Application of machine learning in the prognosis prediction of malignant large bowel obstruction: a two-cohort study

doi:10.21203/rs.3.rs-4354953/v1

Background

The Colon and Rectal NCCN Clinical Practice Guidelines currently identify obstructions as risk factors rather than as specific types. A personalized and intelligent prognostic evaluation system for malignant large bowel obstruction (MLBO) is urgently needed.

Methods

We conducted a retrospective study on 170 MLBO patients who underwent radical excision at two centers. The training and validation sets were randomly derived from the combined data of each center at a 7:3 ratio. We employed machine learning methods, including the logistic regression classifier (LR), linear discriminant analysis classifier (LDA), extreme gradient boosting classifier (XGB), AdaBoost classifier (AB), and light gradient boosting machine classifier (LGBM). These classifiers were based on clinical features (clinical model), radiological features (radiomics model), and their combination (merged model). The best model was identified through the area under the operating characteristic curve (AUC).

Results

Using clinicopathologic parameters, clinicopathologic models XGB achieved an impressive AUC of 0.97 for DFS, and LDA maintained strong performance with an AUC of 0.92 for OS, rather than radio-omics and dual-omics models. Using the Qingdao Center(QD) dataset as a single validation set, the model performance was not ideal due to demographic differences, with AUC values of 0.42 and 0.50 for DFS and OS, respectively. Finally, when cross-training and validating clinicopathological features from two centers were conducted, LDA exhibited exceptional performance for both DFS and OS, with AUCs of 0.96 and 0.95, respectively. Regardless of DFS or OS, the worse prognosis group had higher levels of the following metrics compared to the better prognosis group. [For DFS: pT(p < 0.001), pN(p < 0.006), pM(p < 0.001), monocyte count(0.64 vs. 0.52, p = 0.038), and carbohydrate antigen 199(CA199) (27.59 vs. 15.14, p = 0. 006); For OS: pT(p = 0.002), pN(p = 0.002) and pM(p < 0.001), as well as LVI (p = 0.037), monocyte count(0.68 vs. 0.51, p = 0.005) and CA199 (31.78 vs. 15.88, p = 0.006)].

Conclusions

High-efficacy models for the prognosis prediction of MLBO via clinicopathological features across two centers was constructed. We recommend heightened vigilance for MLBO patients with a high TNM stage, lymphovascular invasion occurrence, elevated CA199 levels, and high monocyte count.

Malignant large bowel obstruction

two centers

machine learning

prognosis prediction

long-term outcome

Colorectal cancer (CRC), the second most commonly diagnosed malignancy and the second leading cause of cancer-related deaths in the United States, continues to pose significant health challenges. Approximately 1.5 million new cases and nearly 52,550 resultant deaths are forecasted for the year 2023(1). A phenotype of CRC, malignant large bowel obstruction (MLBO), is an acute abdominal condition that not only complicates diagnosis but also adds complexity to life-saving strategies, obstruction resolution, and oncologic management. The Colon and Rectal National Comprehensive Cancer Network (NCCN) Clinical Practice Guidelines currently identify obstructions as mere risk factors and indications for emergency surgery. They overlook the need for a comprehensive scoring system specifically tailored for MLBO that accounts for severe microbial dysbiosis, escalating mechanical stress, and unique oncological characteristics(2, 3). Although the World Journal of Emergency Surgery provides complete remediation guidance on the initial process and management of MLBO, a dedicated prognostic scoring system is still lacking(4).

Even though the survival rate for CRC has exceeded 50% at 5 years(5, 6), nearly 40% of patients eligible for potentially curative resection encounters recurrence within first 2–3 year(7). For MLBO, regardless of the staging and adjuvant chemotherapy, both overall survival(OS) and disease-free survival(DFS) were obviously shorter than the non-MLBO CRC, especially for OS with more than 13.9–40% reduction(8, 9). Given this context, there is an urgent need to develop a personalized and intelligent prognostic evaluation system for patients with MLBO. Such a system can revolutionize our understanding and management of this challenging subset of CRC. Multidimensional approaches, leveraging pathological, radiological, immunological, and clinical features, have shown promise in providing precise prognostic predictions for CRC(10–12). The advent of deep learning, especially adept at handling large-scale data, has enabled successful evaluation of immune scores and risk factors, thereby informing CRC guidelines(13). Radiomics from preoperative magnetic resonance imaging(MRI) have been utilized in creating prognosis prediction models for rectal cancer via deep learning algorithms(14). Moreover, microsatellite instability (MSI) in CRC could be precisely predicted using deep learning methods(15). Despite these advancements, the integration of independent risk factors specifically for MLBO appears to lag behind. A comprehensive evaluation of multidimensional features is crucial.

In this study, we successfully established a robust prognostic prediction system for MLBO through the use of machine learning models in cross-cohorts. This model accurately forecasts the prognosis of MLBO patients, offering critical insights that will enable clinicians to design more precise and personalized therapeutic strategies.

2.1 Patients

2.1.1 Fujian Medical University Affiliated Union Hospital (FJ Center)

This study was undertaken at the Fujian Medical University Union Hospital in China from February 2010 to January 2020, with a 3-year postoperative follow-up. Initially, we included 207 patients who were diagnosed with MLBO. The inclusion criteria were as follows: 1) patients who presented classic symptoms of bowel obstruction, such as abdominal pain, vomiting, bloating, or constipation; 2) patients whose bowel obstruction was verified by preoperative computed tomography (CT); and 3) patients whose malignancy was confirmed via postoperative pathology. The exclusion criteria were as follows: 1) perforation, abscess, or ischemia reported on preoperative CT and 2) poor quality or improperly labeled CT images. Ultimately, our study included 127 MLBO patients (80 patients were excluded for the following reasons: 10 cases lacked preoperative CT scan; 34 cases had poor CT image quality; 16 cases lacked postoperative pathology reports of lymphovascular invasion (LVI); 20 cases lost contact or refused to cooperate.) This retrospective study was approved by the Ethics Committee of Fujian Medical University Union Hospital.

