Statistical evaluations were conducted using R software version 4.3.2 and SPSS version 25.0, applying LASSO regression through the “glmnet” package and generating ROC and DCA curves with the “rms” and “rmda” packages. Significance was determined using a two-tailed P value threshold of less than 0.05.

RESULTS

Patients’ characteristics

The data presented in Table 1 outlines the demographic and clinical characteristics of the participants across both the training and test sets. In the training set, 41 patients (25.5%) and in the test set, 15 patients (28.3%) underwent bowel resection (P = 0.684). No marked differences in gender, age, ASA classification, or comorbidities were noted between the groups undergoing or not undergoing resection across both sets. Notably, incidences of peritonitis and bowel obstruction were considerably higher in the training set’s resection group (82.9% vs 55.8%, P = 0.004; 75.6% vs 54.3%, P = 0.026, respectively). The mean concentrations of CRP and D-dimer at the point of diagnosis were significantly higher in the resection group when contrasted with the non-resection group in the training set (38.39 ± 33.04 vs 24.25 ± 35.43, P = 0.001; 1003.41 ± 776.70 vs 665.61 ± 611.18, P = 0.001). However, no notable differences were detected in the test set. The resection group exhibited a higher likelihood of undergoing Bassini repair (41.5% vs 15.0%, P = 0.001) and a reduced likelihood of undergoing Lichtenstein procedure in the training set (14.6% vs 75.0%, P = 0.000), with increased incidences of postoperative complications such as incision infections (14.6% vs 5.0%, P = 0.043) and abdominal abscesses (7.3% vs none, P = 0.015).

Table 1 Patient’s characteristics, mean ± SD/n (%).

Clinical features	Training set (n = 161)		P value	Test set (n = 53)		P value
	Bowel resection (n = 41)	No bowel resection (n = 120)		Bowel resection (n = 15)	No bowel resection (n = 38)
Age (years)	72.15 ± 12.73	68.13 ± 18.16	0.306	63.53 ± 12.49	61.87 ± 15.26	0.898
Gender (male)	33 (80.5)	105 (87.5)	0.396	12 (80.0)	29 (76.3)	0.990
BMI (kg/m2)	23.52 ± 3.69	22.68 ± 3.32	0.177	22.73 ± 3.44	21.68 ± 2.81	0.254
ASA (%)			0.575			0.642
II	15 (36.6)	46 (38.3)		3 (20.0)	12 (31.6)
III	19 (46.3)	61 (50.8)		9 (60.0)	21 (55.3)
IV	7 (17.1)	13 (10.8)		3 (20.0)	5 (13.2)
Comorbidity (%)
Diabetes	3 (7.32)	7 (5.83)	0.999	3 (20.0)	2 (5.26)	0.258
Cardiovascular disease	10 (24.4)	18 (15.0)	0.258	5 (33.3)	6 (15.8)	0.297
Cerebrovascular disease	6 (14.6)	15 (12.5)	0.935	5 (33.3)	5 (13.2)	0.193
Respiratory disease	2 (4.9)	13 (10.8)	0.411	2 (13.3)	3 (7.9)	0.929
Malignant disease	3 (7.3)	12 (10.0)	0.842	0 (0.0)	5 (13.2)	0.340
Incarcerated time (> 4 hours)	32 (78.0)	73 (60.8)	0.046	13 (86.7)	22 (57.9)	0.046
Bowel obstruction	31 (75.6)	65 (54.3)	0.026	11 (73.3)	17 (44.7)	0.116
Peritonitis	34 (82.9)	67 (55.8)	0.004	12 (80.0)	19 (50.0)	0.092
Laboratory parameters
CRP (mg/L)	38.39 ± 33.04	24.25 ± 35.43	0.001	28.03 ± 25.10	18.11 ± 17.20	0.126
WBC (× 109/L)	8.97 ± 3.85	8.58 ± 4.10	0.555	9.47 ± 3.60	8.21 ± 3.01	0.319
PLT (× 109/L)	217.83 ± 76.33	209.47 ± 94.10	0.501	223.13 ± 60.23	236.26 ± 138.80	0.953
ALB (g/L)	35.72 ± 9.54	38.46 ± 7.66	0.156	38.11 ± 7.05	39.39 ± 6.95	0.521
HB (g/L)	132.44 ± 26.10	138.81 ± 19.77	0.095	131.47 ± 24.98	132.00 ± 17.89	0.921
D-D (μg/L)	1003.41 ± 776.70	665.61 ± 611.18	0.001	1011.33 ± 483.38	842.11 ± 470.08	0.213
Surgical approaches
High ligation	18 (43.9)	8 (6.7)	0.000	5 (33.3)	5 (13.2)	0.091
Bassini	17 (41.5)	18 (15.0)	0.001	6 (40.0)	7 (18.4)	0.100
Lichtenstein	6 (14.6)	90 (75.0)	0.000	4 (26.7)	24 (63.2)	0.017
TAPP	0 (0.0)	4 (3.3)	0.236	0 (0.0)	2 (5.3)	0.365
Patch repair	6 (14.6)	94 (78.3)	0.000	4 (26.7)	26 (68.4)	0.006
Complications
Incision infection	6 (14.6)	6 (5.0)	0.043	3 (20)	2 (5.3)	0.098
Effusion or hematoma	2 (4.9)	14 (11.7)	0.210	2 (13.3)	3 (7.9)	0.542
Atelectasis	1 (2.4)	4 (3.3)	0.776	0 (0.0)	2 (5.3)	0.365
Uroschesis	3 (7.3)	5 (4.2)	0.423	2 (13.3)	1 (2.6)	0.129
Abdominal abscess	3 (7.3)	0 (0.0)	0.015	1 (6.7)	0 (0.0)	0.108
Duration of hospitalization (day)	8.29 ± 3.66	7.43 ± 3.18	0.152	9.93 ± 5.37	7.84 ± 3.28	0.090

