Training Set Research Articles

Radiomics-based multimodality prediction of left atrial appendage thrombus using coronary computed tomography angiography

Abstract Background Timely detection of thrombus in the left atrial appendage (LAA) may help guide decision-making. However, existing models have challenges in extracting comprehensive image information. It has been shown that radiomics-based prediction models can be used for accurate identification. This study investigates potential radiomic markers and constructs a highly efficient multimodality prediction model for LAA thrombus in patients undergoing coronary computed tomography angiography (CCTA). Methods We retrospectively enrolled patients with nonvalvular atrial fibrillation (NVAF) who underwent CCTA and transesophageal echocardiography (TEE) from May 2015 to May 2020. These patients were classified into thrombus or non-thrombus groups based on TEE findings. The data were balanced by resampling and assigned to training or test sets. Radiomics was conducted on the extracted features, and a random forest algorithm was employed for feature selection and importance ranking. We constructed different multimodality thrombus prediction models and compared their performance using the area under the receiver operating characteristic curve (AUC) and other efficacy parameters. Results Among 670 patients (60.17±10.98 years, 70.1% male), 5.2% (n=35) had LAA thrombi. A total of 1232 radiomics features were extracted, with 25 features selected using a random forest for modeling. Two radiomics features (Ibp_3D_m1_firstorder Skewness and original_shape_Flatness) ranked the highest. The radiomics-based multimodality prediction model achieved an AUC of 0.964 (95% confidence interval [CI]: 0.947-0.982), an accuracy of 0.926 (95% CI: 0.925-0.926), and an F1 score of 0.924, outperforming other traditional prediction models. Decision curve analysis also indicated that this model provided the best net clinical benefit. Conclusions Radiomics facilitates a comprehensive exploration of predictors associated with LAA thrombi. A radiomics-based multimodality prediction model can significantly enhance the efficiency of non-invasively detecting thrombi.Receiver operating characteristic curveDecision curve analysis

Multimodal machine learning-based technical success prediction in patients undergoing transcatheter aortic valve implantation

Abstract Background Procedural success is fundamental to achieving optimal clinical outcome, and prediction of procedural success and failure is of key importance in the selection of the appropriate treatment strategy in patients with severe aortic stenosis requiring aortic valve replacement. Purpose We aimed to develop a prediction model for procedural success in patients undergoing transcatheter aortic valve implantation (TAVI) using multimodal information and employing machine learning algorithms. Method In a prospective TAVI registry, patient baseline demographics, clinical history, medication, laboratory, procedural characteristics, electrocardiogram, angiography, transthoracic and transesophagel echocardiography, and computed tomography information were collected prospectively and extracted as numeric data. Data preprocessing included the imputation of missing values and the elimination of features with high collinearity. We incorporated several feature selection (FS) and classification algorithms for model development. Datasets for each target were randomly split into train/validation sets (70/10%) and hold-out test sets (20%). All machine learning processes, including preprocessing, feature selection, parameter, and hyperparameter optimization, were carried out on training and validation sets in 10-fold cross-validation. Different metrics, including area under the curve (AUC) in receiver operating characteristic curve analyses, sensitivity, and specificity, were calculated for model evaluation on the holdout test set. Procedural success was defined as technical success in the Valve Academic Research Consortium-3 (VARC-3). Technical failure differentiated between vascular and cardiac complications. Results In a total of 2,576 patients undergoing TAVI with a contemporary device between March 2012 and June 2023, 242 (9.4%) had technical failure: 70 (2.7%) had cardiac and 178 (6.9%) had vascular technical failure. The sensitivity and specificity for overall technical failure were 0.71 and 0.77, respectively, with AUC of 0.76 (95% CI: 0.69-0.84) achieved by the recursive feature elimination and AdaBoost algorithm. For cardiac technical failure, the ANOVA FS and Random Forest (RF) model yielded sensitivity and specificity of 0.71 and 0.77, with an AUC of 0.76 (95% CI: 0.65 to 0.87). Vascular technical failure showed AUC of 0.76 (95% CI: 0.69 to 0.84), with sensitivity and specificity of 0.72 and 0.73, respectively, with achieved using the ANOVA FS and RF model. Conclusion The initial results of the novel prediction model for technical success are promising. Further improvements and external validation are warranted to support the use of the proposed model in clinical practice.Central Illustration

Preventing unnecessary echocardiograms in de novo patients referred to the cardiology outpatient clinic using electrocardiogram-based deep learning