2.1.2 The Affiliated Hospital of Qingdao University (QD Center)

Simultaneously, a study was conducted at The Affiliated Hospital of Qingdao University in China between January 2019 and July 2020, with a 3-year postoperative follow-up. In this center, 43 patients with MLBO were initially considered. The inclusion and exclusion criteria matched those of the Fujian Center Study. After applying these criteria, 42 MLBO patients remained in the study (1 patient was excluded due to a postoperative pathology report indicating diffuse large B-cell lymphoma). This retrospective study was approved by the Ethics Committee of The Affiliated Hospital of Qingdao University.

2.2 Inclusion of Different Omics

2.2.1 Clinical-omics Data from the Two Centers

The preoperative features considered in this study included the following baseline information: age, sex, body mass index (BMI), comorbidities, ColoRectal Obstruction Scoring System (CROSS score), and American Society of Anesthesiologists grade (ASA grade). Blood indices: These indices included hemoglobin (HB), albumin (ALB), lymphocyte (LN), and platelet (PLT) counts and a derived score calculated as HB × ALB × LN/PLT (referred to as the HALP score)(16–19), white blood cell (WBC) count, neutrophil (NE) count, monocyte (MONO) count and carbohydrate antigen 199(CA199). Postoperative features included (1) pathological tumor stage (pT), pathological lymph node metastasis (pN), and pathological distant metastasis (pM) and LVI.

2.2.2 Radiomics of a Single Center

2.2.2.1 Preoperative 3-dimensional (3D) CT Reconstruction

The process of reconstructing the region of interest (ROI) was carried out in three primary steps. Initially, we utilized the labeling software itk-SNAP (itk-SNAP version 3.8 from the website http://www.itksnap.org) to label the images. This software was selected based on its convenience, applicability to our data files, and ability to automatically complete 3D image reconstruction. In the second step, we executed a 3D reconstruction for each scan of the preoperative CT images (consisting of 3440 DICOM-formatted scans). The reconstruction area for the obstructed segment as well as the proximal and distal sections spanned at least 5 cm(20). Finally, the initial reconstruction was performed by three surgeons, followed by a corrective review conducted by a radiologist with two decades of diagnostic reading experience. With the foundation of manually labeling the ROI on each layer of the CT scan, automatic 3D reconstruction was achieved, enabling radiological visualization.

2.2.2.2 Radiomic Feature Extraction

In the initial stage of constructing predictive classifiers, we focused on extracting preoperative clinical and radiomics features. Considering that most of the clinical data were nonnormally distributed, we employed a nonparametric rank sum test to achieve dimensionality reduction. Concurrently, we used PyRadiomics, an open-source Python package, to extract radiomic features from the ROIs(the obstructed segment). This extraction process encompassed seven types of features, including first-order statistics, morphological features, a gray-level co-occurrence matrix, a gray-level run-length matrix, a gray-level size zone matrix, a gray-level dependence matrix, and a neighborhood gray-tone difference matrix.

2.3 Follow-up

Three-year overall survival (OS) was defined as the duration spanning three years from the date of surgery to either the date of the last follow-up visit or the date of death from any cause. Three-year disease-free survival (DFS) was defined as the period extending three years from surgery to the occurrence of any recurrence. Follow-ups were conducted at the oncology outpatient clinic and supplemented with personal patient contacts, interactions with the general practitioner, and/or phone conversations with the patient's family.

Among the two participating institutions, Fujian Medical University Affiliated Union Hospital enrolled 127 patients, and the final follow-up date was December 31, 2022. The Affiliated Hospital of Qingdao University recorded the data of 42 patients (42 OS, 36 DFS), with a final follow-up date of September 1, 2023. A full consort diagram of the study is provided in Fig. 1.

2.4 Machine-learning Classification Model

2.4.1 Initial Data Processing

Addressing the issue of missing clinical data is fundamental to maintaining the integrity of our study. We adopted a scientifically reliable imputation method called MissForest, which can handle all types of data (continuous or discrete). For radiologic features, there were no missing data.

Next, considering that different types of radiomic features could lead to varied computational results due to their distinct units of measurement, we executed a mean normalization operation on the features. This procedure eliminated the unit and scale differences between the features, ensuring that each feature vector held an equal weight. While this process reduces computational effort, it also maintains the consistency of all feature information.

Moreover, given that the high dimensionality of the extracted features could result in classifier overfitting, it was crucial to apply appropriate parameters during feature extraction for optimal classification accuracy. To address this issue, we proposed the deployment of principal component analysis (PCA) for dimensionality reduction. PCA is a method that preserves all data information while reducing dimensions, implementing a stacked reduction of all features. The optimal dimensionality after this reduction process is determined by synchronizing it with the best area under the receiver operating characteristic curve (AUC) result from the validation set of models.

2.4.2 Clinical Data Processing

Initially, we processed the data from the two centers independently, adhering to the standard process outlined earlier. Subsequently, we combined the datasets from both centers and subjected this combined dataset to the same standard process. Accounting for the unequal volume of data between the two centers, it was necessary to create a balanced comparison. To achieve this, we utilized SPSS to randomly select an equivalent amount of data from the FJ Center(FJ) to match the data volume from the QD Center(QD). This allowed us to facilitate a more equitable comparative analysis.

2.4.3 Machine-learning Classifier Filtering

We randomly partitioned our dataset into a training set and a testing set at a 7:3 ratio. To identify the optimal settings for the performance parameters, we conducted 10 random divisions of these training and testing datasets.

Subsequently, we employed several classifiers—a logistic regression classifier (LR), linear discriminant analysis classifier (LDA), XGB classifier (XGB), AdaBoost classifier (AB), and light gradient boosting machine classifier (LGBM)—to predict classification outcomes. The performance of these classifiers was evaluated using the ROC and AUC. Additionally, we calculated the accuracy, precision, recall, and F1-score values to further assess the models' performance on the test set.

Ultimately, we utilized Joblib to preserve the best-performing model. All of the aforementioned data manipulations were conducted in PyCharm Community Edition 2022.2.2.

