Predicting ischemic stroke risk from atrial fibrillation based on multi-spectral fundus images using deep learning

Background Electrocardiographic predictors of atrial fibrillation (AF) from a non-AF ECG–such as p wave abnormalities and supraventricular ectopy–have been extensively documented. However, risk prediction tools for AF utilize little if any of the wealth of information available from the ECG. Better AF prediction from the ECG may improve efficiency of screening and performance of AF risk tools. Deep learning methods have the potential to extract an unlimited number of features from the ECG to improve prediction of AF. Purpose We hypothesize that a deep learning model can identify patterns predictive of AF during normal sinus rhythm. To test the hypothesis, we trained and tested a neural network to predict AF from normal sinus rhythm ambulatory ECG data. Methods We trained a deep convolutional neural network to detect features of AF that are present in single-lead ECGs with normal sinus rhythm, recorded using a Food and Drug Administration (FDA)-cleared, smartphone-enabled device. A cohort of 27,526 patients with at least 50 ECGs recorded between January 7, 2013, and September, 19, 2018, and the FDA-cleared automated findings of Normal and Atrial Fibrillation associated with those ECGs, were used for model development. Specifically, we trained the deep learning model on 1,984,581 Normal ECGs from 19,267 patients with 1) only Normal ECG recordings, or 2) at least 30% ECGs with AF. Of the 27,526 patients, an internal set of 8,259 patients with 841,776 Normal ECGs was saved for testing (validation). Results Among 8,259 patients in the test set, 3,467 patients had at least 30% of their ECGs with an automated finding of AF. When the deep learning model was run on 841,776 Normal ECGs, it was able to predict whether the ECG was from a patient with no AF or with 30% or more AF, with an area under the curve (AUC) of 0.80. Using an operating point with equal sensitivity and specificity, the model's sensitivity and specificity were 73.1%. Using an operating point with high specificity (90.0%), the model's sensitivity was 48.0%. When the model was applied to a randomly-selected, broader cohort of 15,000 patients (at least 50 ECGs recorded, any amount of AF), a positive, non-linear relationship between neural network output and AF burden per patient was observed (Figure). Model Output vs AF Burden Per Patient Conclusions A deep learning model was able to predict AF from ECGs in normal sinus rhythm that were recorded on a smartphone-enabled device. The use of deep learning, if prospectively validated, may facilitate AF screening in patients with paroxysmal disease or warn patients who are at high risk for developing AF. Acknowledgement/Funding AliveCor

Assess the efficacy of deep convolutional neural networks (DCNNs) in detection of critical enteric feeding tube malpositions on radiographs. 5475 de-identified HIPAA compliant frontal view chest and abdominal radiographs were obtained, consisting of 174 x-rays of bronchial insertions and 5301 non-critical radiographs, including normal course, normal chest, and normal abdominal x-rays. The ground-truth classification for enteric feeding tube placement was performed by two board-certified radiologists. Untrained and pretrained deep convolutional neural network models for Inception V3, ResNet50, and DenseNet 121 were each employed. The radiographs were fed into each deep convolutional neural network, which included untrained and pretrained models. The Tensorflow framework was used for Inception V3, ResNet50, and DenseNet. Images were split into training (4745), validation (630), and test (100). Both real-time and preprocessing image augmentation strategies were performed. Receiver operating characteristic (ROC) and area under the curve (AUC) on the test data were used to assess the models. Statistical differences among the AUCs were obtained. p < 0.05 was considered statistically significant. The pretrained Inception V3, which had an AUC of 0.87 (95 CI; 0.80–0.94), performed statistically significantly better (p < .001) than the untrained Inception V3, with an AUC of 0.60 (95 CI; 0.52–0.68). The pretrained Inception V3 also had the highest AUC overall, as compared with ResNet50 and DenseNet121, with AUC values ranging from 0.82 to 0.85. Each pretrained network outperformed its untrained counterpart. (p < 0.05). Deep learning demonstrates promise in differentiating critical vs. non-critical placement with an AUC of 0.87. Pretrained networks outperformed untrained ones in all cases. DCNNs may allow for more rapid identification and communication of critical feeding tube malpositions.