Abstract Introduction The growing demand on European echocardiography labs, driven by the rise in structural heart disease (SHD), is further complicated by a shortage of skilled sonographers, necessitating resource optimisation strategies. Electrocardiogram (ECG)-based deep learning models may assist in streamlining the echocardiography workflow. Purpose This study sought to develop a deep learning model capable of identifying patients at very low risk of SHD in a population of de novo patients referred to the cardiology outpatient clinic. Methods We identified all adult patients who underwent a 12-lead ECG and echocardiogram at a Dutch academic teaching hospital between 2011 and August 2023. Echocardiogram outcomes were extracted from reports, including left ventricular systolic dysfunction (LVSD), left ventricular dilatation (LVD), aortic stenosis (AS), aortic regurgitation (AR), mitral regurgitation (MR), and tricuspid regurgitation (TR), from which a composite outcome label was derived. A cohort of patients newly referred to our cardiology outpatient clinic with a primary complaint of chest pain, dyspnea, palpitations, dizziness, a murmur, or an abnormal ECG was held out as a test set, while the remaining dataset was split patient-wise into training (90%) and validation (10%) sets. A convolutional neural network was trained using only the ECG tracing as input and the composite label as outcome. Results We included 92,210 ECG-echo pairs from 44,785 patients (19,717 female) with 25,543 (27.7%) abnormal ECG-echo pairs. Among 2724 newly referred patients in the test set, 210/1389 males (15.1%) and 201/1335 females (15.1%) were diagnosed with SHD. In the test set, the model for predicting composite outcome demonstrated an area-under-the-curve, negative predictive value, and specificity of 0.74 (95% confidence interval [CI] 0.71-0.79), 0.94 (95% CI 0.92-0.95), and 0.47 (95% CI 0.45-0.49), respectively. Conclusion This study demonstrates the feasibility of employing deep learning models based solely on electrocardiogram (ECG) data to identify patients at very low risk of structural heart disease (SHD). Further validation and prospective studies are warranted to confirm the generalizability and clinical utility of our findings in diverse patient populations and healthcare settings.

AngioPy: an open-source, deep learning-driven tool for coronary artery segmentation

Abstract Background Quantitative coronary angiography (QCA) permits the objective assessment of coronary artery disease. However, traditional edge detection algorithms often struggle with accurate vessel segmentation due to the complexity and variable quality of angiographic images. As a result, manual correction is frequently needed, which is a major limitation of current QCA systems. Purpose We developed angioPy, a deep learning-driven tool for vessel segmentation that employs user-defined ground truth points to minimise manual correction. We compared its performance without correction with an established QCA system. Methods angioPy was developed using 2455 images from the Fractional Flow Reserve versus Angiography for Multivessel Evaluation 2 (FAME 2) study. Deep learning models were constructed using the U-Net architecture with one of three popular backbones for image classification: ResNet101, DenseNet121, and InceptionResNet-v2. Critically, the model was constructed to integrate the user’s clinical expertise through the selection of a 5-10 ground-truth points along the length of the vessel of interest including the desired start and end point of vessel segmentation. These ground truth points were added to a separate channel of the input image. Five-fold cross validation was performed using proportions of 3:1:1 for training, validation, and test sets, respectively. External validation was performed on a separate dataset of 580 images not used for model development. For both datasets, ground truth segmentation masks were provided by interventional cardiologists using specialised in-house software. A separate analysis was performed in a cohort of 74 patients with mild/moderate stenoses that underwent vessel segmentation using both angioPy and an established commercial QCA package (Medis QFR®). Vessel dimensions at stenoses (proximal and distal reference, minimal luminal diameter) were extracted from both segmentation approaches and compared (609 diameters in total). Results The best performing model performed segmentation with an average F1 score of 0.927 (pixel accuracy 0.998, precision 0.925, sensitivity 0.930, specificity 0.999) with 99.2% of masks exhibiting an F1 score > 0.8 (Figure 1). Simliar results were achieved in the external validation set (F1 score 0.924, pixel accuracy 0.997, precision 0.921, sensitivity 0.929, specificity 0.999). Stenosis dimensions measured with angioPy (mean diameter 3.54 ±1.15 mm) exhibited excellent agreement with QCA (r=0.95 [95% CI 0.95-0.96], p<0.001; mean difference -0.18 mm [limits of agreement: -0.84 to 0.49]) (Figure 2). Minimal luminal diameter (mean 2.93 ±0.91 mm) also exhibited strong agreement (r=0.93 [95% CI 0.91-0.95], p<0.001; mean difference -0.06 mm [limits of agreement: -0.70 to 0.59]. Conclusion angioPy performs rapid and accurate coronary segmentation without the need for manual correction. Available as an open-source tool, angioPy has the potential to increase the efficiency and reliability of QCA.Figure 1.angioPy segmentationFigure 2.angioPy vs QCA (Medis)

Fully automated detection of anomalous aortic origin of the coronary arteries in coronary CT images: a novel artificial intelligence-based screening tool

