Machine Learning to Predict Fascial Dehiscence after Exploratory Laparotomy Surgery

Development and validation of a machine learning–based model for varices screening in compensated cirrhosis (CHESS2001): an international multicenter study

To investigate the value of machine learning(ML) in enhancing prostate cancer(PCa) diagnosis. Consecutive systematic prostate biopsies performed from Jan 2003-June 2017 were used as the training cohort, and prospective biopsies performed from July 2017-November 2019 were used as validation cohort. Men were included if PSA was 0.4-50 ng/mL, and information of digital rectal examination (DRE), Transrectal ultrasound(TRUS) prostate volume, TRUS abnormality were known. Clinically significant PCa(csPCa) was defined as Gleason 3 + 4 or above cancers. Area-under-curve (AUC) of receiver-operating characteristics (ROC) was compared between PSA, PSA density, European Randomized Study of Screening for Prostate Cancer (ERSPC) risk calculator (ERSPC-RC), and various ML techniques using PSA, DRE and TRUS information. ML techniques used included XGBoost, LightGBM, Catboost, Support vector machine (SVM), Logistic regression (LR), and Random Forest (RF), where cost sensitive learning was applied. Training and validation cohorts included 3881 and 778 consecutive men, respectively. RF model performed better than other ML techniques and PSA, PSA density and ERSPC-RC for prediction of PCa or csPCa in the validation cohort. In csPCa prediction, AUC of PSA, PSA density, ERSPC-RC and RF was 0.71, 0.80, 0.83 and 0.88 respectively. At 90-95% sensitivity for csPCa, RF model achieved a negative predictive value (NPV) of 97.5-98.0% and avoided 38.3-52.2% unnecessary biopsies. Decision curve analyses (DCA) showed RF model provided net clinical benefit over PSA, PSA density and ERSPC-RC. By using the same clinical parameters, ML techniques performed better than ERSPC-RC or PSA density in csPCa predictions, and could avoid up to 50% unnecessary biopsies.

MP77-06 MRI-BASED MACHINE LEARNING RADIOMICS FOR PREOPERATIVE PREDICTION OF HER2 STATUS IN UROTHELIAL BLADDER CARCINOMA

Optimization of risk stratification is important for facilitating prognoses and therapeutic decisions regarding diffuse large B-cell lymphoma (DLBCL). However, a simple and applicable prognostic tool is lacking for individuals with human immunodeficiency virus (HIV)-related DLBCL in the era of combined antiretroviral therapy (cART). This retrospective multicenter observational study included 147 HIV-related DLBCL patients with histologically confirmed DLBCL from 2013 to 2020. The total group was divided into training (n = 78) and validation (n = 69) cohorts to derive the best prognostic score. Clinicopathological and characteristic biomarkers correlated with clinical outcomes were analyzed. Age, Ann Arbor stage, lactate dehydrogenase (LDH) ratio, bulky disease, and red blood cell distribution width (RDW) ratio retained robust independent correlations with overall survival (OS) in multivariate analysis. A new and practical prognostic model was generated and externally validated, classifying patients into three categories with significantly different survival rates. Moreover, the new index outperformed the International Prognostic Index (IPI) score (area under the curve values of 0.94 vs. 0.81 in the training cohort and 0.85 vs. 0.74 in the validation cohort, C-indices of 0.80 vs. 0.70 in the training cohort and 0.74 vs. 0.70 in the validation cohort, and integrated discrimination improvement values of 0.203 in the training cohort and 0.175 in the validation cohort) and was better at defining intermediate- and high-risk groups. The calibration curves performed satisfactorily for predicting 3-year OS in the training and validation cohorts. We developed and validated a simple and feasible prognostic model for patients with HIV-related DLBCL that had more discriminative and predictive accuracy than the IPI score for risk stratification and individualized treatment in the cART era.