2.6 Statistics

Normally distributed continuous variables are represented as the mean (with standard deviation). Variability in such cases was investigated using an independent samples t test. For other continuous and categorical variables that did not conform to a normal distribution, the data are expressed as median values (with interquartile ranges) or frequencies (percentages), and variability was examined via nonparametric rank sum tests. A p value less than 0.05 was considered to indicate statistical significance. The majority of the statistical analyses and graph generation were conducted using SPSS for Windows software (version 25.0, Chicago, Illinois, USA). A full workflow of the study is provided in Fig. 2.

3.1 Single-center Model Using Radiomics and Clinical-omics Integration

3.1.1 Training and Validation of a Single-center Predictive Model by Radiomics

MLBO patients were randomized to two groups as follows: 88 training patients (nonrecurrence group (NRG): recurrence group (RG) = 37:51) and 39 test patients (NRG: RG = 16:23). There were 88 training patients (survival group (SG): nonsurvival group (NSG) = 43:45) and 39 test patients (SG:NSG = 18:21).

For DFS, the predictive models were dependent on 20 features after PCA. For the test dataset, the LR exhibited an AUC of 0.61 (95% confidence interval (CI): 0.48–0.79), an accuracy of 0.52, a precision of 0.64, a recall of 0.64 and an F1 score of 0.64. The LDA was inferior to the LR in terms of AUC(0.61, 95% CI: 0.49–0.80), accuracy (0.51), precision (0.67), recall (0.67) and F1 score (0.67) (details in Fig. 3, Table 1).

For OS, the predictive models were dependent on 5 features after PCA. For the test dataset, the LDA showed excellent predictive performance, with an AUC of 0.79 (95% CI: 0.56–0.87), an accuracy of 0.70, a precision of 0.75, a recall of 0.74 and an F1 score of 0.74. The LR was inferior to the LDA in terms of the AUC (0.76, 95% CI: 0.53–0.83), accuracy (0.75), precision (0.70), recall (0.69) and F1-score (0.69) (details in Fig. 3, Table 1).

Table 1

Training and Validation of Single-center Predictive Models for Radiomics
Models	AUC (95% Cl)	Accuracy	Precision	Recall	F1-score
DFS- FJ- Radiomics
LR	0.61[0.48–0.79]	0.52	0.64	0.64	0.64
LDA	0.61[0.49–0.80]	0.51	0.67	0.67	0.67
AB	0.55[0.35–0.79]	0.48	0.48	0.49	0.48
XGB	0.66[0.36–0.78]	0.47	0.61	0.62	0.61
LGBM	0.66[0.37–0.79]	0.47	0.66	0.67	0.66
OS- FJ- Radiomics
LR	0.76[0.53–0.83]	0.75	0.70	0.69	0.69
LDA	0.79[0.56–0.87]	0.70	0.75	0.74	0.74
AB	0.79[0.58–0.93]	0.63	0.85	0.85	0.85
XGB	0.72[0.57–0.92]	0.69	0.71	0.69	0.69
LGBM	0.75[0.54–0.92]	0.67	0.65	0.64	0.64
Long-term prognosis prediction results based on a FJ(Fujian Center) radiomics machine-learning models for patients with malignant large bowel obstruction. Abbreviations: AUC = area under the receiver operating characteristic curve, CI = confidence interval, DFS = 3-year disease-free survival, OS = 3-year overall survival, Classifiers: LR = Logistic Regression, LDA = Linear Discriminant Analysis, AB = adaBoosting, XGB = xgBoosting, LGBM = LightGBM.

3.1.2 Training and Validation of a Single-center Predictive Model by Clinical-omics

Finally, the models for predicting DFS depended on 5 features after PCA. For the test dataset, the XGB showed excellent predictive performance, with an AUC of 0.97 (95% CI: 0.72–0.97), an accuracy of 0.76, a precision of 0.90, a recall of 0.90 and an F1 score of 0.90. The LR was inferior to the XGB in terms of AUC(0.96, 95% CI: 0.78–0.98), accuracy (0.78), precision (0.90), recall (0.90) and F1 score (0.90) (details in Fig. 3, Table 2).

The predictive models for OS were dependent on 5 features after PCA. For the test dataset, the LDA showed excellent predictive performance, with an AUC of 0.92 (95% CI: 0.64–0.90), an accuracy of 0.74, a precision of 0.77, a recall of 0.77 and an F1 score of 0.77. The LR was inferior to the LDA in terms of the AUC (0.91, 95% CI: 0.65–0.92), accuracy (0.79), precision (0.79), recall (0.79) and F1-score (0.79) (details in Fig. 3, Table 2).

3.1.3 Training and Validation of Single-center Predictive Models for Clinical-omics and Radiomics

To develop the classification model, we employed PCA to decrease the radiological data from 1316 dimensions to 20 dimensions. Subsequently, we used the k-means method to transform the 20-dimensional features into one classification feature. Finally, we combined these categorical features with the 5 clinical features to generate the ultimate predictive model data source.

For DFS, the LR showed excellent predictive performance, with an AUC of 0.96 (95% CI: 0.75–0.97), an accuracy of 0.77, a precision of 0.87, a recall of 0.87 and an F1-score of 0.87. The LDA was inferior to the LR in terms of AUC(0.96, 95% CI: 0.74–0.96), accuracy (0.76), precision (0.85), recall (0.85) and F1 score (0.85) (details in Fig. 3, Table 2).

For OS, the predictive models were dependent on 6 features after PCA. For the test dataset, the LR showed excellent predictive performance, with an AUC of 0.92 (95% CI: 0.66–0.92), an accuracy of 0.73, a precision of 0.80, a recall of 0.79 and an F1 score of 0.80. The LDA was inferior to the LR in terms of the AUC (0.92, 95% CI: 0.67–0.93), accuracy (0.72), precision (0.82), recall (0.82) and F1-score (0.82) (details in Fig. 3, Table 2).

3.2 Testing of Predictive Models with Data from the QD

In contrast to our previous findings, once the QD dataset was defined as the single-validation set, the performance of the models seemed unsatisfactory, with AUC values of 0.42 and 0.50 for DFS and OS, respectively(details in Fig. 3, Table 2).