Abstract Background Anomalous Aortic Origin of the Coronary Artery (AAOCA) is a rare congenital disease that can be easily overlooked during contrast-enhanced Coronary Computed Tomography Angiography (CCTA) image analysis. Objective In this study, we aimed to develop a novel artificial intelligence-based screening tool to automate the detection of AAOCA in CCTA images. Method We enrolled 127 patients from the coronary artery anomaly registry from two centers with AAOCA diagnosis based on CCTA as well as 413 normal cases. A deep learning model was developed to automatically segment the aorta and the left ventricle (LV), after which the images were automatically cropped to the aortic root and LV intersection with a box size of 6 cm3. The images were then clipped to Hounsfield Units ranging from -400 to 750 and normalized to -1 to1. These cropped regions were fed into different residual convolutional neural networks (ResNet, with different depths 10, 18, 34 and 50) to detect AAOCA. Model development (parameter and hyperparameters optimization) was performed on the training and validation sets (70/10%), and finally, it was tested on the 20% untouched test set. Various classification metrics were reported for the detection of AAOCA. Results ResNet-10 achieved an AUC of 0.91 (sensitivity: 0.89, specificity: 0.86), while ResNet-18 showed improved performance with an AUC of 0.95 (sensitivity: 0.89, specificity: 0.95). ResNet-34 achieved an AUC of 0.94 (sensitivity: 0.83, specificity: 0.96), and ResNet-50 excelled with the highest AUC of 0.96 (sensitivity: 0.89, specificity: 0.95), indicating its superior capability in accurately identifying AAOCA. Conclusion We developed a high-performance AI tool for the fully automated detection of AAOCA using CCTA images. This tool can be used to evaluate large patient cohorts and detect AAOCA in extensive databases. Moreover, it can operate seamlessly alongside clinical assessments of CCTA scans, providing real-time alerts to medical personnel about potential AAOCA findings. This enhances diagnostic efficiency and supports early intervention in cardiovascular care.

Sex-specific models to predict 5-year mortality after ST-elevation myocardial infarction using machine learning: insight from FAST-MI registry

Abstract Background Most prognostic stratification tools in acute myocardial infarction (AMI) have been derived from populations including both women and men and including only a small proportion of women. Purpose Using supervised machine learning (ML), we examined the performance of 2 models assessing 5-year mortality after AMI, one derived from a population of women only and one from a population of men only, to determine whether sex-specific models improved outcome prediction. Methods This cohort study used the French registry on acute ST-elevation and non-ST-elevation myocardial infarction (FAST-MI) 2010 and 2015 surveys. This multicentric registry led in more than 200 French hospitals, enrolled all consecutive patients with acute myocardial infarction during a 1-month recruitment period. Our analysis included all men and women presenting STEMI who underwent invasive coronary angiography (ICA). To build and validate the models, the data set were split with a 70%/30% ratio into training and testing sets. The primary outcome was 5-year all-cause mortality. Then, 52 clinical, laboratory, ECG, echocardiographic, and ICA parameters were evaluated for feature selection using Boruta algorithm. Different supervised machine learning algorithms, including random forest (RF), were assessed for model building, and their performance were compared in women and men. To compare prediction accuracy according to sex, we evaluate the performance of the women-based ML model in men from the same cohort. Results 1,189 consecutive women and 3,685 men with STEMI (mean age 61±13 and 69±15 years, respectively) were recruited; 12% of men and 20% of women experienced 5-year all-cause mortality. Using Boruta algorithm, the 10 most important variables for prediction were selected (Figure 1). For women-based ML model building, the RF algorithm exhibited the best performance to predict mortality with an area under the receiver-operating characteristic curve (ROC-AUC) of 0.82 (95% CI: 0.77 – 0.88), an area under the precision-recall curve (PR-AUC) of 0.59 (95% CI: 0.54 – 0.64); and a F1-score of 0.58. The women-based ML model exhibited lower performance in men (ROC-AUC: 0.78; PR-AUC: 0.43; Figure 2). Conversely, the men-based model exhibited better accuracy in men than in women. Conclusion In a large multicentric cohort of STEMI patients, a women-based ML-model exhibited a good accuracy to predict 5-year all-cause mortality in women, but a drop of accuracy when applied to men. In mirror, the accuracy of the men-based model was lower in women. This suggests that sex-specific models might be superior to general models to predict mortality after AMI.Features selection and ML model buildingML model performance according to sex

Predicting atrial fibrillation within 7 years follow-up using machine learning models

Abstract Background The prevalence of atrial fibrillation (AF), the most common type of cardiac arrhythmia, has increased globally. AF is also the leading cardiac factor for several cardiovascular and cerebrovascular events. Early prediction and treatment of AF prevents major cardiovascular and cerebrovascular events. Purpose Our objective is to predict AF risk within 7-years follow-up using only twelve basic factors and applying machine learning (ML) models. Methods Data from the German epidemiological trial on ankle brachial index (getABI) were utilized in this work. GetABI includes 6,457 patients with mean age of 72.53 years. 3,985 individuals were assessed for AF of which 393 patients suffered from AF at 7-years follow-up. The feature selection was initially done manually, ending up in 45 features. The 10 best features were extracted from this dataset according to the SHAP (SHapley Additive exPlanations) values. The final features were body mass index, sensitive CRP, cholesterol, LDL-cholesterol, troponin I, NT-pro-BNP, glomerular filtration rate, ratio of triglycerides and HDL-cholesterol, gamma-GT, heart rate, including age and sex as major factors. The handling of missing values was achieved by SimpleImputer, estimating the mean values. XGBoost and Random Forest predicted AF. Logistic Regression was also used as a baseline model. The data were split into training and test set and ML models were applied for 100 iterative runs, extracting results with mean values. Results XGBoost was the most accurate model, predicting the AF with 71.19 % mean accuracy, 66.37 % mean sensitivity and 75.61 % mean specificity. Figure 1 depicts the performance of all models. Additionally, Receiver Operating Characteristic Curve (ROC) and Area Under Curve (AUC) for each model are presented in Figure 2. XGBoost also outperformed all the others models, achieving the highest mean AUC value, equal to 70.99 %. Contrariwise, logistic regression had the lowest overall performance. Conclusion ML models, specifically XGBoost, achieved accurate AF prediction in 3,985 individuals utilizing only twelve clinical-routine data.Evaluation results of each used model.ROC curve and AUC value of each model.