Background A universal screening strategy for atrial fibrillation (AF) is not established despite its rising incidence. Emerging consumer and medical-grade artificial intelligence (AI) electrocardiography (ECG) devices hold promise for AF screening. Latest research suggests that AI can extract clinically relevant features from sinus rhythm (SR)-ECG, potentially identifying patients nearing an AF diagnosis. This positions AI as a novel tool for proactive AF screening that could enhance early detection and intervention. Purpose To assess the effectiveness of AI algorithms in detecting patients approaching an AF diagnosis using SR-ECG and analyse the factors influencing their performance. Methods A systematic search was conducted in PubMed, EMBASE, Scopus, Web of Science, Google Scholar, ClinicalTrials.gov, European Union Clinical Trials Register, and International Clinical Trials Registry Platform. To enhance inclusivity, references of included studies were also included. Eligible studies were published from January 2014 to May 2024, used AI to detect AF patients from SR-ECGs in patients without prior AF history, tested the model on unseen data, and confirmed AF via traditional methods within one year maximum. Summarized performance metrics were compared using the Mann-Whitney U test with subgroup analyses by AI type and AF confirmation timeframe. Study quality was assessed with QUADAS-2. Results The search yielded 3568 articles, of which 14 studies describing 33 AI models were included. A total of 4,182,953 ECGs from 1,459,653 patients were analysed by the models. Median model performance for AF detection using SR-ECG was: accuracy 58.0%, sensitivity 62.0%, specificity 57.8%, precision 52.0%, and area under the curve (AUC) 0.740. Classical machine learning was used by 12 (36.4%) models and deep learning by 21 models (63.6%), of which 13 used a convolutional neural network. Deep learning outperformed traditional machine learning in accuracy (83.0% vs 57.0%; p=0.027), sensitivity (72.6% vs 54.5%; p=0.009), and AUC (0.830 vs 0.610; p&amp;lt;0.001). The confirmatory AF event was captured within a maximum of 31 days by 13 (29.4%) models, and within a 31-day to one-year period by 20 (60.6%) models. Models using a 31-day confirmation window showed higher accuracy (83.2% vs 56.0%; p=0.001) and AUC (0.851 vs 0.630; p&amp;lt;0.001) than those using a one-year timeframe. A formal meta-analysis was deemed unsuitable due to considerable methodological heterogeneity among studies. Conclusions Deep learning algorithms and shorter confirmation windows demonstrate improved AF identification performance, highlighting timely follow-up. Expanding AI usage through accessible platforms may enhance ECG screening in underserved areas, addressing health disparities. Further research is needed to assess the clinical benefits and cost-effectiveness of such widespread application.Performance by AI typePerformance by confirmation timeframe

Classification of Bisyllabic Lexical Stress Patterns Using Deep Neural Networks

Computer-aided diagnosis of ocular disorders like glaucoma can be effectively performed with retinal fundus images. Since the advent of machine learning and later deep learning techniques in medical image analysis, research focusing on automated and early detection of glaucoma has gained prominence. In this paper, we show how a deep Convolutional Neural Network model can be used in visualising the features related to Glaucoma that are consistent with (a subset of) those used by ophthalmologists for diagnosis. However, since the doctors do not necessarily depend entirely on these features but back them up with other important measurements such as visual field testing (HVF), any deep neural network such as CNN may not necessarily be expected to yield 100% accuracy. Further, other important parameters such as Intraocular Pressure (IOP), which may be built into a deep learning model, may not necessarily correlate well with the presence of Glaucoma. In this work, a deep CNN model has been trained and tested on a large number of high-quality fundus images, both normal and glaucomatous. We compare the results with transfer learning models such as VGG16, ResNet50, and MobileNetV2. On an average, we obtained an accuracy of 93.75% in identifying glaucoma, focusing only on the features of the fundus image.