655 Background: Pancreatic neuroendocrine tumors (PNETs) have varying biologic behavior based on factors including tumor grade. Many prognostic factors can only be determined post-pancreatectomy; thus, are unable to guide clinical decision making preoperatively. This is particularly relevant in small, non-functional (NF)-PNETs. Imaging analysis with radiomics may be able to elucidate biologic data such as tumor grade, without the need for tissue sampling. Our study aims to create an automatic pipeline from segmentation to radiomics-signature building to preoperatively predict tumor grade. Methods: Patients that underwent resection between 2003-21 with adequate preoperative arterial phase CT scans were divided into training and validation cohorts. In the training cohort, the pancreas and tumor region were manually segmented and used to train an auto-segmentation model. The validation cohort then underwent automatic segmentation. A total of 255 radiomics features were extracted from the tumor region. Correlated features were removed, and the minimum redundancy maximum relevance (mRMR) method was used to select the final feature set. These features were used to train a Support Vector Machine (SVM)-based classifier through the training cohort to predict grade, which was dichotomized into grade I versus II/III. The radiomic based prediction model was then evaluated in the automatically segmented validation cohort. Results: A total of 186 patients were identified and divided into training (n = 140) and validation (n = 46) cohorts. Median age was 57 (27, 85) and 52% were male. 113 (62%) and 70 (38%) patients were grades I and II/III, respectively. Median tumor size was 24 mm (6, 200) and 43 (26%) patients had positive nodal disease. The auto-segmentation model was able to accurately segment the tumor region in 33 (72%) patients of the validation cohort. In the training cohort (n = 140), the radiomics-signature produced an area under the curve (AUC) of 0.87 (0.81, 0.93). Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were 0.94 (0.9, 0.98), 0.74 (0.66, 0.81), 0.72 (0.65, 0.80) and 0.97 (0.94, 1.00) respectively. In the auto-segmented validation cohort (n = 33), the radiomics model produced an AUC of 0.75 (0.61, 0.90). Sensitivity, specificity, PPV and NPV were 0.88 (0.77, 0.99), 0.62 (0.46, 0.79), 0.71 (0.56, 0.87) and 0.83 (0.71, 0.96), respectively. Conclusions: The proposed auto-segmentation model was able to automatically identify tumor regions with a high degree of accuracy. Similarly, the subsequent radiomics-signature was also able to strongly discern PNET grade. This automatic pipeline bypasses manual segmentation and could help incorporate a radiomics-signature into preoperative clinical decision-making for PNETs.

Background Androgen receptor signaling inhibitors (ARSIs), including abiraterone acetate and enzalutamide, were approved for metastatic castration-resistant prostate cancer (mCRPC) based on landmark trials showing improved survival in both the pre- and post-chemotherapy settings. However, multiple real-world studies have demonstrated increased cardiovascular (CV) morbidity with these agents. Purpose To develop and validate a machine learning-based CV risk prediction model in prostate cancer (PCa) patients receiving ARSIs. Methods A nationwide cohort study was conducted utilizing data from the National Health Insurance Research Database containing the Taiwan Cancer Registry. The study population comprised 4,739 PCa patients who received abiraterone or enzalutamide between January 1, 2014, and February 28, 2022. The cohort was divided into a training set (70%, n=3,318) and a validation set (30%, n=1,421). The index date was defined as the initiation of either ARSI. A machine learning technique with random survival forest (RSF) model incorporating 16 variables was developed to predict major adverse cardiovascular events (MACEs), which were defined as a composite outcome encompassing heart failure (HF) hospitalization, myocardial infarction, ischemic stroke, and cardiovascular mortality. Results Over a mean follow-up period of 2.1 years, MACEs occurred in 10.9% and 11.3% of the training and validation cohorts, respectively. The RSF model utilized variable importance (VIMP) to rank the prognostic capability of 16 clinically relevant variables routinely available during ARSI treatment. These variables encompassed patient's age, ARSI type, androgen deprivation therapy (ADT) modality, eight comorbidities (not including cardiovascular disease because its components have been included), four prior cardiovascular events requiring hospitalization, and previous docetaxel use. The final RSF model identified five key predictive indicators: age <65 or ≥75 years, heart failure, stroke, hypertension, and myocardial infarction. The model exhibited robust performance, achieving an area under the curve (AUC) of 85.1% in the training set and demonstrating strong external validity with an AUC of 85.5% in the validation cohort. A positive correlation was observed between the number of risk factors and the incidence of MACEs. In the training cohort, stratification by risk factor count yielded the following 4-year MACE rates: 6.9% for patients with 0-1 risk factors, 15.3% for those with 2 risk factors, and 24.8% for individuals with ≥3 risk factors (P trend <0.001). This finding was consistent in the validation cohort. Conclusions This machine learning approach identified five predictors of MACEs in PCa patients receiving ARSIs. The risk stratification model provides clinicians with an evidence-based tool for individualized cardiovascular risk assessment prior to initiating ARSI therapy.Variable Importance Distribution Risk Stratification Based on Predictors

