Convolutional Neural Network Model Research Articles

Robust Multi-Class Classification for Real-Time Agricultural Applications Using Efficient and Adaptive Deep Learning

Plant diseases severely affect agricultural productivity, necessitating accurate and rapid detection methods. This research presents a robust, multi-class plant disease classification framework using adaptive deep learning. We utilize pre-trained convolutional neural networks (CNNs), specifically Xception, InceptionResNetV2, InceptionV3, ResNet50, and the proposed EfficientNetB3-based Adaptive Augmented Deep Learning (EfficientNetB3-AADL) model. Our approach leverages transfer learning combined with extensive data augmentation and trimming techniques to enhance model performance and mitigate overfitting. The EfficientNetB3-AADL architecture incorporates convolutional and max pooling layers, regularization strategies, and a dense feature learning layer, optimized to classify 52 disease categories from a publicly available leaf image dataset. The model’s performance is extensively evaluated using metrics such as accuracy, precision, recall, and F1 score. Notably, EfficientNetB3-AADL achieves superior accuracy over 98%, outperforming other CNN models. The proposed methodology highlights the efficacy of compound scaling and adaptive data augmentation in ensuring robust and efficient disease classification, suitable for real-time agricultural applications. This advancement supports sustainable farming by offering a scalable, computationally efficient solution for early and accurate disease detection in diverse crop species.

Single Well Production Prediction Model of Gas Reservoir Based on CNN-BILSTM-AM

In the prediction of single-well production in gas reservoirs, the traditional empirical formula of gas reservoirs generally shows poor accuracy. In the process of machine learning training and prediction, the problems of small data volume and dirty data are often encountered. In order to overcome the above problems, a single-well production prediction model of gas reservoirs based on CNN-BILSTM-AM is proposed. The model is built by long-term and short-term memory neural networks, convolutional neural networks and attention modules. The input of the model includes the production of the previous period and its influencing factors. At the same time, the fitting production and error value of the traditional gas reservoir empirical formula are introduced to predict the future production data. The loss function is used to evaluate the deviation between the predicted data and the real data, and the Bayesian hyperparameter optimization algorithm is used to optimize the model structure and comprehensively improve the generalization ability of the model. Three single wells in the Daniudi D28 well area were selected as the database, and the CNN-BILSTM-AM model was used to predict the single-well production. The results show that compared with the prediction results of the convolutional neural network (CNN) model, long short-term memory neural network (LSTM) model and bidirectional long short-term memory neural network (BILSTM) model, the error of the CNN-BILSTM-AM model on the test set of three experimental wells is reduced by 6.2425%, 4.9522% and 3.0750% on average. It shows that on the basis of coupling the empirical formula of traditional gas reservoirs, the CNN-BILSTM-AM model meets the high-precision requirements for the single-well production prediction of gas reservoirs, which is of great significance to guide the efficient development of oil fields and ensure the safety of China’s energy strategy.

An intelligent spectral identification approach for the simultaneous detection of endocrine-disrupting chemicals in aquatic environments.

With the rapid progression of industrialization, the application and release of endocrine disruptors (EDCs), including bisphenol A (BPA), octylphenol and nonylphenol have significantly increased, presenting substantial health hazards. Conventional analytical techniques, such as high-performance liquid chromatography and gas chromatography-mass spectrometry, are highly sophisticated but suffer from complex procedures and high costs. To overcome these limitations, this study introduces an innovative spectral methodology for the simultaneous detection of multiple aquatic multicomponent EDCs. By leveraging chemical machine vision, specifically with convolutional neural network (CNN) models, we employed a long-path holographic spectrometer for rapid, cost-effective identification of BPA, 4-tert-octylphenol, and 4-nonylphenol in aqueous samples. The CNN, refined with the ResNet-50 architecture, demonstrated superior predictive performance, achieving detection limits as low as 3.34, 3.71 and 4.36 μg/L, respectively. The sensitivity and quantification capability of our approach were confirmed through the analysis of spectral image Euclidean distances, while its universality and resistance properties were validated by assessments of environmental samples. This technology offers significantly advantages over conventional techniques in terms of efficiency and cost, offering a novel solution for EDC monitoring in aquatic environments. The implications of this research extend beyond improved detection speed and cost reduction, presenting new methodologies for analyzing complex chemical systems and contributing to environmental protection and public health.