BMI: Body mass index; ASA: American Society of Anesthesiologists; HB: Hemoglobin; WBC: White blood cell; PLT: Platelet; ALB: Albumin; CRP: C-reactive protein; D-D: D-dimer; TAPP: Transabdominal preperitoneal patch plasty.

Clinical model construction

Univariate analyses identified incarceration duration, peritonitis, bowel obstruction, and elevated levels of CRP and D-dimer as factors significantly associated with the risk of bowel resection (P < 0.05) (Table 1). Notably, incarceration exceeding four hours also differed significantly between the groups in the test set (P < 0.05). These variables were subsequently incorporated into a multivariable LR analysis, which confirmed bowel obstruction (P = 0.012, OR = 3.153, 95%CI: 1.291-7.700), signs of peritonitis (P = 0.003, OR = 4.207, 95%CI: 1.608-11.011), and elevated levels of D-dimer (P = 0.042, OR = 1.001, 95%CI: 1.000-1.001) and CRP (P = 0.019, OR = 1.013, 95%CI: 1.002-1.024) as independent predictors of bowel resection (Table 2). Based on these findings, a clinical model predicting the risk of bowel resection was developed using a LR approach. The model demonstrated an AUC of 0.760 (95%CI: 0.679-0.840) in the training set and 0.747 (95%CI: 0.598-0.897) in the test set (Supplementary Figure 2).

Table 2 Multivariable logistic regression analysis of clinical features between patients with bowel resection and without bowel resection in the training set.

Clinical features	P value	OR	95%CI
Incarcerated time, > 4 hours	0.060	2.406	0.965-6.000
Bowel obstruction (%)	0.012	3.153	1.291-7.700
Peritonitis (%)	0.003	4.207	1.608-11.011
CRP (mg/L)	0.019	1.013	1.002-1.024
D-D	0.042	1.001	1.000-1.001

CRP: C-reactive protein; D-D: D-dimer; OR: Odds ratio; CI: Confidence interval.