Heartbeat classification is a crucial step for arrhythmia diagnosis during electrocardiographic (ECG) analysis. The new scenario of wireless body sensor network- (WBSN-) enabled ECG monitoring puts forward a higher-level demand for this traditional ECG analysis task. Previously reported methods mainly addressed this requirement with the applications of a shallow structured classifier and expert-designed features. In this study, modified frequency slice wavelet transform (MFSWT) was firstly employed to produce the time-frequency image for heartbeat signal. Then the deep learning (DL) method was performed for the heartbeat classification. Here, we proposed a novel model incorporating automatic feature abstraction and a deep neural network (DNN) classifier. Features were automatically abstracted by the stacked denoising auto-encoder (SDA) from the transferred time-frequency image. DNN classifier was constructed by an encoder layer of SDA and a softmax layer. In addition, a deterministic patient-specific heartbeat classifier was achieved by fine-tuning on heartbeat samples, which included a small subset of individual samples. The performance of the proposed model was evaluated on the MIT-BIH arrhythmia database. Results showed that an overall accuracy of 97.5% was achieved using the proposed model, confirming that the proposed DNN model is a powerful tool for heartbeat pattern recognition.

Deep learning is being incorporated in many modern software systems. Deep learning approaches train a deep neural network (DNN) model using training examples, and then use the DNN model for prediction. While the structure of a DNN model as layers is observable, the model is treated in its entirety as a monolithic component. To change the logic implemented by the model, e.g. to add/remove logic that recognizes inputs belonging to a certain class, or to replace the logic with an alternative, the training examples need to be changed and the DNN needs to be retrained using the new set of examples. We argue that decomposing a DNN into DNN modules— akin to decomposing a monolithic software code into modules—can bring the benefits of modularity to deep learning. In this work, we develop a methodology for decomposing DNNs for multi-class problems into DNN modules. For four canonical problems, namely MNIST, EMNIST, FMNIST, and KMNIST, we demonstrate that such decomposition enables reuse of DNN modules to create different DNNs, enables replacement of one DNN module in a DNN with another without needing to retrain. The DNN models formed by composing DNN modules are at least as good as traditional monolithic DNNs in terms of test accuracy for our problems.

Background Combining the artificial intelligence algorithm with the known clinical risk factors may provide enhanced accuracy for prediction of the hidden atrial fibrillation (AF) in patients with embolic stroke with undetermined source (ESUS). Purpose We aimed to develop enhanced prediction models for AF with deep learning algorithm (DLA) and clinical predictors in patients with ESUS. The DLA was created to identify the patients with paroxysmal AF based on their electrocardiograms (ECG) during sinus rhythm. Methods We analyzed the 221 patients who underwent insertable cardiac monitor (ICM) for AF detection following ESUS. The DLA was previously developed with sinus rhythm ECGs of 10,605 paroxysmal AF patients and 50,522 non-AF patients. The convolutional neural network was used for the DLA. The primary endpoint was defined as any AF episode lasting over 5 min by ICM. The atrial ectopic burden (AEB) was calculated as the percentage of the number of conducted QRS by atrial ectopy on Holter monitoring. Results AF (≥5 min) was detected by ICM in 32 patients (14.5%) during follow-up period of 15.1±8.6 months. AF patients had higher AEB (0.199% vs 0.023%, p&amp;lt;0.001), larger left atrial diameters (LAD, 41.2 mm vs 35.7 mm, p&amp;lt;0.001), and larger left atrial volume index (LAVI, 46.4 ml/m2 vs 32.3 ml/m2, p&amp;lt;0.001) than those without AF. The means of calculated probabilities of AF by DLA were higher in patients with AF than those without AF (63.8% vs 40.2%, p&amp;lt;0.001). In the receiver operating characteristic curve analysis, the areas under the curve (AUC) were the highest in DLA (0.824) followed by AEB (0.784), LAVI (0.780), and LAD (0.768). The multivariable model with AEB, LAVI, and DLA demonstrated excellent prediction accuracy for paroxysmal AF (AUC: 0.902, Figure 1) Conclusions In patients with ESUS, the DLA outperformed other clinical risk factors for prediction of AF. Combining DLA with AEB, LAD and LAVI could is a potential useful tool to predict AF in ESUS patients. Funding Acknowledgement Type of funding sources: None.