Machine learning-based prediction of soil organic matter via smartphone

Color is an important property of soil that indicates soil composition and fertility. Soil organic matter (SOM) of darker color soils, which have rich humus and minerals, is higher than others such as red soils. The aim of this study is to construct models to predict SOM for a range of colors using a smartphone as an image-capturing device. Random forest of classification (RFC), random forest of logical regression (RFLR), convolutional neural network (CNN) and MobileNet models are compared, which is better for SOM prediction. Soil photos were collected by smartphone in a camera obscura with a steady light. After treatment by OpenCV, photos were separated by SOM content into five groups. The prediction accuracies of RFC, RFLR, CNN and MobileNet were 0.9743, 0.5614, 0.9600 and 0.7915, respectively. Based on these results, the RFC model for SOM has the best performance both in training and validation. The proposed combination of smartphone and machine learning-based prediction models provide a fast, economic, and robust approach to monitor, detect, and predict SOM contents in precision and intelligence agriculture.

Evaluation of Electrical Properties and Uniformity of Single Wall Carbon Nanotube Dip-Coated Conductive Fabrics Using Convolutional Neural Network-Based Image Analysis

This study proposes a convolutional neural network (CNN)-based image analysis method to evaluate the electrical properties and uniformity of conductive fabrics treated with single-walled carbon nanotube (SWCNT) dip-coating. The conductive fabric was produced by dip-coating cotton-blended spandex with SWCNT, and the surface images were scanned and preprocessed to obtain image data, while resistance measurements were conducted to obtain labels and build the dataset. SEM analysis revealed that as the number of dip-coating cycles increased, particle density and path formation improved. The CNN model learned the relationship between surface images and resistance values, achieving a high predictive performance, with an R-squared (R²) value of 0.9422. The model demonstrated prediction accuracies of 99.1792% for the coefficient of variation (CV) of uniformly coated fabrics and 96.8877% for non-uniformly coated fabrics. Additionally, p-value analysis of all fabric samples yielded a result of 0.96044, indicating no statistically significant difference between the predicted and actual values. The proposed CNN-based model can accurately evaluate the electrical uniformity of conductive fabrics, showing potential for contributing to quality control and process optimization in production.

Process–Material–Performance Trade-off Exploration of Materials Sintering with Machine Learning Models

AbstractProcess-induced porosity, defects, and residual stresses lead to mechanical performance degradation in fiber-reinforced composite and other heterogeneous structures. Physical and chemical processes create complex process–material–performance relationships. Predicting porosity and residual stresses in this context requires computationally burdensome forward simulations and obtaining optimal process settings and calibrating properties of new materials requires solving inverse problems with predictions from the forward simulations. In this paper, we parameterized the process–material–performance space and created a dataset based on physics models that are valid for sintering ceramic powders. The dataset was used to train several machine learning models that captured the process–material–performance relationships. The trained ML models were applied in process optimization, calibration of properties for new material systems, and estimating performance for a given process and material. Support vector regression (SVR), convolutional neural networks (CNNs), and a Gaussian mixture model (GMM) called REPAIRS were selected, and their prediction accuracy was determined. While the SVR and CNN models require training several models, we show that the GMM model captures the process–material–performance relationships with a single machine-learned model and partial system completion methods. The paper describes root-mean-square error and mean absolute percentage errors of the inferences from the models on a validation dataset.

Deep learning-based automated measurement of hip key angles and auxiliary diagnosis of developmental dysplasia of the hip.