Predictive slope stability early warning model based on CatBoost

A model for predicting slope stability is developed using Categorical Boosting (CatBoost), which incorporates 6 slope features to characterize the state of slope stability. The model is trained using a symmetric tree as the base model, utilizing ordered boosting to replace gradient estimation, which enhances prediction accuracy. Comparative models including Support Vector Machine (SVM), Light Gradient Boosting Machine (LGBM), Random Forest (RF), and Logistic Regression (LR) were introduced. Five performance evaluation metrics are utilized to assess the predictive capabilities of the CatBoost model. Based on CatBoost model, the predicted probability of slope instability is calculated, and the early warning model of slope instability is further established. The results suggest that the CatBoost model demonstrates a 6.25% disparity in accuracy between the training and testing sets, achieving a precision of 100% and an Area Under Curve (AUC) value of 0.95. This indicates a high level of predictive accuracy and robust ordering capabilities, effectively mitigating the problem of overfitting. The slope instability warning model offers reasonable classifications for warning levels, providing valuable insights for both research and practical applications in the prediction of slope stability and instability warning.

External validation of a machine learning based classification algorithm for ambulatory heart rhythm diagnostics using smartphone-photoplethysmography - the SMARTBEATS algorithm study

Abstract Introduction Smartphone-photoplethysmography (PPG) can be used for heart rhythm monitoring. According to current guidelines, ECG verification is required for a new diagnosis of atrial fibrillation (AF). Accurate machine learning (ML) based automatic algorithms have the potential to facilitate clinical applications generating large numbers of PPG recordings, by offloading the need for manual over-read. Purpose The aim of this study was to evaluate the diagnostic performance of an automatic ML-based algorithm for heart rhythm diagnostics using smartphone-PPG recorded by patients with AF in an unsupervised ambulatory setting. Methods Patients undergoing direct current cardioversion (DCCV) at a university hospital for treatment of AF or atrial flutter (AFL) were asked to perform one-minute heart rhythm recordings post-cardioversion at least twice daily for 30 days. All participants were provided with an iPhone 7 smartphone running the CORAI Heart Monitor PPG application simultaneously with a single-lead ECG recording (KardiaMobile). The automatic algorithm uses support vector machines (SVM) to classify heart rhythm from smartphone-PPG recordings. The SVM classifier evaluated here was trained on ambulatory PPG recordings made by patients in a separate cardioversion cohort, ensuring complete independence between training and validation sets. PPG recordings in the validation cohort were analyzed by the automatic algorithm and diagnostic performance was calculated by comparing heart rhythm classification output from the SVM classifier to the heart rhythm diagnosis from the simultaneous ECG recordings. ECG recordings were interpreted independently by two cardiologists and were set as gold standard. Results The SVM classifier was trained on data from 153 patients with 11,749 PPG recordings, of which 5,873 (50.0%) were labelled as AF and 5,876 (50.0%) as sinus rhythm (SR), as assessed by manual reading. In the validation cohort, 280 patients, with a median age of 69.0 years (31% women), registered 18,005 simultaneous PPG and ECG recordings. Recordings interpreted as AFL on ECG (2.0%), as having insufficient quality on ECG (4.9%) or PPG (2.8%), and low certainty algorithmic classifications (1.2%) were excluded, leaving 16,057 PPG recordings. Included ECG recordings consisted of 71.2% interpreted as (SR) and 28.8% as AF. Algorithm classification of the PPG recordings in the validation cohort diagnosed AF (sensitivity) in 99.7% and SR (specificity) in 99.7% of the recordings, with an overall accuracy of 99.7%. F1-score was 99.4% and area under the ROC curve (AUC) was 0.999. Conclusion A machine learning based algorithm for automatic heart rhythm diagnostics showed excellent diagnostic performance for smartphone-PPG recordings in an unsupervised ambulatory setting, minimizing the need for ECG verification in cardioversion populations.