To develop a predictive model by 18F-FDG PET/CT radiomic features and to validate the predictive value of the model for distinguishing solitary lung adenocarcinoma from tuberculosis. A total of 235 18F-FDG PET/CT patients with pathologically or follow-up confirmed lung adenocarcinoma (n = 131) or tuberculosis (n = 104) were retrospectively and randomly divided into a training (n = 163) and validation (n = 72) cohort. Based on the Transparent Reporting of Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD), this work was belonged to TRIPOD type 2a study. The Mann-Whitney U test and least absolute shrinkage and selection operator (LASSO) algorithm were used to select the optimal predictors from 92 radiomic features that were extracted from PET/CT, and the optimal predictors were used to build the radiomic model in the training cohort. The meaningful clinical variables comprised the clinical model, and the combination of the radiomic model and clinical model was a complex model. The performances of the models were assessed by the area under the receiver operating characteristic curve (AUC) in the training and validation cohorts. In the training cohort, 9 radiomic features were selected as optimal predictors to build the radiomic model. The AUC of the radiomic model was significantly higher than that of the clinical model in the training cohort (0.861 versus 0.686, p < 0.01), and this was similar in the validation cohort (0.889 versus 0.644, p < 0.01). The AUC of the radiomic model was slightly lower than that of the complex model in the training cohort (0.861 versus 0.884, p > 0.05) and validation cohort (0.889 versus 0.909, p > 0.05), but there was no significant difference. 18F-FDG PET/CT radiomic features have a significant value in differentiating solitary lung adenocarcinoma from tuberculosis.

Neoadjuvant therapies have been shown to decrease tumor burden, increase resection rate, and improve the outcomes among patients with locally advanced gastric cancer (GC). However, not all patients are equally responsive; therefore, differentiating potential respondents from nonrespondents is clinically important. To use pretreatment computed tomography (CT)-pixelated feature-difference extraction techniques to identify diagnostically relevant features that could predict patients' response to neoadjuvant chemotherapy at diagnosis. This multicenter cohort study included patients with locally advanced GC who were treated from January 2010 to July 2017 at 2 hospitals in southern China (training cohort) and 1 hospital in northern China (external validation cohort). Their clinicopathological data, pretreatment CT images, and pathological reports were retrieved and analyzed. Data analysis was conducted from December 2017 to May 2021. All patients underwent 2 to 4 cycles of fluorouracil in combination with a platinum-based neoadjuvant chemotherapy regimen. All gastrectomies were performed according to the Japanese Classification of Gastric Carcinoma (14th edition) guidelines. Reliability of clinicopathological and radiomics-based features were assessed with area under receiver operating characteristic curve (AUC) and Mann-Whitney U test. A total of 323 patients (242 [74.9%] men; median [range] age, 58 [24-82] years) were included in the study, with 250 patients (77.4%) in the training cohort and 73 (22.6%) in the validation cohort. The baseline pretreatment characteristics of the training and validation cohorts were well-balanced. The number of respondents in the training and validation cohort was 122 (48.8%) and 40 (54.8%), respectively, and the number of nonrespondents was 128 (51.2%) and 33 (45.2%), respectively. No clinicopathological variables were significantly associated with treatment response. Using radiomics, 20 low-intercorrelated features from a total of 7477 features were used to construct a radiomics signature that demonstrated significant association with treatment response. Good discrimination performance of the radiomics signature for predicting treatment response in the training (AUC, 0.736; 95% CI, 0.675-0.798) and external validation (AUC, 0.679; 95% CI, 0.554-0.803) cohorts was observed. Decision curve analysis confirmed the clinical utility of the radiomics signature. In this study, the proposed radiomics signature showed potential as a clinical aid for predicting the response of patients with locally advanced GC before treatment, thereby allowing timely planning for effective treatments for potential nonrespondents.