Anteroposterior pelvic radiographs remains the most widely employed method for diagnosing developmental dysplasia of the hip. This study aims to evaluate the accuracy of an artificial intelligence model in measuring angles in pelvic radiographs of the hip. The assessment seeks to demonstrate the efficacy of the artificial intelligence model in diagnosing both developmental dysplasia of the hip and borderline developmental dysplasia of the hip through the analysis of pelvic radiographs. A total of 1,029 patients, including 273 men and 757 women, were retrospectively included in this study. The anteroposterior pelvic radiographs were randomly divided into three sets: the training set (720 cases), the validation set (103 cases), and the test set (206 cases). Key anatomical points on the anteroposterior pelvic radiographs were identified. The Sharp, Tönnis, and Center Edge angles were calculated automatically based on the corresponding criteria. The hip development status was compared between measurements obtained from the artificial intelligence model and those defined manually by two radiologists. The area under the receiver operating characteristic curve was utilized to assess the diagnostic performance of the artificial intelligence model. The results obtained from both manual measurements and the artificial intelligence model demonstrated no significant differences in the Sharp, Tönnis, and Center edge angles (all p > 0.05). The intra-class correlation coefficients and correlation coefficient r values exceeded 0.75, indicating that both the artificial intelligence model and manual measurements exhibited good repeatability and a positive correlation. Notably, the artificial intelligence model provided measurements more faster than those conducted by radiologists (p = 0.001). The artificial intelligence model also demonstrated high diagnostic accuracy, sensitivity, and specificity for developmental dysplasia of the hip. The performance of the artificial intelligence model in diagnosing developmental dysplasia of the hip was robust. Additionally, the results from the artificial intelligence model and manual measurements were largely consistent with clinical diagnosis results (p = 0.01). The artificial intelligence model can effectively evaluate hip conditions by measuring the Sharp, Tönnis, and Center edge angles, which are consistent closely with clinical diagnosis results. The results of the artificial intelligence model measurements demonstrate a high degree of consistency with those obtained through manual measurements. The angles of Sharp, Tönnis, and Center edge, as evaluated by the deep learning-based convolutional neural network model, exhibit robust diagnostic performance in identifying both developmental dysplasia of the hip and borderline developmental dysplasia of the hip. Consequently, the artificial intelligence model has the potential to fully replace manual measurements of these critical hip angles, providing a more efficient and precise alternative for diagnosing both conditions of the hip.

Sex differences in brain MRI using deep learning toward fairer healthcare outcomes

This study leverages deep learning to analyze sex differences in brain MRI data, aiming to further advance fairness in medical imaging. We employed 3D T1-weighted Magnetic Resonance images from four diverse datasets: Calgary-Campinas-359, OASIS-3, Alzheimer's Disease Neuroimaging Initiative, and Cambridge Center for Aging and Neuroscience, ensuring a balanced representation of sexes and a broad demographic scope. Our methodology focused on minimal preprocessing to preserve the integrity of brain structures, utilizing a Convolutional Neural Network model for sex classification. The model achieved an accuracy of 87% on the test set without employing total intracranial volume (TIV) adjustment techniques. We observed that while the model exhibited biases at extreme brain sizes, it performed with less bias when the TIV distributions overlapped more. Saliency maps were used to identify brain regions significant in sex differentiation, revealing that certain supratentorial and infratentorial regions were important for predictions. Furthermore, our interdisciplinary team, comprising machine learning specialists and a radiologist, ensured diverse perspectives in validating the results. The detailed investigation of sex differences in brain MRI in this study, highlighted by the sex differences map, offers valuable insights into sex-specific aspects of medical imaging and could aid in developing sex-based bias mitigation strategies, contributing to the future development of fair AI algorithms. Awareness of the brain's differences between sexes enables more equitable AI predictions, promoting fairness in healthcare outcomes. Our code and saliency maps are available at https://github.com/mahsadibaji/sex-differences-brain-dl.

ECG-based machine learning model for AF identification in patients with first ischemic stroke.