Radiomic signature construction

A total of 1561 radiomic features were extracted from the incarcerated bowel VOIs. These features were categorized into 14 shape features, 306 first-order features, 374 textural features, and 867 transformation features. Diffusion maps based on high-dimension features indicated an inadequate capacity for distinguishing between the resection and non-resection groups in training datasets (Figure 2A). The LASSO feature selection technique was therefore employed to identify the optimal regularization parameter λ, also delineating the association between λ and model coefficients (Figure 2B and C). This process led to the selection of thirteen radiomic features with nonzero coefficients for the construction of the radiomic signature (Figure 2D). Supplementary Figure 3 showed the differences in selected radiomic features between the training and test sets. The interrelationships among these features are presented in Figure 3A and B, with Wavelet_LLL_firstorder_RootMeanSquare and Wavelet_LLL_firstorder_Maximum standing out as particularly influential in test sets (Figure 3C and D). Various radiomic models, developed through eight distinct algorithms, underwent evaluation for their predictive accuracy. It was determined that the LR model delivered the most superior predictive results, evidenced by AUCs of 0.828 (95%CI: 0.753-0.902) for the training set and 0.791 (95%CI: 0.668-0.914) for the test set (Table 3). The calculated probabilities from the LR model were then utilized to generate the radiomic signature for each patient, slated for future integration into the nomogram.

Figure 2 Selection of radiomic features using the least absolute shrinkage and selection operator regression model. A: Diffusion maps displaying patient clusters based on radiomic features in training set; B: Least absolute shrinkage and selection operator (LASSO) coefficient profile plot showing various log (λ) values; The vertical dashed lines indicate the 13 radiomic features with nonzero coefficients selected using the optimal λ value; C: Selection of the LASSO model’s tuning parameter (λ) utilizing 10-fold cross-test via the minimum criterion, with vertical lines indicating the optimal λ value. 10-fold cross-test was employed for parameter tuning; D: Distribution of the 13 radiomic features with nonzero coefficients. R: Bowel resection group; NR: Non-bowel resection group; MSE: Mean squared error; PC: Principal component; LASSO: Least absolute shrinkage and selection operator.

Figure 3 A nomogram based on Rad-score and clinical indicators for predicting bowel resection risk in patients with incarcerated inguinal hernia. A and B: Heatmaps illustrating the distribution of the 13 selected radiomic features in both cohorts; C and D: Box plots of the most significantly different radiomic features in the training and test cohorts; E: Construction of the radiomic-clinical nomogram; F and G: Calibration curves for the radiomic nomogram in the training and test sets. R: Bowel resection group; NR: Non-bowel resection group.

Table 3 Predictive performance of radiomic signature in multiple models.

Model	Set	AUC (95%CI)	ACC	SEN	SPE	PPV	NPV
LR	Training	0.828 (0.7523-0.902)	0.832	0.610	0.908	0.694	0.872
	Test	0.791 (0.668-0.915)	0.698	0.933	0.622	0.483	0.958
SVM	Training	0.919 (0.859-0.979)	0.863	0.927	0.842	0.667	0.971
	Test	0.742 (0.599-0.885)	0.717	0.800	0.703	0.500	0.897
KNN	Training	0.840 (0.780-0.901)	0.783	0.756	0.792	0.554	0.905
	Test	0.613 (0.463-0.764)	0.528	0.867	0.417	0.361	0.882
RF	Training	1.000 (0.999-1.000)	0.988	1.000	0.983	0.953	1.000
	Test	0.603 (0.434-0.771)	0.717	0.267	0.944	0.500	0.756
ET	Training	1.000 (1.000-1.000)	1.000	1.000	1.000	1.000	1.000
	Test	0.583 (0.418-0.748)	0.604	0.600	0.622	0.375	0.793
XGBoost	Training	1.000 (1.000-1.000)	1.000	1.000	1.000	1.000	1.000
	Test	0.665 (0.517-0.813)	0.623	0.867	0.526	0.419	0.909
LightGBM	Training	0.963 (0.936-0.989)	0.907	0.927	0.900	0.760	0.973
	Test	0.627 (0.479-0.776)	0.528	1.000	0.351	0.375	1.000
MLP	Training	0.829 (0.760-0.899)	0.770	0.780	0.767	0.533	0.911
	Test	0.686 (0.536-0.836)	0.604	0.867	0.514	0.406	0.905