Individualized Prediction and Risk Factors of Recurrence in Chinese Patients with Sebaceous Carcinoma: A Multicenter Study of 418 Patients

Background Machine learning (ML), a subset of artificial intelligence (AI), uses algorithms to analyze data and predict outcomes without extensive human intervention. In healthcare, ML is gaining attention for enhancing patient outcomes. This study focuses on predicting additional hospital days (AHD) for patients with cervical spondylosis (CS), a condition affecting the cervical spine. The research aims to develop an ML-based nomogram model analyzing clinical and demographic factors to estimate hospital length of stay (LOS). Accurate AHD predictions enable efficient resource allocation, improved patient care, and potential cost reduction in healthcare. Methods The study selected CS patients undergoing cervical spine surgery and investigated their medical data. A total of 945 patients were recruited, with 570 males and 375 females. The mean number of LOS calculated for the total sample was 8.64 ± 3.7 days. A LOS equal to or <8.64 days was categorized as the AHD-negative group (n = 539), and a LOS > 8.64 days comprised the AHD-positive group (n = 406). The collected data was randomly divided into training and validation cohorts using a 7:3 ratio. The parameters included their general conditions, chronic diseases, preoperative clinical scores, and preoperative radiographic data including ossification of the anterior longitudinal ligament (OALL), ossification of the posterior longitudinal ligament (OPLL), cervical instability and magnetic resonance imaging T2-weighted imaging high signal (MRI T2WIHS), operative indicators and complications. ML-based models like Lasso regression, random forest (RF), and support vector machine (SVM) recursive feature elimination (SVM-RFE) were developed for predicting AHD-related risk factors. The intersections of the variables screened by the aforementioned algorithms were utilized to construct a nomogram model for predicting AHD in patients. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve and C-index were used to evaluate the performance of the nomogram. Calibration curve and decision curve analysis (DCA) were performed to test the calibration performance and clinical utility. Results For these participants, 25 statistically significant parameters were identified as risk factors for AHD. Among these, nine factors were obtained as the intersection factors of these three ML algorithms and were used to develop a nomogram model. These factors were gender, age, body mass index (BMI), American Spinal Injury Association (ASIA) scores, magnetic resonance imaging T2-weighted imaging high signal (MRI T2WIHS), operated segment, intraoperative bleeding volume, the volume of drainage, and diabetes. After model validation, the AUC was 0.753 in the training cohort and 0.777 in the validation cohort. The calibration curve exhibited a satisfactory agreement between the nomogram predictions and actual probabilities. The C-index was 0.788 (95% confidence interval: 0.73214–0.84386). On the decision curve analysis (DCA), the threshold probability of the nomogram ranged from 1 to 99% (training cohort) and 1 to 75% (validation cohort). Conclusion We successfully developed an ML model for predicting AHD in patients undergoing cervical spine surgery, showcasing its potential to support clinicians in AHD identification and enhance perioperative treatment strategies.

BackgroundVessels encapsulating tumor clusters (VETC) is a newly described vascular pattern that is distinct from microvascular invasion (MVI) in patients with hepatocellular carcinoma (HCC). Despite its importance, the current pathological diagnosis report does not include information on VETC and hepatic plates (HP). We aimed to evaluate the prognostic value of integrating VETC and HP (VETC-HP model) in the assessment of HCC.MethodsA total of 1255 HCC patients who underwent radical surgery were classified into training (879 patients) and validation (376 patients) cohorts. Additionally, 37 patients treated with lenvatinib were studied, included 31 patients in high-risk group and 6 patients in low-risk group. Least absolute shrinkage and selection operator (LASSO) regression analysis was used to establish a prognostic model for the training set. Harrell’s concordance index (C-index), time-dependent receiver operating characteristics curve (tdROC), and decision curve analysis were utilized to evaluate our model's performance by comparing it to traditional tumor node metastasis (TNM) staging for individualized prognosis.ResultsA prognostic model, VETC-HP model, based on risk scores for overall survival (OS) was established. The VETC-HP model demonstrated robust performance, with area under the curve (AUC) values of 0.832 and 0.780 for predicting 3- and 5-year OS in the training cohort, and 0.805 and 0.750 in the validation cohort, respectively. The model showed superior prediction accuracy and discrimination power compared to TNM staging, with C-index values of 0.753 and 0.672 for OS and disease-free survival (DFS) in the training cohort, and 0.728 and 0.615 in the validation cohort, respectively, compared to 0.626 and 0.573 for TNM staging in the training cohort, and 0.629 and 0.511 in the validation cohort. Thus, VETC-HP model had higher C-index than TNM stage system(p < 0.01).Furthermore, in the high-risk group, lenvatinib alone appeared to offer less clinical benefit but better disease-free survival time.ConclusionsThe VETC-HP model enhances DFS and OS prediction in HCC compared to traditional TNM staging systems. This model enables personalized temporal survival estimation, potentially improving clinical decision-making in surveillance management and treatment strategies.