Peanut leaf diseases have a major impact on peanut yield and quality. Timely, rapid, and accurate early diagnosis and control of peanut leaf diseases are key to ensuring high quality and yield of peanuts. This work focuses on the early diagnosis of peanut diseases and pests and conducts systematic research on the hardware system for imaging and spectral sensing of peanut plant leaves, as well as the software for deep learning classification algorithms. First, we designed a system that can separately obtain multispectral reflectance and fluorescence images and collect multispectral images of three asymptomatic peanut leaf diseases, including scab, scorch spot, and anthracnose. Second, we constructed a convolutional neural network to extract the basic features of spectral images. Third, an adaptive channel attention mechanism is introduced to update the weights of different channels. Fourth, a sparse second-order attention mechanism driving network is constructed to enhance the discriminative ability of deep feature information. Finally, the classification is completed utilizing the Softmax classifier. The experimental results demonstrate that the spectral image information improves the robustness of deep learning models to data transformation and achieves a high-precision classification score of 98.45% for asymptomatic peanut leaf diseases. Compared to traditional optical devices and software algorithms, the proposed multispectral imaging system and deep learning algorithm significantly improve detection ability and classification accuracy, which can assist botanists in making more accurate diagnoses of peanut leaf diseases.

Deep neural networks (DNNs) are increasingly deployed in various fields. Despite their spectacular advances, DNNs are known to suffer from adversarial vulnerabilities. The robustness of DNNs is then threatened by leading them to misclassifications with unexpected inputs (adversarial examples). The fuzzing technique frequently used for testing traditional software has recently been adopted to evaluate the robustness of DNNs. Current DNN fuzzing techniques focus on image classification DNNs and generate test cases by mutations, e.g., image transformations and adversarial perturbations. However, mutation-based test cases usually lack diversity and have distribution deflection from the original DNN input space, which impacts the evaluation of DNNs. In this paper, we propose a generation-based fuzzing frame-work FuzzGAN to detect adversarial flaws existing in DNNs. We integrate the testing purpose and the guidance of the neuron coverage into the original objectives of auxiliary classifier generative adversarial networks. Hence, FuzzGAN learns the representation of a DNN's input space and generates test cases without the limitation of any concrete seed input. We evaluate the performance of FuzzGAN on two DNN models that have classical network structures and are trained on public datasets. The experiment results demonstrate that FuzzGAN can generate realistic, diverse and valid test cases and achieve high neuron coverage. Moreover, these test cases can be used to improve the performance of the target DNN through adversarial retraining.

Develop a predictive model utilizing weakly supervised deep learning techniques to accurately forecast major pathological response (MPR) in patients with resectable non-small cell lung cancer (NSCLC) undergoing neoadjuvant chemoimmunotherapy (NICT), by leveraging whole slide images (WSIs). This retrospective study examined pre-treatment WSIs from 186 patients with non-small cell lung cancer (NSCLC), using a weakly supervised learning framework. We employed advanced deep learning architectures, including DenseNet121, ResNet50, and Inception V3, to analyze WSIs on both micro (patch) and macro (slide) levels. The training process incorporated innovative data augmentation and normalization techniques to bolster the robustness of the models. We evaluated the performance of these models against traditional clinical predictors and integrated them with a novel pathomics signature, which was developed using multi-instance learning algorithms that facilitate feature aggregation from patch-level probability distributions. Univariate and multivariable analyses confirmed histology as a statistically significant prognostic factor for MPR (P-value< 0.05). In patch model evaluations, DenseNet121 led in the validation set with an area under the curve (AUC) of 0.656, surpassing ResNet50 (AUC = 0.626) and Inception V3 (AUC = 0.654), and showed strong generalization in external testing (AUC = 0.611). Further evaluation through visual inspection of patch-level data integration into WSIs revealed XGBoost's superior class differentiation and generalization, achieving the highest AUCs of 0.998 in training and robust scores of 0.818 in validation and 0.805 in testing. Integrating pathomics features with clinical data into a nomogram yielded AUC of 0.819 in validation and 0.820 in testing, enhancing discriminative accuracy. Gradient-weighted Class Activation Mapping (Grad-CAM) and feature aggregation methods notably boosted the model's interpretability and feature modeling. The application of weakly supervised deep learning to WSIs offers a powerful tool for predicting MPR in NSCLC patients treated with NICT.