High-throughput prediction of oral acute toxicity in Rat and Mouse of over 100,000 polychlorinated persistent organic pollutants (PC-POPs) by interpretable data fusion-driven machine learning global models

This study utilized available oral acute toxicity data in Rat and Mouse for polychlorinated persistent organic pollutants (PC-POPs) to construct data fusion-driven machine learning (ML) global models. Based on atom-centered fragments (ACFs), the collected high-throughput data overcame the applicability limitations, enabling accurate toxicity prediction for a wide range of PC-POPs series compounds using only single models. The data variances in the Rat training and test sets were 1.52 and 1.34, respectively, while for the Mouse, the values were 1.48 and 1.36, respectively. Genetic algorithm (GA) was used to build multiple linear regression (MLR) models and pre-screen descriptors, addressing the “black-box” problem prevalent in ML and enhancing model interpretability. The best ML models for Rat and Mouse achieved approximately 90 % prediction reliability for over 100,000 true untested compounds. Ultimately, a warning list of highly toxic compounds for eight categories of polychlorinated atom-centered fragments (PCACFs) was generated based on the prediction results. The analysis of descriptors revealed that dioxin analogs generally exhibited higher toxicity, because the heteroatoms and ring systems increased structural complexity and formed larger conjugated systems, contributing to greater oral acute toxicity. The present study provides valuable insights for guiding the subsequent in vivo tests, environmental risk assessment and the improvement of global governance system of pollutants.

Plaque phenotyping using deep learning in optical coherence tomography imaging

Abstract Background Artificial intelligence (AI) techniques have emerged as powerful tools in various computer vision tasks, particularly in classification, detection, and segmentation. Purpose This study aimed to develop a deep learning model for categorizing optical coherence tomography (OCT) frames into different plaque phenotypes, including thin-cap fibroatheroma (TCFA) and fibroatheroma. Methods A prospective multicenter observational study enrolled 168 patients who underwent percutaneous coronary intervention (PCI) and received OCT imaging for non-target lesions with angiographically estimated diameter stenosis greater than 30% (177 vessels). Among them, 129 patients (135 vessels) underwent follow-up OCT imaging at 1 year. A total of 45,072 OCT frames were randomly divided into training and validation sets in a 4:1 ratio. An EfficientNet-B3 convolutional neural network model was developed to classify OCT frames into TCFA, thick-cap fibroatheroma, non-fibroatheroma, adaptive intimal thickening, and normal vessel categories. Results A confusion matrix comparing manually assessed plaque phenotypes with AI-predicted phenotypes was presented in Table. In the validation datasets (9,015 frames), the overall accuracy of the model was 85.1%. For predicting the presence of TCFA, sensitivity was 81.1%, specificity was 98.8%, positive predictive value was 63.9%, and negative predictive value was 99.5%. Regarding the presence of fibroatheroma, sensitivity was 90.3%, specificity was 89.4%, positive predictive value was 93.1%, and negative predictive value was 85.4%. Conclusions The deep learning algorithm demonstrated high accuracy in detecting TCFA and fibroatheroma with excellent reproducibility. This automated process enables operators to promptly identify high-risk lesions during PCI, potentially leading to improved clinical outcomes.

Cross-validatory Z-Residual for Diagnosing Shared Frailty Models

Accurate model performance assessment in survival analysis is imperative for robust predictions and informed decision-making. Traditional residual diagnostic tools like martingale and deviance residuals lack a well-characterized reference distribution for censored regression, making numerical statistical tests based on these residuals challenging. Recently, the introduction of Z-residuals for diagnosing survival models addresses this limitation. However, concerns arise from conventional methods that utilize the entire dataset for both model parameter estimation and residual assessment, which may cause optimistic biases. This paper introduces cross-validatory Z-residuals as an innovative approach to address these limitations. Employing a cross-validation (CV) framework, the method systematically partitions the dataset into training and testing sets to reduce the optimistic bias. Our simulation studies demonstrate that, for goodness-of-fit tests and outlier detection, cross-validatory Z-residuals are significantly more powerful (e.g. power increased from 0.2 to 0.6). and more discriminative (e.g. AUC increased from 0.58 to 0.85) than Z-residuals without CV. We also compare the performance of Z-residuals with and without CV in identifying outliers in a real application that models the recurrence time of kidney infection patients. Our findings suggest that cross-validatory Z-residuals can identify outliers, which Z-residuals without CV fail to identify. The CV Z-residual is a more powerful tool than the No-CV Z-residual for checking survival models, particularly in goodness-of-fit tests and outlier detection. We have published a generic function, which is collected in an R package called Zresidual, for computing CV Z-residual for the output of the widely used survival R package.

DELINEATE-Regurgitation: deep learning for automated assessment of aortic, mitral, and tricuspid regurgitation from echocardiography