TABLE 2 Training, Validation and Testing of Predictive Models for Clinical-omics, Radiomics and Dual-omics on two centers
Models	AUC (95% Cl)	Accuracy	Precision	Recall	F1-score
DFS-FJ- Clinical-omics
LR	0.96[0.78 - 0.98]	0.78	0.90	0.90	0.90
LDA	0.96[0.77 - 0.98]	0.77	0.87	0.87	0.87
AB	0.89[0.71 - 0.97]	0.72	0.81	0.79	0.80
XGB	0.97[0.72 - 0.97]	0.76	0.90	0.90	0.90
LGBM	0.94[0.71 - 0.97]	0.70	0.82	0.82	0.82
OS-FJ- Clinical-omics
LR	0.91[0.65 - 0.92]	0.73	0.79	0.79	0.79
LDA	0.92[0.64 - 0.90]	0.74	0.77	0.77	0.77
AB	0.82[0.64 - 0.90]	0.74	0.80	0.79	0.79
XGB	0.84[0.62 - 0.90]	0.67	0.72	0.72	0.72
LGBM	0.88[0.62 - 0.90]	0.72	0.76	0.77	0.77
DFS-FJ- Dual-omics
LR	0.96[0.75 - 0.97]	0.77	0.87	0.87	0.87
LDA	0.96[0.74 - 0.96]	0.76	0.85	0.85	0.85
AB	0.89[0.71 - 0.95]	0.74	0.81	0.79	0.80
XGB	0.94[0.70 - 0.95]	0.71	0.85	0.85	0.84
LGBM	0.91[0.69 - 0.95]	0.72	0.79	0.79	0.79
OS-FJ- Dual-omics
LR	0.92[0.66 - 0.92]	0.73	0.80	0.79	0.80
LDA	0.92[0.67 - 0.93]	0.72	0.82	0.82	0.82
AB	0.85[0.64 - 0.92]	0.73	0.77	0.77	0.77
XGB	0.87[0.64 - 0.92]	0.67	0.77	0.77	0.77
LGBM	0.87[0.63 - 0.91]	0.72	0.76	0.72	0.71
DFS-QD- Clinical -omics
XGB-saved	0.54
OS-QD- Clinical -omics
LDA-saved	0.48
The establishment of machine learning models based on routine bicenter (clinical-omics, radiomics and dual-omics) data with the goal of long-term prognosis of patients with malignant large bowel obstruction. Abbreviations: AUC = area under the receiver operating characteristic curve, CI = confidence interval, DFS = 3-year disease-free survival, OS = 3-year overall survival, FJ = Fujian Center, QD = Qingdao Center, Classifiers: LR = Logistic Regression, LDA = Linear Discriminant Analysis, AB = adaBoosting, XGB = xgBoosting, LGBM = LightGBM, XGB-saved = XGB trained and saved for DFS using clinical-omics from Fujian Center, LDA-saved = LDA trained and saved for OS using clinical-omics from Fujian Center.

To further determine the reasons for the contradictory results between the two centers, we randomly selected instances from the FJ that corresponded to the QD using SPSS with various groupings (DFS 36:36, OS 42:42). Patients in the FJ for DFS had more advanced pT stage (p = 0.007) and greater WBC (7.85 vs. 6.38, p = 0.038) and NE (5.90 vs. 3.74, p = 0.017) counts. Conversely, lower BMIs (20.42 vs. 23.06, p = 0.013), ASA grades (p < 0.001), and CROSS scores (p < 0.001) were detected in the FJ. For OS, patients in the FJ had a lower LVI (+) rate (FJ 29%: QD 52%, p = 0.027), ASA grade (p < 0.001), and CROSS score (p < 0.001) (details in Fig. 4, Supplementary Table 1).

3.3 Machine Learning Model Construction on Two-Center Merged Datasets

3.3.1 Differential Analysis of Two-center Merged Data on DFS and OS

Owing to the different clinicopathological features, we merged two-center data for further investigation. In terms of DFS, a total of 163 patients were classified into two groups: the recurrence group (RG, n = 87) and the nonrecurrence group (NRG, n = 76). Patients in the RG exhibited significantly greater levels of pT (p < 0.001), pN (p < 0.006), and pM (p < 0.001), as well as elevated MONO counts(0.64 vs. 0.52, p = 0.038) and CA199 (27.59 vs. 15.14, p = 0.006), but had a lower ASA grade (p = 0.008) (details in Fig. 4, Supplementary Table 2).

In terms of OS, 169 patients were classified into two groups: the survival group (SG, n = 86) and the nonsurvival group (NSG, n = 83). Patients in the NSG had higher levels of pT (p = 0.002), pN (p = 0.002), and pM (p < 0.001), as well as LVI (p = 0.037), MONO counts(0.68 vs. 0.51, p = 0.005), and CA199 (31.78 vs. 15.88, p = 0.006) (details in Fig. 4, Supplementary Table 2).

3.3.2 Training and Validation of Two-center Predictive Models

For DFS, MLBO patients were randomized into two groups as follows: 114 training patients (NRG: RG = 55:59) and 49 test patients (NRG: RG = 21:28). Finally, the predictive models for DFS were dependent on 7 features after PCA. For the test dataset, the LDA showed excellent predictive performance, with an AUC of 0.96 (95% CI: 0.76–0.95), an accuracy of 0.88, a precision of 0.90, a recall of 0.88 and an F1 score of 0.88. The LR was inferior to the LDA in terms of the AUC (0.96, 95% CI: 0.74–0.94), accuracy (0.84), precision (0.85), recall (0.84) and F1-score (0.84) (details in Fig. 5, Table 3).