Introduction Paroxysmal atrial fibrillation (AF) is detected in approximately one-third of patients diagnosed with cryptogenic stroke undergoing prolonged ECG monitoring. Several risk factors for AF have been described, including left atrium (LA) enlargement, LA deformation, age, cardiovascular risk factors, and biomarkers as N-pro-BNP and troponin. Specifically, LA enlargement is a direct marker of atrial cardiopathy, which predisposes to AF. However, the methods to calculate LA enlargement vary. While LA biplane volume indexed to body surface area (BSA) remains the standardized method of choice according to current guidelines, it may underestimate the degree of enlargement in overweight patients. Additionally, overweight produces a systemic inflammatory state which is also associated with AF. Purpose The primary objective of the study was to define the best parameter among LA end systolic volume (LAESV), LAESV indexed to BSA (LAESV/BSA), and LAESV indexed to ideal BSA for identifying patients admitted for cryptogenic stroke with undetected paroxysmal AF. The secondary objective was to assess the accuracy of these parameters in both normal-weight and overweight patients. Methods Consecutive patients over 65 years-old diagnosed with cryptogenic stroke in our tertiary hospital were enrolled. All patients underwent 15-days ambulatory ECG monitoring to screen for undetected paroxysmal AF. Clinical, echocardiographic and laboratory data were prospectively collected. Ideal BSA was calculated using the Devine method. The accuracy of LA volume parameters in classifying patients with undetected paroxysmal AF was assessed using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. The analysis was stratified for overweight status (BMI &amp;lt; 25 or BMI ≥ 25). Results A total of 66 patients (Mean age 78.4 ± 7.6 years; 48% male) were included. Of them, 39 (59%) had hypertension, 16 (24%) had diabetes and 39 (59%) were overweight. AF was present in 21 (32%) patients. The mean LAESV, LAESV/BSA, LAESV/ideal BSA were 56.2 ± 23.0 ml, 34.2 ± 12.6 ml/m2, and 34.7 ± 14.1 ml/m2, respectively. Figure 1 illustrates the ROC curves for LA volume parameters in the main cohort and stratified by overweight status. Overall, no relevant differences in AUC were observed among LAESV, LAESV/BSA, LAESV/ideal BSA in each scenario. However, BMI significantly altered the discrimination performance of LA volume measures, which resulted optimal in normal-weight patients (AUC from 0.838 to 0.886) and absent in overweight patients (AUC from 0.838 to 0.521 to 0.564). Conclusions LAESV, LAESV/BSA, LAESV/ideal BSA show similar discrimination performance in identifying undetected AF in patients with cryptogenic stroke. However, the presence of overweight considerably impacts on the predictive value of LA volume parameters in this clinical scenario.ROC curves

Pineoblastomas (PBs) are rare central nervous system tumors primarily affecting children and adolescents, with limited data on clinical characteristics and survival outcomes. Prognosis prediction models for this disease are lacking. The purpose of this study was to develop deep learning (DL) models for predicting 3-year survival in patients with pineoblastoma. Patients with pineoblastomas of all ages were identified from the Surveillance, Epidemiology, and End Results (SEER) database (1975-2019). Deep neural networks (DNN) were trained and tested at a ratio of 7:3 in a 5-fold cross-validated fashion. Multivariate CPH models were constructed for comparison. The primary outcomes were 3-year overall survival (OS) and disease-specific survival (DSS). All the variables were included in the analysis. Receiver operating characteristic (ROC) curve analysis and calibration plots were used to evaluate the model performance. A total of 145 patients were included in this study. The area under the curve (AUC) for the DNN models was 0.92, 0.91, and 0.749 for OS and 0.76 for DSS. The DNN models exhibited good calibration: the OS model (slope = 0.94, intercept = 0.07) and DSS model (slope = 0.81, intercept = 0.20). Our DNN models showed a more accurate prediction of survival outcomes in patients with pineoblastoma than the widely used CPH models. These results indicate the potential of DL algorithms to improve outcome prediction in patients with rare tumors.