Abstract Background Valvular heart diseases, including aortic regurgitation (AR), mitral regurgitation (MR), and tricuspid regurgitation (TR) are significant contributors to cardiovascular morbidity and mortality worldwide. Artificial intelligence (AI) may improve echocardiographic assessment, but no models to date have successfully evaluated valvular regurgitation. Purpose The goal of this study was to develop an AI system to improve the accuracy of echocardiographic AR, MR, and TR interpretation. Methods Complete transthoracic echocardiogram (TTE) studies performed at a single center were split into train (n=65,301), validation (14,018), and test sets (13,746). The Deep Learning for Echo Analysis, Tracking, and Evaluation of Valvular Regurgitation (DELINEATE-Regurgitation) system intakes entire TTE studies, identifies color Doppler clips showing each valve, and yields a study level classification of AR, MR, and TR on a 6-grade scale (none/trace, mild, mild-moderate, moderate, moderate-severe, severe) using the cardiologist interpretation as a reference standard. This system used a hybrid neural network, leveraging the spatiotemporal feature learning capabilities of convolutional neural networks with the sequential processing strengths of transformer networks. The end-to-end hybrid approach allows the model to learn nuanced representations of the data including features within and across color Doppler videos. Model accuracy was assessed using quadratically weighted Cohen’s kappa (k), area under the receiver operator characteristic curve (AUROC) for the detection of both "moderate or greater" and "severe" regurgitation, and AI-cardiologist agreement the 6-grade scale. Additionally, an analysis was conducted using a 4-grade regurgitation scale (none/trace, mild, moderate, and severe) by rolling up intermediate classes (mild-moderate and moderate-severe) into the next higher class for both cardiologist determinations and model predictions. Results The DELINEATE-Regurgitation system demonstrated a high accuracy in the classification of AR (k=0.857), MR (k=0.867), and TR (k=0.847) (Figure 1). DELINEATE predicted the same severity as cardiologists (AR 91.8%, MR 79.9%, TR 77.7%) or within ±1 grade accuracy (AR 99.5%, MR 98.6%, TR 98.6%). The models had excellent binary discrimination in the detection of moderate or greater (AR AUROC 99.4, MR 98.6, TR 98.1) and severe (AR 99.7, MR 99.3, TR 99.6) regurgitation (Figure 2). In the 4-grade scale, the model had exact agreement with cardiologists in 92.3% of AR, 82.4% of MR, and 80.6% of TR cases. Panel adjudication of AI-cardiologist disagreement and correlation with quantitative metrics at higher levels of regurgitation will be presented. Conclusions The DELINEATE-Regurgitation AI system demonstrates a high level of accuracy in the assessment of AR, MR, and TR using color Doppler TTE videos. Future work will focus on automating quantification and developing strategies for optimal clinical deployment.Figure 1 - Confusion MatricesFigure 2 - ROC Curves

Individual Tree Crown Detection and Classification of Live and Dead Trees Using a Mask Region-Based Convolutional Neural Network (Mask R-CNN)

Mapping the distribution of living and dead trees in forests, particularly in ecologically fragile areas where forests serve as crucial ecological environments, is essential for assessing forest health, carbon storage capacity, and biodiversity. Convolutional neural networks, including Mask R-CNN, can assist in rapid and accurate forest monitoring. In this study, Mask R-CNN was employed to detect the crowns of Casuarina equisetifolia and to distinguish between live and dead trees in the Pingtan Comprehensive Pilot Zone, Fujian, China. High-resolution images of five plots were obtained using a multispectral Unmanned Aerial Vehicle. Six band combinations and derivatives, RGB, RGB-digital surface model (DSM), Multispectral, Multispectral-DSM, Vegetation Index, and Vegetation-Index-DSM, were used for tree crown detection and classification of live and dead trees. Five-fold cross-validation was employed to divide the manually annotated dataset of 21,800 live trees and 7157 dead trees into training and validation sets, which were used for training and validating the Mask R-CNN models. The results demonstrate that the RGB band combination achieved the most effective detection performance for live trees (average F1 score = 74.75%, IoU = 70.85%). The RGB–DSM combination exhibited the highest accuracy for dead trees (average F1 score = 71.16%, IoU = 68.28%). The detection performance for dead trees was lower than for live trees, which may be due to the similar spectral features across the images and the similarity of dead trees to the background, resulting in false identification. For the simultaneous detection of living and dead trees, the RGB combination produced the most promising results (average F1 score = 74.18%, IoU = 69.8%). It demonstrates that the Mask R-CNN model can achieve promising results for the detection of live and dead trees. Our study could provide forest managers with detailed information on the forest condition, which has the potential to improve forest management.

Evaluating XGBoost’s predictive accuracy on surgical site Infections in cardiac surgery

Abstract Background Surgical site infections (SSIs) pose a significant threat in cardiac surgery (CS), profoundly impacting patient prognosis. SSI incidence varies widely, reported between 3.5% and 26.8%. The study used a decade-long dataset from SSI surveillance in a specialized hospital in Italy, applying machine learning (ML) techniques to predict the SSI incidence among CS patients. Methods Data collected from 2013-2023 through the surveillance system of SSI in patients undergoing CS were used to train a predictive ML algorithm (XGBoost). Data included information on demographics, risk factors, surgery variables (incision site, prophylaxis, etc.), and the infection outcome derived from follow-up interviews at 30 days post-surgery, or 90 days for patients with prosthetic materials, according to the ECDC case definition of SSI. We used the R libraries “caret”, “smotefamily” and “xgboost” to train the algorithm. A train-test split of 70-30 was applied. Both downsampling of the majority class (no SSI) and oversampling of SSI cases with SMOTE were used in the training set to address class imbalance. Results A total of 10,534 subjects (65.9% males, mean age 68.3 years) who underwent CS (64.3% with prosthetic materials) were included, among which 533 SSIs were identified (mean incidence of 5.06%), with 430 cases (80.7%) occurring after discharge (38.7% deep SSI) and 103 (19.3%) before discharge (55.5% deep SSI). The trained XGBoost algorithm achieved an AUC of 0.62, sensitivity of 69%, and specificity of 50% in the prediction of SSI on the test dataset. Conclusions These findings suggest that while the XGBoost model provides a fair predictive capability, there is significant room for improvement in sensitivity and specificity. The use of artificial intelligence offers a promising opportunity to identify patients at risk of SSI. However, further research and comprehensive data are needed to refine the predictive model and effectively improve prevention measures. Key messages • Predominantly occurring post-discharge, SSIs in cardiac surgery necessitate enhanced prevention strategies, highlighting the critical phase after hospital discharge where surveillance is paramount. • Artificial intelligence may enhance post-discharge SSI surveillance, promising to identify patients at increased risk. Further research is needed to refine these AI tools for optimal sensitivity.