For OS, MLBO patients were randomized into two groups as follows: 118 training patients (NSG: SG = 62:56) and 51 test patients (NSG: SG = 21:30). Finally, the predictive models for OS were dependent on five features after PCA. For the test dataset, the LDA showed excellent predictive performance, with an AUC of 0.95 (95% CI: 0.76–0.96), an accuracy of 0.88, a precision of 0.88, a recall of 0.88 and an F1 score of 0.88. The LR was inferior to the LDA in terms of the AUC (0.95, 95% CI: 0.76–0.96), accuracy (0.86), precision (0.86), recall (0.86) and F1-score (0.86) (details in Fig. 5, Table 3)

TABLE 3 Training and Validation Results of the Predictive Models on the Two-center Merged Clinical-omics

Models	AUC (95% Cl)	Accuracy	Precision	Recall	F1-score
DFS-Two-center Merged
LR	0.95[0.74 - 0.94]	0.84	0.85	0.84	0.84
LDA	0.96[0.76 - 0.95]	0.88	0.90	0.88	0.88
AB	0.92[0.74 - 0.95]	0.84	0.84	0.84	0.84
XGB	0.92[0.74 - 0.95]	0.86	0.86	0.86	0.86
LGBM	0.93[0.73 - 0.95]	0.82	0.82	0.82	0.82
OS- Two-center Merged
LR	0.95[0.76 - 0.96]	0.86	0.86	0.86	0.86
LDA	0.95[0.73 - 0.94]	0.84	0.84	0.84	0.84
AB	0.92[0.72 - 0.94]	0.80	0.82	0.80	0.81
XGB	0.89[0.72 - 0.94]	0.84	0.84	0.84	0.84
LGBM	0.90[0.72 - 0.94]	0.84	0.85	0.84	0.84
To effectively predict DFS and OS, five machine-learning models were trained and validated by two-center merged clinical-omics date. Abbreviations: DFS = 3-year disease-free survival, OS = 3-year overall survival, AUC = area under the receiver operating characteristic curve, CI = confidence interval, classifiers: LR = Logistic Regression, LDA = Linear Discriminant Analysis, AB = adaBoosting, XGB = xgBoosting, LGBM = LightGBM.

In this study, we specifically engineered machine learning models tailored for MLBO patients. This model accurately predicts the OS and DFS of MLBO patients. MLBO is a complex condition arising from CRC with obstruction as a risk factor. Emergency interventions that weigh oncological outcomes, surgical tolerance, and the emergency situation under the concept of damage control are needed(4). Unfortunately, the current literature provides limited guidance for the prognosis of MLBO, and debates persist regarding the efficacy of stent-bridging procedures versus emergency procedures(10, 21, 22). These factors underscore the urgent need for a scoring system capable of accurately predicting long-term prognosis in MLBO patients. Our approach integrates the emergency situation, radiological features, and oncological characteristics into a unified model. Several prognostic models have been proposed for evaluating MLBO. Classical diagnostic and therapeutic standards for CRC include TNM staging(2, 3), while inflammation, a pivotal clinical feature of MLBO, has been closely linked to prognosis(23-25). HALP baseline metrics, which are used to assess the nutritional status of cancer patients, have proven to be reliable prognostic indicators(26, 27). Moreover, the ASA grade was developed for cardiopulmonary evaluation(28), and CROSS scores were used to measure the degree of obstruction(29). However, no existing prognostic evaluation system specifically addresses MLBO by concurrently considering these multifaceted aspects.

Our study successfully reconstructed the obstructed segment in MLBO and conducted a professional conversion of image data. This builds on previous applications for educational purposes, such as 3D-printed models, reconstruction of mesenteric vessels, and MRI reconstruction of rectal cancer(30-33). However, the ultimate predictive results from the classifier were not entirely satisfactory. The LR for DFS displayed an AUC of 0.61 (95% CI: 0.48-0.79), while the LDA demonstrated decent performance for OS, with an AUC of 0.79 (95% CI: 0.56-0.87). The relatively small sample size used in this research could limit the efficacy of radiomics. In contrast, the clinicopathological feature classifier outperforms the radiological feature classifiers.

At our center, we constructed a novel predictive model considering factors such as LVI, nutritional information, tumor markers, ASA grade, and the CROSS score, all underpinned by TNM stage. After conducting data downscaling and self-clustering, the model yielded exceptional accuracy in predicting survival outcomes. XGB exhibited remarkable performance, with an AUC of 0.97 (95% CI: 0.72-0.97) for DFS. Moreover, the LDA showed excellent performance, with an AUC of 0.92 (95% CI: 0.64-0.90) for OS. Both classifiers have proven to be reliable within the medical field(34-37). To further optimize the model, we integrated clinical genomics and radiomics data. However, despite our efforts, radiomics did not significantly enhance the overall effectiveness of the model. The predictive performance of the LR remained outstanding, with an AUC of 0.96 (95% CI: 0.75-0.97) for DFS and 0.92 (95% CI: 0.66-0.92) for OS. Given these compelling results, we concluded that single clinicopathological classifiers exhibit robust efficacy(38, 39). To further validate our predictive model, we applied it to an external validation cohort. However, as a single validation set, the predictive efficacy was negatively impacted, yielding an AUC of 0.42 for DFS and 0.50 for OS.

To understand these conflicting results, we carried out a randomized matched difference study between centers to identify potential discrepancies. Regarding DFS, our center exhibited a higher recurrence rate, potentially attributable to higher T-stage and inflammation levels. Simultaneously, our center recorded lower CROSS scores and ASA grades for OS, likely due to more severe degrees of bowel obstruction and greater cardiopulmonary tolerance. A lower BMI was also observed at our center, reflecting body type differences between southern and northern regions in China(40). For OS, our center reported a lower incidence of LVI, indicating comparatively less tumor dissemination(41). These discrepancies undermine the predictive efficacy when treating all QD's data as a single validation set. Numerous multicenter studies often compare data variability without merging the data, which can lead to mixed outcomes. Some validation sets performed well(42-44), while others did not meet expectations(45).

Therefore, we initiated training and testing using hybrid data from both centers, resulting in improved predictive accuracy for DFS (using the LDA, with an AUC of 0.96, 95% CI of 0.76–0.95) and OS (using the LDA, with an AUC of 0.95, 95% CI of 0.76–0.96). Based on these findings, we propose that multicenter data should be merged prior to initiating training. In our internal variability study involving merged data from two centers, we discovered that higher TNM stage, the occurrence of LVI, high CA199 levels, and high MONO levels were closely correlated with recurrence and mortality in MLBO patients. This finding is consistent with previous studies and anticipated outcomes(25, 46, 47). The severity of the tumor burden and inflammatory response mirrors the progression of MLBO(48-50). This congruence further supports the robustness of our predictive model and accentuates the significance of these factors in determining the prognosis of MLBO patients.