Left atrial appendage thrombus (LAAT) may be detected by transesophageal echocardiography (TOE) in patients with atrial fibrillation (AF) despite continuous anticoagulation therapy. We examined the factors predisposing to LAAT in patients treated with the anticoagulants dabigatran and rivaroxaban. We retrospectively evaluated 1,256 AF patients from three centres who underwent TOE before electrical cardioversion (n = 611, 51.4%) or catheter ablation (n = 645, 48.6%) from January 2013 to December 2019 and had been on at least three weeks of continuous dabigatran (n = 603, 48%) or rivaroxaban (n = 653, 52%) therapy. Preprocedural TOE diagnosed LAAT in 51 patients (4.1%), including 30 patients (5%) treated with dabigatran and 21 patients (3.2%) treated with rivaroxaban (p=0.1145). In multivariate logistic regression, predictors of LAAT in patients treated with dabigatran were non-paroxysmal AF (vs. paroxysmal AF) (OR = 6.2, p=0.015), heart failure (OR = 3.22, p=0.003), and a eGFR <60 ml/min/1.73 m2 (OR = 2.65, p=0.012); the predictors in patients treated with rivaroxaban were non-paroxysmal AF (vs. paroxysmal AF) (OR = 5.73, p=0.0221) and heart failure (OR = 3.19, p=0.116). In ROC analysis of the dabigatran group, the area under the curve (AUC) for the CHA2DS2-VASc-RAF score was significantly higher (0.78) than those for the CHADS2, CHA2DS2-VASc, and R2CHADS2 scores (0.67, 0.70, and 0.72, respectively). In the rivaroxaban group, the CHA2DS2-VASc-RAF score also performed significantly better (AUC of 0.77) than the CHADS2, CHA2DS2-VASc, and R2CHADS2 scores (AUC of 0.66, 0.64, and 0.67, respectively). The risk of LAAT was the same for patients in both treatment groups. In all patients, non-paroxysmal AF or heart failure, and in patients treated with dabigatran an eGFR <60 ml/min/1.73 m2, were independent predictors of LAAT. The new CHA2DS2-VASc-RAF scale had the highest predictive value for LAAT in the entire study population.

This research examines the efficacy of ensemble Machine Learning (ML) models, mainly focusing on Deep Neural Networks (DNNs), in predicting the need for cardiovascular surgery, a critical aspect of clinical decision-making. It addresses key challenges such as class imbalance, which is pivotal in healthcare settings. The research involved a comprehensive comparison and evaluation of the performance of previously published ML methods against a new Deep Learning (DL) model. This comparison utilized a dataset encompassing 50,000 patient records from a large hospital between 2015-2022. The study proposes enhancing the efficacy of these models through feature selection and hyperparameter optimization, employing techniques like grid search. A novel aspect of this research was the comparison of a newly developed DNN model with existing ensemble models based on similar cardiovascular datasets. The results indicated the DNN model's superior predictive accuracy, demonstrating an Area Under the Curve (AUC) of 74%, alongside notable precision (68%) and recall (72%) for the minority class, which indicates patients requiring surgery. The model further achieved a 70% F1-Score and a balanced accuracy rate of 72%, significantly outperforming the existing ensemble models in every key performance metric. The study underscores the transformative potential of DNNs in predictive modeling for cardiovascular care and highlights the importance of integrating advanced ML techniques into clinical workflows. Future research should delve into the practical application and integration of these models.