The recurrence rate of strokes associated with atrial fibrillation (AF) can be substantially reduced through the administration of oral anticoagulants. However, previous studies have not demonstrated a clear benefit from the universal application of oral anticoagulants in patients with embolic stroke of undetermined source. Timely detection of AF remains a challenge in patients with stroke. This study aims to develop a convolutional neural network (CNN) model to accurately identify patients with AF using a 12-lead sinus-rhythm electrocardiogram (ECG) recorded around the time of the first ischemic stroke. Additionally, this study also evaluates the model's ability to predict future occurrence of AF. A CNN model was trained with ECG data from patients at Taipei Veterans General Hospital. External validation was performed on ischemic stroke patients from National Taiwan University Hospital. The model's performance was assessed for detecting AF at the stroke event and predicting future AF occurrences. The model demonstrated an area under curve (AUC) of 0.91 for internal validation and 0.69 for external validation in identifying AF at the stroke event, with sensitivity and negative predictive value both achieving 97%. Kaplan-Meier survival analysis of patients without a prior diagnosis of AF revealed a significant increase in future AF incidence among the high-risk group identified by the model (adjusted hazard ratio: 4.06; 95% confidence interval: 2.74-6.00). The CNN model effectively identifies AF in stroke patients using 12-lead ECGs and predicts future AF events, facilitating early anticoagulation therapy and potentially reducing recurrent stroke risk. Further prospective studies are warranted to confirm these findings.

Abstract 4116410: A Novel Deep Learning Approach for Prediction of Right Heart Failure After Left Ventricular Assist Device Implantation using Pulmonary Artery Pressure Tracings

Introduction: Left ventricular assist device (LVAD) offers a lifeline for advanced heart failure patients, but ~30% experience post-operative right heart failure (RHF) with significant morbidity and mortality. Current methods for predicting RHF risk have limited accuracy. This study is the first of its kind to explore the feasibility of using a convolutional neural network (CNN) deep learning model with pulmonary artery pressure (PAP) tracings acquired during pre-operative hemodynamic assessment to identify patients at risk for early RHF. Methods: A CNN model was pre-trained on PAP tracings from the MIMIC-III Waveform Database to leverage transferable knowledge and improve model performance. The model learned general features and temporal relationships in hemodynamic tracing morphology for RHF prediction. It was then adapted for classification by modifying the output layers. Post-LVAD RHF was definitively adjudicated for each instance included from our 246 patient single center LVAD database. RHF was adjudicated using the 2020 Academic Research Consortium defintion. The database included 205 non-RHF instances and 41 RHF instances. We developed a novel signal processing method utilizing computer vision to extract PAP waveform data from pre-operative right heart catheterization reports. Data was reviewed to ensure accurate pressure tracing representation. SMOTE-ENN (synthetic oversampling) and class weighting addressed dataset imbalance during training to enhance model generalizability and RHF prediction accuracy. Results: The model achieved strong performance in classifying RHF from non-RHF outcomes, as evidenced by consistent AUC scores of 0.87 and 0.85 on validation and unseen test data, respectively. These findings indicate the potential for hemodynamic tracings to predict RHF after LVAD implantation. Notably, the model maintained its performance on unseen data, highlighting its generalizability. Conclusions: This study explored the feasibility of using a deep learning model for RHF risk stratification in patients with LVADs. The model, pre-trained on PAP tracings, achieved promising classification performance assessed by AUC. Further validation with larger, more diverse datasets will refine model performance and enhance generalizability for clinical application. Leveraging hemodynamic tracings within a deep learning framework holds promise for improved clinical outcome prediction.

Abstract 4130756: Two Leads are All We Need: Optimizing Choice of Electrocardiogram Leads for Artificial Intelligence-based Based Detection of Left Ventricular Systolic Dysfunction