LR: Logistic regression; SVM: Support vector machine; KNN: K-nearest neighbor; RF: Random forest; XGBoost: EXtreme gradient boosting; LightGBM: Light gradient boosting machine; ET: Extra trees; MLP: Multilayer perceptron; AUC: Area under the curve; CI: Confidence interval; ACC: Accuracy; SEN: Sensitivity; SPE: Specificity; PPV: Positive predictive value; NPV: Negative predictive value.

Construction and performance of the combination model

Considering the marked associations identified between peritonitis, bowel obstruction, and elevated levels of CRP and D-dimer with bowel resection in patients with IIH (P < 0.05) (Table 2), a clinical signature was formulated using the LR model. This signature, which amalgamated the four clinical risk factors and Rads-score, is depicted in the resulting nomogram (Figure 3E). The calibration curves demonstrated sufficient calibration for both the training and test sets (Figure 3F and G). The integrated model exhibited an AUC of 0.864 (95%CI: 0.800-0.929) within the training set and 0.800 (95%CI: 0.669-0.931) in the test set, indicating superior efficacy compared to independent clinical and radiomic models (Figure 4A and B, Table 4). DCA of the combined nomogram for both the training and testing sets indicated significant net benefits surpassing outcomes from the treat-all-patients or treat-none approaches in the IIH patient set (Figure 4C and D).

Figure 4 Receiver operating characteristic curves and decision curve analysis. A and B: Receiver operating characteristic curves for the clinical factor model, the radiomic signature, and the radiomic nomogram in the training sets (A) and test sets (B); C and D: Decision curve analysis for the radiomic nomogram model in the training (C) and test (D) sets. AUC: Area under the curve; CI: Confidence interval; DCA: Decision curve analysis.

Table 4 Predictive performance of clinical model, radiomic signature and nomogram in the training and test sets.

Group	Signature	AUC (95%CI)	ACC	SEN	SPE	PPV	NPV
Training set	Clinic signature	0.760 (0.679-0.840)	0.714	0.707	0.723	0.460	0.878
	Rad signature	0.828 (0.753-0.902)	0.832	0.610	0.908	0.694	0.872
	Nomogram	0.864 (0.800-0.929)	0.845	0.707	0.892	0.690	0.899
Test set	Clinic signature	0.747 (0.598-0.897)	0.811	0.533	0.946	0.727	0.833
	Rad signature	0.791 (0.668-0.915)	0.698	0.933	0.622	0.483	0.958
	Nomogram	0.800 (0.669-0.931)	0.830	0.667	0.895	0.714	0.872

AUC: Area under the curve; CI: Confidence interval; ACC: Accuracy; SEN: Sensitivity; SPE: Specificity; PPV: Positive predictive value; NPV: Negative predictive value.

DISCUSSION

The study assessed the efficacy of an unenhanced CT-based radiomic nomogram in forecasting the risk of bowel resection for patients with IIH. Comprising thirteen robust features, the radiomic signature proficiently categorized patients according to their likelihood of requiring bowel resection. A user-friendly radiomic nomogram, which incorporates both clinical and radiomic data, was shown to offer superior discriminatory power in both the training set (AUC = 0.864, 95%CI: 0.800-0.929) and test set (AUC = 0.800, 95%CI: 0.669-0.931), outperforming conventional clinical indicators.