Prediction of microvascular invasion in hepatocellular carcinoma with conventional ultrasound, Sonazoid-enhanced ultrasound, and biochemical indicator: a multicenter study

PurposeTo develop and validate a preoperative prediction model based on multimodal ultrasound and biochemical indicator for identifying microvascular invasion (MVI) in patients with a single hepatocellular carcinoma (HCC) ≤ 5 cm.MethodsFrom May 2022 to November 2023, a total of 318 patients with pathologically confirmed single HCC ≤ 5 cm from three institutions were enrolled. All of them underwent preoperative biochemical, conventional ultrasound (US), and contrast-enhanced ultrasound (CEUS) (Sonazoid, 0.6 mL, bolus injection) examinations. Univariate and multivariate logistic regression analyses on clinical information, biochemical indicator, and US imaging features were performed in the training set to seek independent predictors for MVI-positive. The models were constructed and evaluated using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis in both validation and test sets. Subgroup analyses in patients with different liver background and tumor sizes were conducted to further investigate the model’s performance.ResultsLogistic regression analyses showed that obscure tumor boundary in B-mode US, intra-tumoral artery in pulsed-wave Doppler US, complete Kupffer-phase agent clearance in Sonazoid-CEUS, and biomedical indicator PIVKA-II were independently correlated with MVI-positive. The combined model comprising all predictors showed the highest AUC, which were 0.937 and 0.893 in the validation and test sets. Good calibration and prominent net benefit were achieved in both sets. No significant difference was found in subgroup analyses.ConclusionsThe combination of biochemical indicator, conventional US, and Sonazoid-CEUS features could help preoperative MVI prediction in patients with a single HCC ≤ 5 cm.Critical relevance statementInvestigation of imaging features in conventional US, Sonazoid-CEUS, and biochemical indicators showed a significant relation with MVI-positivity in patients with a single HCC ≤ 5 cm, allowing the construction of a model for preoperative prediction of MVI status to help treatment decision making.Key PointsMVI status is important for patients with a single HCC ≤ 5 cm.The model based on conventional US, Sonazoid-CEUS and PIVKA-II performs best for MVI prediction.The combined model has potential for preoperative prediction of MVI status.Graphical

Asphalt pavement raveling recognition based on deep learning algorithms

Abstract Pavement raveling can lead to the development of issues such as potholes. Monitoring and providing real-time early warnings for road surface developments is crucial for timely maintenance planning and effectively controlling the occurrence and progression of surface issues. To achieve intelligent identification of raveling distress on asphalt pavement surfaces, we utilized a portable laser-based pavement surface scanning device to collect raveling images, followed by histogram equalization to eliminate noise caused by environmental factors. Additionally, data enhancement strategies such as rotation and mirroring were used to enrich training data and expand the sample size. A Convolutional Neural Network (CNN) was employed to perform multi-layer convolution and pooling operations on the images to quickly extract features of distress. Finally, five pre-trained transfer learning models—ResNet-18, DenseNet, VGG16, Inception, and EfficientNet—were introduced to identify the asphalt surface disease data and the normal surface data. The results showed that the accuracy and loss of the training sets of several models exhibited a positive trend, with the EfficientNet model demonstrating stable and superior performance under the same training conditions. The precision matrix, confusion matrix, precision, recall, and F1 score were used as evaluation indexes to assess the effectiveness of the models for identifying pavement raveling. The EfficientNet and Inception models had the best accuracy and comprehensive performance, with accuracy rates of 0.96 and 0.95, respectively. The precision, recall, and F1 score are 0.95, 0.96, and 0.96 for the EfficientNet model respectively. A total of 146 images correctly predicted raveling disease and only 4 normal images were incorrectly predicted as diseased. For example, the EfficientNet model’s identification accuracy for pockmark disease on the PC platform was 0.963, verifying the model’s feasibility in practical applications. The principal contribution of this paper lies in the efficient and precise identification of asphalt pavement raveling, which adversely affects vehicular traveling comfortability. This provides theoretical and technical support for the intelligent identification of asphalt pavement defects and enhancing road asset maintenance management in the pavement engineering industry.