Introduction: Artificial intelligence (AI) models built with convolutional neural networks (CNN) accurately detect left ventricular systolic dysfunction (LVSD) from 12-lead electrocardiogram (ECG). While AI models can also detect LVSD from 1-lead (but with lower accuracy), the optimal numbers and combinations of the leads remain unclear. Aims: To identify the optimal numbers and combinations of ECG leads for detecting LVSD. Methods: A total of 75,033 ECGs recorded within 14 days of an echocardiogram were collected. The data was randomly divided into derivation, validation, and test sets in a 5:2:3 ratio without patient overlap. The same split was used throughout the study. While all available ECGs were included in derivation and validation sets, the test set included 1 ECG per patient closest to the echocardiogram. CNN models were trained to detect LVSD, defined as left ventricular ejection fraction <40%, using all possible combinations of 1-, 2-, and 3-lead extracted from 12-lead ECG. The final model for each lead combination was chosen according to the area under the receiver operating curve (AUROC) on the validation set across the 20 epochs. The models were trained 4 times to evaluate the variance caused by initialization vectors. Results: The 12-lead model showed AUROC of 0.893 (95%CI, 0.887-0.899). Multiple models trained with 2-lead achieved similar performance, with aVR-V4 displaying the highest AUROC (0.892; 95%CI, 0.889-0.895), followed by I-V4 (0.889; 95%CI, 0.883-0.895) and I-II (0.888; 95%CI, 0.884-0.893). The results were similar for 3-lead (aVR-aVL-V4: AUROC 0.895; 95%CI, 0.893-0.897, I-aVR-V4: 0.892; 95%CI, 0.889-0.896, and I-II-V4: 0.891; 95%CI, 0.886-0.895). The choice of leads greatly affected the performance. For the 1-lead models, aVR showed the highest AUROC (0.876; 95%CI, 0.873-0.879), but III only achieved an AUROC of (0.766; 95%CI, 0.761-0.772). The combinations, including a chest lead and a limb lead (especially aVR), tend to achieve higher performance (Fig 1). Conclusion: A combination of 2-lead was sufficient to achieve comparable accuracy as the 12-lead for the AI model to detect LVSD. Wearable devices collecting a combination of a limb and a chest lead may enhance the detection of LVSD.

Spatial heterogeneity response of soil salinization inversion cotton field expansion based on deep learning

Soil salinization represents a significant challenge to the ecological environment in arid areas, and digital mapping of soil salinization as well as exploration of its spatial heterogeneity with crop growth have important implications for national food security and salinization management. However, the machine learning models currently used are deficient in mining local information on salinity and do not explore the spatial heterogeneity of salinity impacts on crops. This study developed soil salinization inversion models using CNN (Convolutional Neural Network), LSTM (Long Short-Term Memory Network), and RF (Random Forest) models based on 97 field samples and feature variables extracted from Landsat-8 imagery. By evaluating the accuracy, the best-performing model was selected to map soil salinity at a 30m resolution for the years 2013 and 2022, and to explore the relationship between soil electrical conductivity (EC) values and the expansion of cotton fields as well as their spatial correlation. The results indicate that:(1) The CNN performs best in prediction, with an R2 of 0.84 for the training set and 0.73 for the test set, capable of capturing more local salinity information. (2) The expansion of cotton fields has reduced the level of soil salinization, with the area of severely salinized and saline soils in newly added cotton fields decreasing from 177.91 km2 and 381.46 km2 to 19.49 km2 and 1.12 km2, respectively. (3) Regions with long-term cotton cultivation and newly reclaimed cotton fields exhibit high sensitivity and vulnerability to soil salinity. This study explores the excellent performance of deep learning in salinity mapping and visualizes the spatial distribution of cotton fields that are highly sensitive to soil salinity, providing a scientific theoretical basis for accurate salinity management.

Spatiotemporal feature extraction method combining common spatial pattern and hybrid spatial‐temporal attention‐Net in motor imagery–brain‐computer interface system

AbstractIn recent years, motor imagery (MI) based on brain‐computer interface (BCI) has gained the attention of researchers and been widely used in many fields, such as medical rehabilitation and entertainment. The temporal and spatial features of electroencephalography (EEG) are very important for MI‐BCI classification. This study proposed a better spatiotemporal feature extraction method by combining the common spatial pattern (CSP) with a hybrid spatiotemporal attention convolutional neural network (HSTA‐Net) model. The method first intercepts the MI‐EEG into multiple time windows and projects them into the feature space and feeds into the HSTA‐Net in parallel. Subsequently, the scaled dot product attention module and multi‐head attention module are used to extract spatial feature attention. Finally, the features in multiple time windows obtained through HSTA‐Net are integrated and used for classification. The proposed method achieves a classification accuracy of 77.3% on the BCI Competition IV2a dataset. The experimental results show that the proposed model has better classification performance. The CSP‐HSTA‐Net model can quickly adapt to the distribution of MI‐EEG data during the training process, which increases the interclass distance of the extracted features and decreases its intraclass distance; it can better accomplish the MI‐BCI classification task.