Development and Validation of a MicroRNA Signature Predictive of Local-Regional Recurrence and Overall Survival after Resection of Pancreatic Ductal Adenocarcinoma (PDAC)

BackgroundPostoperative pulmonary complications (PPCs) after thoracoscopic surgery are common. This retrospective study aimed to develop a nomogram to predict PPCs in thoracoscopic surgery.MethodsA total of 905 patients who underwent thoracoscopy were randomly enrolled and divided into a training cohort and a validation cohort at 80%:20%. The training cohort was used to develop a nomogram model, and the validation cohort was used to validate the model. Univariate and multivariable logistic regression were applied to screen risk factors for PPCs, and the nomogram was incorporated in the training cohort. The discriminative ability and calibration of the nomogram for predicting PPCs were assessed using C-indices and calibration plots.ResultsAmong the patients, 207 (22.87%) presented PPCs, including 166 cases in the training cohort and 41 cases in the validation cohort. Using backward stepwise selection of clinically important variables with the Akaike information criterion (AIC) in the training cohort, the following seven variables were incorporated for predicting PPCs: American Society of Anesthesiologists (ASA) grade III/IV, operation time longer than 180 min, one-lung ventilation time longer than 60 min, and history of stroke, heart disease, chronic obstructive pulmonary disease (COPD) and smoking. With incorporation of these factors, the nomogram achieved good C-indices of 0.894 (95% confidence interval (CI) [0.866–0.921]) and 0.868 (95% CI [0.811–0.925]) in the training and validation cohorts, respectively, with well-fitted calibration curves.ConclusionThe nomogram offers good predictive performance for PPCs after thoracoscopic surgery. This model may help distinguish the risk of PPCs and make reasonable treatment choices.

Develop and evaluate a machine learning (ML) model based on CT radiomics for the identification of non-tuberculous mycobacterial lung disease (NTM-LD) and pulmonary tuberculosis (PTB). Retrospectively, chest CT images with NTM-LD and PTB patients confirmed at Medical Center 1 between January 2019 to December 2024 were collected. The dataset was divided into a training cohort and a validation cohort in a 7:3 ratio. Additionally, patients from medical center 2 were collected for external test. A radiomics model was constructed using five machine learning algorithms: Logistic Regression (LR), Random Forest (RF), Quadratic Discriminant Analysis (QDA), YeoJohnson_LR, and YeoJohnson_LR (LDA). Receiver operating characteristic (ROC) and area under the curve (AUC) were used to evaluate the diagnostic efficacy of the five models, and the optimal prediction model was obtained. The optimal model was compared with three radiologists in the testing cohort. A total of 1,512 cases were included, including 1,407 cases from Center 1 (NTM-LD: 547; PTB: 860) and 105 patients from Center 2 (NTM-LD: 32; PTB: 73). Patients in the NTM-LD group were significantly older than those in the PTB group (p < 0.001). There was a significant gender difference between the NTM-LD group and the PTB group (p = 0.005). By comparing the five models, it was found that the YeoJohnson_LR (LDA) model was the best-performing prediction model, with an accuracy of 0.8286 in the external test. The AUCs of the YeoJohnson_LR (LDA) model on the training, validation, and test cohort were 0.8421, 0.8037, and 0.8233, respectively. In comparison with radiologists, the YeoJohnson_LR (LDA) model demonstrated gains of 3.12 ~ 15.62% in sensitivity and 6.85 ~ 12.33% in specificity. The YeoJohnson_LR (LDA) model can be used to distinguish NTM-LD from PTB, assisting in rapid clinical diagnosis and benefiting patients with NTM-LD.

Predictive model for microvascular invasion of hepatocellular carcinoma among candidates for either hepatic resection or liver transplantation