The incidence of inguinal hernia is high, making it one of the most common surgical emergencies in general surgery[2,3]. It is crucial that early recognition of inguinal hernia, especially when accompanied by intestinal necrosis, informs the decision-making process, particularly the choice of emergent surgery. When ischemia or necrosis is detected in the incarcerated bowel, surgeons should prioritize timely intervention-including reduction, necrotic tissue excision, and hernia repair-to reduce complications and mortality[24]. Moreover, preoperative evaluation of the risk for bowel resection aids in selecting the appropriate surgical approach. In clinical practice, tension-free repair and laparoscopic surgeries are typically discouraged for patients with incarcerated hernias requiring resection of necrotic intestinal segments, due to increased risks of mesh infection and potential surgical failure or reoperation[25]. However, accumulating evidence indicates that laparoscopic mesh repair is a safe and viable option for select patients with acute incarcerated groin hernia[26]. In our retrospective set, 21%-31.6% of patients without bowel resection underwent laparoscopic patch repair, showing no increase in postoperative complications compared to the bowel resection group, thus demonstrating the practicality of laparoscopic mesh repair in the surgical treatment of these hernias. Accurate pre-surgical determination of bowel resection risk could allow for the choice of laparoscopic repair to decrease the likelihood of hernia recurrence and reoperation; in cases with a high risk of bowel resection, a laparotomy incision is preferred for thorough exploration and intestinal resection to maintain a contamination-free surgical field[27]. Establishing predictive models for the likelihood of bowel resection in incarcerated hernias is beneficial for surgical planning.

The clinical risk factors for bowel resection in patients with IIH are still debated, despite extensive research. Our study explored the correlation between peritonitis and bowel resection in IIH patients, positing that using a specific ICD code for peritonitis provides a more comprehensive and objective view of intestinal necrosis and intra-abdominal sepsis compared to relying solely on symptom descriptions[28]. We found a significant association between the presence of peritonitis and the likelihood of bowel resection, aligning with previous findings[29]. However, it is important to note that nearly half of the patients with peritonitis in our study did not undergo bowel resection (Table 1), echoing results from earlier research[29]. This suggests that while peritonitis is a strong indicator of peritoneum inflammation, it is not a reliable marker for bowel necrosis and subsequent resection. Additionally, a close correlation was observed between bowel obstruction and resection; however, it is critical to correlate evidence of bowel obstruction with the overall clinical context. In our study, bowel obstruction was classified as a secondary condition resulting from IIH, identified through specific ICD code in electronic records, with diagnoses based on clinical and radiological findings. The National Audit of Small Bowel Obstruction Steering Group found that approximately 36.5% of patients with abdominal wall hernia complicated by small bowel obstruction did not undergo bowel resection[30], suggesting that bowel obstruction does not necessarily imply the need for resection, although it is typical of acute incarceration. Our study did not find a significant association between factors such as age or gender and the risk of bowel resection, with these factors showing inconsistent performance in this context[31,32]. Additionally, discrepancies were noted in the incarceration duration between patients with and without bowel resection in both groups, yet LR failed to establish a significant link between incarceration duration and the need for bowel resection (P = 0.06, OR = 2.406, 95%CI: 0.965-6.000), likely due to the limited sample size.

The clinical relevance of blood-based biomarkers for identifying patients at risk of bowel resection remains unclear. This study aimed to identify single biomarkers and optimize combinations thereof for this clinical issue. Our data indicated that elevated levels of CRP and D-dimer predict bowel resection in IIH patients. These findings could be attributed to pathological changes occurring during bowel incarceration, such as compromised intestinal microcirculation, secondary infections, blood clotting, micro-thrombosis, blood disorders, tissue ischemia, hypoxia, and ultimately necrosis, leading to persistent inflammation and coagulation disorders[33]. Previous research on CRP yielded similar results, suggesting a coexistence of inflammatory status and bowel incarceration[34,35]. However, the utility of D-dimer in diagnosing intestinal ischemia is contentious, given its low specificity due to detectability in various clinical conditions[36]. This study did not find positive correlations between bowel resection and WBC, or lymphocyte counts, consistent with previous findings on the inconsistent predictive value of WBC and neutrophil counts for bowel necrosis in IIH patients[35,37]. To date, few practical biomarkers have been developed, and the cut-off values for specific laboratory tests to detect intestinal changes vary, limiting their clinical utility. Given the limitations of single clinical factors in predicting bowel resection risk, we selected four key variables peritonitis, bowel obstruction, CRP, and D-dimer to create a clinical model using LR methods. However, the performance of our clinical model was not exceptional, with AUC values of 0.760 (95%CI: 0.679-0.840) and 0.747 (95%CI: 0.598-0.897) for the training and test sets, respectively. This result did not show a clear advantage over those obtained from several novel single indicators developed previously. Therefore, we are considering whether integrating radiographic information could enhance the performance of the predictive model, as relying solely on clinical factors is insufficient.