Despite these promising findings, our study has several limitations. First, this was a retrospective investigation conducted with a relatively small sample size. Second, the selection of classifiers was confined to those that are currently mature and widely used. Future studies might necessitate the development of custom classifiers for more specific or nuanced applications.

High-efficacy models for the prognosis prediction of MLBO via clinicopathological features across two centers was constructed. Preprocessing was needed to balance the bias between centers in cross-training and those in validation. We recommend heightened vigilance for MLBO patients with a high TNM stage, LVI occurrence, elevated high levels of CA199 and MONO.

CRC = Colorectal cancer, MLBO = malignant large bowel obstruction, NCCN = National Comprehensive Cancer Network, MRI = magnetic resonance imaging, MSI = microsatellite instability, FJ Center = Fujian Medical University Affiliated Union Hospital, QD Center = The Affiliated Hospital of Qingdao University, CT = computed tomography, LVI = lymphovascular invasion, BMI = body mass index, CROSS score = ColoRectal Obstruction Scoring System, ASA grade = American Society of Anesthesiologists grade, HB = hemoglobin, ALB = albumin, LN = lymphocyte count, PLT = platelet, WBC = white blood cell count, NE = neutrophil count, MONO = monocyte count, CA199 = carbohydrate antigen 199, pT = pathological tumor grade, pN = pathological lymph node metastasis, pM = pathological distant metastasis, 3D = three-dimensional, ROI = region of interest, OS = three-year overall survival, DFS = three-year disease-free survival, PCA = principal component analysis, AUC = area under the receiver operating characteristic curve, LR = logistic regression classifier, LDA = linear discriminant analysis classifier, XGB = XGB classifier, AB = AdaBoost classifier, LGBM = light gradient boosting machine classifier, NRG = nonrecurrence group, RG = recurrence group, SG = survival group, NSG = nonsurvival group, CI = confidence interval.

Ethics approval and consent to participate: This retrospective study was approved by the Ethics Committee of Fujian Medical University Union Hospital(2023KJT012) and The Affiliated Hospital of Qingdao University.

Consent for publication: Not applicable.

Availability of data and materials: The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests: The authors declare that they have no competing interests.

Funding: This research was supported in part by the Natural Science Foundation of Fujian Province, China(No. 2023J01633 to Xian-Qiang Chen), the Young Scientists Fund of the National Natural Science Foundation of China (No. 82203646 to Jun-Rong Zhang) and the Young Scientists Fund of the National Natural Science Foundation of China (No. 82302467 to Zhen-Lu Li).

Authors' contributions: Shuai Chen, Jun-Rong Zhang, Zhen-Lu Li, Chang-Liang Wu and Xian-Qiang Chen had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design: Shuai Chen, Jun-Rong Zhang, Zhen-Lu Li, Chang-Liang Wu, Xian-Qiang Chen. Acquisition, analysis, or interpretation of data: Shuai Chen, Jun-Rong Zhang, Zhen-Lu Li. Drafting of the manuscript: Shuai Chen, Jun-Rong Zhang, Zhen-Lu Li. Critical revision of the manuscript for important intellectual content: Shuai Chen, Jun-Rong Zhang, Xian-Qiang Chen. Statistical analysis: Shuai Chen, Jun-Rong Zhang. Administrative, technical, or material support: Shuai Chen, Jun-Rong Zhang, Zhen-Lu Li, Cang-Dian Huang, Peng-Sheng Tu, Wen-Xuan Chen, Xin-Chang Shang-Guan. Supervision: Shuai Chen, Jun-Rong Zhang, Zhen-Lu Li, Cang-Dian Huang, Peng-Sheng Tu, Wen-Xuan Chen, Xin-Chang Shang-Guan, Chang-Liang Wu, Xian-Qiang Chen.

Acknowledgements: Not applicable.

Authors' information (optional): Not applicable.