Development of a nomogram-based model incorporating radiomic features from follow-up longitudinal lung CT images to distinguish invasive adenocarcinoma from benign lesions: a retrospective study

PurposeTo develop and validate a radiomic model for differentiating pulmonary invasive adenocarcinomas from benign lesions based on follow-up longitudinal CT images.MethodsThis is a retrospective study including 336 patients (161 with invasive adenocarcinomas and 175 with benign lesions) who underwent baseline (T0) and follow-up (T1) CT scans from January 2016 to June 2022. The patients were randomized in a 7:3 ratio into training and test sets. Radiomic features were extracted from lesion volumes of interest on longitudinal CT images at T0 and T1. Differences in radiomic features between T1 and T0 were defined as delta-radiomic features. Logistic regression was used to build models based on clinicoradiological (CR), T0, T1, and delta radiomic features and compute signatures. Finally, a nomogram based on the CR, T0, T1 and delta signatures was constructed. Model performance was evaluated for calibration, discrimination, and clinical utility.ResultsThe T1 radiomic model was superior to the other independent models. In the training set, it had an area under the curve (AUC) of 0.858), superior to the CR model (AUC 0.694), the T0 radiomic model (AUC 0.825), and the delta radiomic model (AUC 0.734). In the test set, it had an AUC of 0.817, again outperforming the CR model (AUC 0.578), the T0 radiomic model (AUC 0.789), and the delta radiomic model (AUC 0.647). The nomogram incorporating the CR, T0, T1 and delta signatures showed the best predictive performance in both the training (AUC: 0.906) and test sets (AUC: 0.856), and it exhibited excellent fit with calibration curves. Decision curve analysis provided additional validation of the clinical utility of the nomogram.ConclusionA nomogram utilizing radiomic features extracted from longitudinal CT images can enhance the discriminative capability between pulmonary invasive adenocarcinomas and benign lesions.

Predictors of irreversible renal dysfunction in patients with idiopathic retroperitoneal fibrosis.

Idiopathic retroperitoneal fibrosis (iRPF) can lead to irreversible kidney damage. This study aimed to investigate predictors of irreversible renal dysfunction in patients with iRPF. Eighty-three patients with newly diagnosed iRPF were enrolled between January 2010 and Sep 2022 at Zhongshan Hospital of Fudan University, including 60 in the training set and 23 in the validation set. They were regularly contacted or followed up via outpatient examinations by specialist doctors, who documented their condition and treatment progress. Predictors of irreversible renal dysfunction were identified using univariate and multivariate regression, logistic model, and receiver operating curve analyses. In the training set, over a median follow-up of 29 months, 16.7% of patients had an estimated glomerular filtration rate (eGFR) of < 60 ml/min/1.73m2 at the last follow-up, and 25% had hydronephrosis or required prolonged double-J stents. A prognostic score was developed by assigning 1, 1, and 2 points for peripheral CD19+ B cells <9.3%, serum creatinine (sCr) ≥120 μmol/l, and no response at 6 months, respectively. A score of ≥ 2 for predicting irreversible renal dysfunction had sensitivity and specificity of 100% and 92%, respectively. In the validation set, 21.7% of patients suffered from irreversible renal dysfunction. The sensitivity and specificity for predicting irreversible renal dysfunction were 100% and 94.4%, respectively. A prognostic score based on factors including CD19+ B cells <9.3% and sCr ≥120 μmol/l at baseline, and no response at 6 months, is suitable for predicting irreversible renal dysfunction in iRPF.

Construction and validation of a line chart for gestational diabetes mellitus based on clinical indicators

BackgroundGestational diabetes mellitus (GDM) is a common complication of mid–to-late pregnancy. Here, we constructed a predictive model for GDM based on a combination of clinical characteristics and relevant serum markers.MethodsData from full-term singleton vaginal deliveries from January 2022 to January 2023 were retrospectively collected from the obstetrics department. The data collected were segregated and assigned to training, validation, and external test sets. Maternal demographic characteristics, living and working habits, and haematological indicators, such as liver function and lipids were collected using a questionnaire designed for the study. The “rms” package in R was used to explore GDM-associated factors through stepwise regression at P < 0.05. A predictive model was developed based on the results of multifactorial logistic regression analysis. We then evaluated the differentiation of the column-line graphical model and performed internal and external validation. To assess the accuracy of the bar graphical model, we plotted calibration and decision curves.ResultsData from 265 pregnant women were included in the training and internal validation sets, and data from 113 pregnant women were included in the external validation set. The logistic regression algorithm screened 8 indicators as predictors. A prediction model was constructed with ALT, TBA, TC, and TG levels while considering whether GDM affects appetite, the husband– wife relationship, family history, and parental relationships as predictors. The Hosmer–Lemeshow goodness-of-fit test revealed that the chi-square values for the modelling, internal validation, and external validation groups (χ2 = 5.964, 3.249, and 12.182, respectively) were all P > 0.05. The ROC curve AUCs for the three groups were 0.93 (95% CI: 0.89–0.97), 0.72 (95% CI: 0.62–0.81), and 0.68 (95% CI: 0.53–0.83), respectively.ConclusionIn this study, a GDM prediction model was constructed to achieve high performance in GDM risk prediction based on routine obstetric tests and information.