Abdominal CT is prevalent in hospitals, serving as an indispensable tool for diagnosing inguinal hernias and facilitating surgical planning and decision-making[38]. CT scans characteristically depict the incarcerated bowel within an inguinal hernia defect as collapsed, with a notable transition point at the hernia neck and evident proximal bowel dilation suggesting obstruction[39]. Traditional CT indicators, such as bowel wall thickening and the presence of free extraluminal fluid within the hernia sac, signal potential intestinal ischemia but are insufficiently specific to confirm strangulation. In a multi-institutional prospective study by Kulvatunyou et al[40], the PPVs for these CT signs in anticipating bowel resection due to intestinal necrosis proved modest. Kohga et al[14] examined enhanced CT images from 76 IIH patients and identified that the lack of bowel wall enhancement reliably predicted the need for intestinal resection. However, the prohibitive costs, extended duration of exams, and the invasiveness of enhanced CT render it impractical for emergency evaluations of IIH. Additionally, the accuracy of CT signs is highly dependent on the radiologist’s expertise, and subjective interpretations could delay the diagnosis of strangulated inguinal hernia and necessary surgical interventions. Unlike conventional CT signs, radiomics extracts quantitative data from imaging, utilizing a high-dimensional feature space that unveils information imperceptible to the human eye and provides insights into the tissue microenvironment, thus reducing the subjectivity of radiological assessments[41,42]. Artificial intelligence-driven radiomics has demonstrated efficacy in preoperative staging of tumors and lymph nodes in gastrointestinal malignancies, showing significant predictive capabilities[43,44]. Recent applications of radiomics in benign conditions such as inflammatory bowel diseases and cardiovascular diseases have highlighted its potential in texture analysis to identify subtle, infarcted pathological changes and differentiate between acute and chronic ischemia vs normal tissue[45,46]. Previously, we developed a predictive model using CT radiomics features from 289 patients, aimed at identifying bowel ischemia in cases of acute bowel obstruction, which achieved AUC values of 0.892 (95%CI: 0.837-0.947) and 0.781 (95%CI: 0.619-0.944) in the training and test sets, respectively[18]. This study builds on those methodologies to investigate the role of an unenhanced CT radiomics nomogram in predicting bowel resection risk in IIH patients.