Siegel R, Miller K, Wagle N, Jemal A. Cancer statistics, 2023. CA: a cancer journal for clinicians. 2023;73(1):17-48.
*Al B. Benson I, MD/Chair † Robert H. Lurie Comprehensive Cancer Center of Northwestern University, *Alan P . Venook MV-CUHDFCCC, Mohamed Adam MUHDFCCC, Mahmoud M. Al-Hawary MфUoMRCC, Yi-Jen Chen M, PhD § City of Hope National Medical Center, Kristen K. Ciombor MV-ICC, et al. NCCN Clinical Practice Guidelines in Oncology Rectal Cancer Version 6.2023. Journal of the National Comprehensive Cancer Network : JNCCN. 2023.
*Al B. Benson I, MD/Chair † Robert H. Lurie Comprehensive Cancer Center of Northwestern University, *Alan P . Venook MV-CUHDFCCC, Mohamed Adam MUHDFCCC, Mahmoud M. Al-Hawary MфUoMRCC, Yi-Jen Chen M, PhD § City of Hope National Medical Center, Kristen K. Ciombor MV-ICC, et al. NCCN Clinical Practice Guidelines in Oncology Colon Cancer Version 4.2023. Journal of the National Comprehensive Cancer Network : JNCCN. 2023.
Pisano M, Zorcolo L, Merli C, Cimbanassi S, Poiasina E, Ceresoli M, et al. 2017 WSES guidelines on colon and rectal cancer emergencies: obstruction and perforation. World journal of emergency surgery : WJES. 2018;13:36.
Rebecca L S, Kimberly D M, Ahmedin J. Cancer statistics, 2020. CA Cancer J Clin. 2020;70.
Roberta DA, Milena S, Michel P C, Silvia F, Paolo B, Daniela P, et al. Cancer survival in Europe 1999-2007 by country and age: results of EUROCARE--5-a population-based study. Lancet Oncol. 2013;15.
Vanesa B-B, Sonia P-D, Teresa G-R, Cristina G-M, Remedios P-P, Loreto Y-G-D, et al. Colorectal cancer recurrence and its impact on survival after curative surgery: An analysis based on multistate models. Dig Liver Dis. 2023.
Fadi AB, Randa T, Mohanad G, Amir M, Gal O, Yael K. Obstructive colon cancers at endoscopy are associated with advanced tumor stage and poor patient outcome. A retrospective study on 398 patients. Eur J Gastroenterol Hepatol. 2020;33.
Seung Han K, Se Hyun J, Han Jo J, Hyuk Soon C, Eun Sun K, Bora K, et al. Colonic stenting as a bridge to surgery for obstructive colon cancer: is it safe in the long term? Surg Endosc. 2022;36.
Wang R, Dai W, Gong J, Huang M, Hu T, Li H, et al. Development of a novel combined nomogram model integrating deep learning-pathomics, radiomics and immunoscore to predict postoperative outcome of colorectal cancer lung metastasis patients. Journal of hematology & oncology. 2022;15(1):11.
Ullah I, Yang L, Yin F, Sun Y, Li X, Li J, et al. Multi-Omics Approaches in Colorectal Cancer Screening and Diagnosis, Recent Updates and Future Perspectives. Cancers. 2022;14(22).
Tsai P, Lee T, Kuo K, Su F, Lee T, Marostica E, et al. Histopathology images predict multi-omics aberrations and prognoses in colorectal cancer patients. Nature communications. 2023;14(1):2102.
Foersch S, Glasner C, Woerl A, Eckstein M, Wagner D, Schulz S, et al. Multistain deep learning for prediction of prognosis and therapy response in colorectal cancer. Nature medicine. 2023;29(2):430-9.
Jiang X, Zhao H, Saldanha O, Nebelung S, Kuhl C, Amygdalos I, et al. An MRI Deep Learning Model Predicts Outcome in Rectal Cancer. Radiology. 2023;307(5):e222223.
Cao R, Yang F, Ma S, Liu L, Zhao Y, Li Y, et al. Development and interpretation of a pathomics-based model for the prediction of microsatellite instability in Colorectal Cancer. Theranostics. 2020;10(24):11080-91.
Chen X, Xue L, Wang W, Chen H, Zhang W, Liu K, et al. Prognostic significance of the combination of preoperative hemoglobin, albumin, lymphocyte and platelet in patients with gastric carcinoma: a retrospective cohort study. Oncotarget. 2015;6(38):41370-82.
Jiang H, Li H, Li A, Tang E, Xu D, Chen Y, et al. Preoperative combined hemoglobin, albumin, lymphocyte and platelet levels predict survival in patients with locally advanced colorectal cancer. Oncotarget. 2016;7(44):72076-83.
Peng D, Zhang C, Tang Q, Zhang L, Yang K, Yu X, et al. Prognostic significance of the combination of preoperative hemoglobin and albumin levels and lymphocyte and platelet counts (HALP) in patients with renal cell carcinoma after nephrectomy. BMC urology. 2018;18(1):20.
Guo Y, Shi D, Zhang J, Mao S, Wang L, Zhang W, et al. The Hemoglobin, Albumin, Lymphocyte, and Platelet (HALP) Score is a Novel Significant Prognostic Factor for Patients with Metastatic Prostate Cancer Undergoing Cytoreductive Radical Prostatectomy. Journal of Cancer. 2019;10(1):81-91.
Mao B, Ma J, Duan S, Xia Y, Tao Y, Zhang L. Preoperative classification of primary and metastatic liver cancer via machine learning-based ultrasound radiomics. European radiology. 2021;31(7):4576-86.
Webster P, Aldoori J, Burke D. Optimal management of malignant left-sided large bowel obstruction: do international guidelines agree? World journal of emergency surgery : WJES. 2019;14:23.
Suzuki Y, Moritani K, Seo Y, Takahashi T. Comparison of decompression tubes with metallic stents for the management of right-sided malignant colonic obstruction. World journal of gastroenterology. 2019;25(16):1975-85.
Chen X, Xue C, Hou P, Lin B, Zhang J. Lymphocyte-to-monocyte ratio effectively predicts survival outcome of patients with obstructive colorectal cancer. World journal of gastroenterology. 2019;25(33):4970-84.
Schmitt M, Greten F. The inflammatory pathogenesis of colorectal cancer. Nature reviews Immunology. 2021;21(10):653-67.
Yamamoto T, Kawada K, Obama K. Inflammation-Related Biomarkers for the Prediction of Prognosis in Colorectal Cancer Patients. International journal of molecular sciences. 2021;22(15).
Xu H, Zheng X, Ai J, Yang L. Hemoglobin, albumin, lymphocyte, and platelet (HALP) score and cancer prognosis: A systematic review and meta-analysis of 13,110 patients. International immunopharmacology. 2023;114:109496.
Zhou J, Yang D. Prognostic Significance of Hemoglobin, Albumin, Lymphocyte and Platelet (HALP) Score in Hepatocellular Carcinoma. Journal of hepatocellular carcinoma. 2023;10:821-31.
Horvath B, Kloesel B, Todd M, Cole D, Prielipp R. The Evolution, Current Value, and Future of the American Society of Anesthesiologists Physical Status Classification System. Anesthesiology. 2021;135(5):904-19.