Radiomics, a relatively novel technology, was utilized in this study where we initially analyzed CT images of 161 IIH patients and outlined VOIs. Defining the VOIs was challenging due to the diverse shapes of incarcerated bowel within the hernia sac, necessitating precision by two adept radiologists in demarcating the lumen wall of incarcerated bowel to exclude the bowel cavity, inner contents, adipose tissue, and surrounding blood vessels. Following this, 1561 radiomics features were extracted from the outlined intestinal VOIs. We employed the principal component analysis technique, a method of linear transformation, to project the radiomics features onto a reduced-dimensional space to visualize the distribution of IIH cases, though the results were underwhelming (Figure 2A). Consequently, we adopted a dimensionality reduction strategy, the LASSO, which facilitated the identification of a subset of features correlated with the risk of bowel resection in IIH patients[47]. The LASSO approach, a form of linear regression, minimizes the influence of less critical features, spotlighting those essential for the predictive model. Through LASSO regression, we condensed the features to 13 of significance. Notably, the features Wavelet_LLL_firstorder_RootMeanSquare and Wavelet_LLL_firstorder_Maximum, predominant in the test set, captured the statistical nuances of the low-frequency elements of the image, reflecting the texture and intensity distribution within the “LLL” wavelet sub-band[48]. Future studies should examine the links between Wavelet_LLL_firstorder radiomics features and tissue ischemia, affirming their relevance in assessing the risk of bowel resection in IIH patients. In our research, we deployed eight distinct ML classifiers: LR, SVM, KNN, RF, ET, XGBoost, LightGBM, and MLP. Ultimately, the LR classifier was chosen for its optimal performance compared to the clinical model, particularly in AUC analysis, though it displayed a lower positive predictive value in the test set. Consequently, we integrated radiomics and clinical indicators to forge a combined predictive model, an innovative approach unexplored in IIH research. Our findings indicated a marked enhancement in performance relative to clinical models, with the AUC rising from 0.747 (95%CI: 0.598-0.897) to 0.800 (95%CI: 0.669-0.931) in the test set, underscoring the effectiveness of amalgamating radiomic and clinical data in forecasting bowel resection risk. The improved PPV when melding the radiomics model with clinical signatures underscores the overall robustness of the fusion model. Using a variety of realistic threshold probabilities, DCA showed that the nomogram provided a net benefit that the clinical factor model could not match when it came to predicting bowel resection in IIH patients. Junior surgeons and radiologists may find this novel combined nomogram beneficial for preoperative assessment of intestinal resection risk in IIH patients. Moreover, radiomics provides a method for quantifying image characteristics to yield objective results, and when integrated with auto-segmentation techniques, it promises more precise and efficient methodologies for evaluating IIH severity. This approach holds substantial promise as a supplementary tool for clinicians.

We attempted to implement this model in clinical practice. After obtaining informed consent from two patients, we utilized this model for the patients’ condition assessment, achieving favorable outcomes. A 63-year-old male patient diagnosed with IIH scored 77.5 on his Rad-score and 65 on his clinic-score, totaling 142.5 risk points (Figure 5A-C). This patient had a greater than 75% probability of requiring bowel resection. Upon reviewing the patient’s scores, we promptly arranged for emergency surgery and performed a bowel resection due to the already necrotic incarcerated bowel. Another 72-year-old male patient diagnosed with IIH scored 45 on his Rad-score and 50 on his clinic-score, for a total of approximately 95 risk points, indicating a less than 50% probability of bowel resection (Figure 5D-F). Subsequently, we successfully performed a laparoscopic mesh repair, resulting in favorable outcomes for this patient.

Figure 5 Demonstrations of bowel resection results predicted by the novel fusion model. A-C: Case 1: A 63-year-old male patient diagnosed with incarcerated inguinal hernia (IIH), including the patient’s nomogram, preoperative computed tomography (CT) images, and intraoperative findings of the incarcerated bowel; D-F: Case 2: A 72-year-old male patient diagnosed with IIH, featuring the patient’s nomogram, preoperative CT images, and intraoperative findings from laparoscopic mesh repair procedures for the patient without bowel resection.

Several caveats should be mentioned with regard to this research. First, there is a lack of statistical power due to the small sample size, which is particularly problematic when it comes to bowel resection surgery. Secondly, the study was limited to patients with inguinal oblique and inguinal hernias affecting the small intestine, omitting evaluations of femoral hernias, incisional hernias, and other groin hernias involving tissue entrapment. Thirdly, the study relied on data from a single center, lacking validation from multi-center data. Furthermore, as the study focused solely on patients undergoing emergency surgery, it predominantly utilized plain CT scans, and further research on enhanced CT scans is essential to strengthen the evidence base for future clinical applications.

CONCLUSION

In summary, we have effectively created and confirmed a radiomics-based model that combines radiomic and clinical characteristics to assess the likelihood of bowel resection in individuals with IIH. This model offers significant potential for clinicians to develop individualized and precise treatment plans.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge Zhu L for developing the software tools, Liu SL for validating the analytical methods, Wang ZB for visualization and investigative support, Liu JN for meticulous data curation, Zhou XM for supervisory guidance, and Hu JL for methodological expertise. All acknowledged contributors have reviewed the manuscript and endorse its data and conclusions.