Ohki T, Yoshida S, Yamamoto M, Isayama H, Yamada T, Matsuzawa T, et al. Determining the difference in the efficacy and safety of self-expandable metallic stents as a bridge to surgery for obstructive colon cancer among patients in the CROSS 0 group and those in the CROSS 1 or 2 group: a pooled analysis of data from two Japanese prospective multicenter trials. Surgery today. 2020;50(9):984-94.
To G, Hawke J, Larkins K, Burke G, Costello D, Warrier S, et al. A systematic review of the application of 3D-printed models to colorectal surgical training. Techniques in coloproctology. 2023;27(4):257-70.
Mari F, Nigri G, Pancaldi A, De Cecco C, Gasparrini M, Dall'Oglio A, et al. Role of CT angiography with three-dimensional reconstruction of mesenteric vessels in laparoscopic colorectal resections: a randomized controlled trial. Surgical endoscopy. 2013;27(6):2058-67.
Petkovska I, Tixier F, Ortiz EJ, Golia Pernicka JS, Paroder V, Bates DD, et al. Clinical utility of radiomics at baseline rectal MRI to predict complete response of rectal cancer after chemoradiation therapy. Abdom Radiol (NY). 2020;45(11):3608-17.
Antunes JT, Ofshteyn A, Bera K, Wang EY, Brady JT, Willis JE, et al. Radiomic Features of Primary Rectal Cancers on Baseline T(2) -Weighted MRI Are Associated With Pathologic Complete Response to Neoadjuvant Chemoradiation: A Multisite Study. J Magn Reson Imaging. 2020;52(5):1531-41.
Verma N, Choudhury A, Singh V, Duseja A, Al-Mahtab M, Devarbhavi H, et al. APASL-ACLF Research Consortium-Artificial Intelligence (AARC-AI) model precisely predicts outcomes in acute-on-chronic liver failure patients. Liver international : official journal of the International Association for the Study of the Liver. 2023;43(2):442-51.
Syleouni M, Karavasiloglou N, Manduchi L, Wanner M, Korol D, Ortelli L, et al. Predicting second breast cancer among women with primary breast cancer using machine learning algorithms, a population-based observational study. International journal of cancer. 2023;153(5):932-41.
Khorrami M, Prasanna P, Gupta A, Patil P, Velu P, Thawani R, et al. Changes in CT Radiomic Features Associated with Lymphocyte Distribution Predict Overall Survival and Response to Immunotherapy in Non-Small Cell Lung Cancer. Cancer immunology research. 2020;8(1):108-19.
Sheng N, Wang Y, Huang L, Gao L, Cao Y, Xie X, et al. Multi-task prediction-based graph contrastive learning for inferring the relationship among lncRNAs, miRNAs and diseases. Briefings in bioinformatics. 2023;24(5).
Yang S, Jang H, Park I, Lee H, Lee K, Oh G, et al. Machine-Learning Algorithms Using Systemic Inflammatory Markers to Predict the Oncologic Outcomes of Colorectal Cancer After Surgery. Annals of surgical oncology. 2023;30(13):8717-26.
Kudo S, Ichimasa K, Villard B, Mori Y, Misawa M, Saito S, et al. Artificial Intelligence System to Determine Risk of T1 Colorectal Cancer Metastasis to Lymph Node. Gastroenterology. 2021;160(4):1075-84.e2.
Li Y, Yang L, Yin L, Liu Q, Wang Y, Yang P, et al. Trends in Obesity and Metabolic Status in Northern and Southern China Between 2012 and 2020. Frontiers in nutrition. 2021;8:811244.
Wang X, Cao Y, Ding M, Liu J, Zuo X, Li H, et al. Oncological and prognostic impact of lymphovascular invasion in Colorectal Cancer patients. Int J Med Sci. 2021;18(7):1721-9.
Liu Z, Wang Y, Shen F, Zhang Z, Gong J, Fu C, et al. Radiomics based on readout-segmented echo-planar imaging (RS-EPI) diffusion-weighted imaging (DWI) for prognostic risk stratification of patients with rectal cancer: a two-centre, machine learning study using the framework of predictive, preventive, and personalized medicine. The EPMA journal. 2022;13(4):633-47.
Liu Z, Guo C, Dang Q, Wang L, Liu L, Weng S, et al. Integrative analysis from multi-center studies identities a consensus machine learning-derived lncRNA signature for stage II/III colorectal cancer. EBioMedicine. 2022;75:103750.
Ma S, Lu H, Jing G, Li Z, Zhang Q, Ma X, et al. Deep learning-based clinical-radiomics nomogram for preoperative prediction of lymph node metastasis in patients with rectal cancer: a two-center study. Frontiers in medicine. 2023;10:1276672.
Wesdorp N, Zeeuw M, van der Meulen D, van 't Erve I, Bodalal Z, Roor J, et al. Identifying Genetic Mutation Status in Patients with Colorectal Cancer Liver Metastases Using Radiomics-Based Machine-Learning Models. Cancers. 2023;15(23).
Ho V, Chung L, Wilkinson K, Lea V, Lim S, Abubakar A, et al. Prognostic Significance of MRE11 Overexpression in Colorectal Cancer Patients. Cancers. 2023;15(9).
Li C, Zhao K, Zhang D, Pang X, Pu H, Lei M, et al. Prediction models of colorectal cancer prognosis incorporating perioperative longitudinal serum tumor markers: a retrospective longitudinal cohort study. BMC medicine. 2023;21(1):63.
Li C, Zhang D, Pang X, Pu H, Lei M, Fan B, et al. Trajectories of Perioperative Serum Tumor Markers and Colorectal Cancer Outcomes: A Retrospective, Multicenter Longitudinal Cohort Study. EBioMedicine. 2021;74:103706.
Sefrioui D, Beaussire L, Gillibert A, Blanchard F, Toure E, Bazille C, et al. CEA, CA19-9, circulating DNA and circulating tumour cell kinetics in patients treated for metastatic colorectal cancer (mCRC). British journal of cancer. 2021;125(5):725-33.
Bertocchi A, Carloni S, Ravenda P, Bertalot G, Spadoni I, Lo Cascio A, et al. Gut vascular barrier impairment leads to intestinal bacteria dissemination and colorectal cancer metastasis to liver. Cancer cell. 2021;39(5):708-24.e11.

No competing interests reported.

SupplementaryMaterials.docx

Application of machine learning in the prognosis prediction of malignant large bowel obstruction: a two-cohort study

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Background

Methods

2.1 Patients

2.1.1 Fujian Medical University Affiliated Union Hospital (FJ Center)

2.1.2 The Affiliated Hospital of Qingdao University (QD Center)

2.2 Inclusion of Different Omics

2.2.1 Clinical-omics Data from the Two Centers

2.2.2 Radiomics of a Single Center

2.2.2.1 Preoperative 3-dimensional (3D) CT Reconstruction

2.2.2.2 Radiomic Feature Extraction

2.3 Follow-up

2.4 Machine-learning Classification Model

2.4.1 Initial Data Processing

2.4.2 Clinical Data Processing

2.6 Statistics

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1