Trading Community Analysis of Countries’ Roll-On/Roll-Off Shipping Networks Using Fine-Grained Vessel Trajectory Data

Roll-on/roll-off vessels (RO/RO vessels) are playing an increasingly critical role in international automobile transport, facilitating the efficient movement of vehicles and heavy machinery across continents. Despite this growing significance, there is still limited research specifically focused on the RO/RO shipping network and its impact on global trade. This paper studies the global RO/RO shipping network using AIS data on RO/RO vessels collected from 2020 to 2023. We construct a method based on the complex network theory and the graph feature extraction method to quantitatively assess the features of the RO/RO shipping network. This method assesses the complexity, sparsity, homogeneity, modularity, and hierarchy of the RO/RO shipping network across various ports and countries and employs the graph convolutional neural network (GCN) model to extract network features for community detection. This process enables the identification of port clusters that are frequently linked to RO/RO vessels, as well as regional transport modes. The paper’s findings support these conclusions: (1) From 2020 to 2023, the number of nodes in the RO/RO shipping network increased by 22%, primarily concentrated in African countries. The RO/RO shipping network underwent restructuring after the pandemic, with major complex network parameters showing an upward trend. (2) The RO/RO shipping network is complex, with a stable graph density of 0.106 from 2020 to 2023. The average degree increased by 7% to 4.224. Modularity decreased by 6.5% from 0.431 in 2022 to 0.403, while the hierarchy coefficient rose to 0.575, suggesting that post-pandemic, community routes have become more diverse, reflecting the reconstruction and maturation of the overall network. (3) The model yielded a silhouette coefficient of 0.548 and a Davies–Bouldin index of 0.559 using an improved automatic feature extraction method. In comparison between 2020 and 2023, the changes in the two indicators are small. This shows that GINs can effectively extract network features and give us results that we can understand for community detection. (4) In 2023, key communities divide the RO/RO shipping network, with one community handling 39% of global routes (primarily Europe–Asia), another community handling 23% (serving Asia–Pacific, Africa, and the Middle East), and a third community managing 38% (linking Asia, Europe, and South America).

An End-to-End Brain Computer Interface System for Mental Workload Estimation through Hybrid Deep Learning Model

AbstractIn this paper, a new cascade one-dimensional convolutional neural network (1DCNN) and bidirectional long short-term memory (BLSTM) model has been developed for binary and ternary classification of mental workload (MWL). MWL assessment is important to increase the safety and efficiency in brain–computer interface (BCI) systems and professions, where multi-tasking is required. Keeping in mind the necessity of MWL assessment, a two-fold study is presented, firstly binary classification is done to classify MWL into low and high classes. Secondly, ternary classification is applied to classify MWL into low, moderate, and high classes. The cascaded1DCNN-BLSTM deep learning architecture has been developed and tested over the Simultaneous task EEG workload (STEW) dataset. Unlike recent research in MWL, handcrafted feature extraction and engineering are not done, rather end-to-end deep learning is used over 14 channel EEG signals for classification. Accuracies exceeding the previous state-of-the-art studies have been obtained. In binary and ternary classification accuracies of 96.77% and 95.36%have been achieved with sevenfold cross-validation, respectively.

Abstract Su803: Sudden Cardiac Arrest Prediction via Deep Learning Electrocardiogram Analysis

Background: Sudden cardiac arrest (SCA) is a potentially fatal event that often occurs without prior indications. Hypothesis: It was hypothesized that electrocardiogram (ECG) data could be used in conjunction with Deep Learning (DL) algorithms to predict the occurrence of SCA’s. Aims: To improve SCA outcomes and enable preventative strategies, the ECG data from the Nightingale Open Science - Subtyping Cardiac Arrest dataset were utilized to create a DL model to serve as a screening tool for sudden cardiac arrest. Methods: A publicly-available dataset containing ten seconds of 12-lead ECGs from individuals who had a SCA and persons who did not have a SCA, and information about time from ECG to arrest, and age and sex were utilized for individualized prediction of SCA. Subsets of the initial data were used to train, validate, and test deep convolution neural network models that processed cardiac cycles from single-lead ECG voltage waveforms. Results: The cohort comprised 221 (mean age: 68.5 and #females: 75) patients who experienced SCA and 1046 controls (mean age: 69.7 and #females: 497) who did not have a cardiac arrest. The SCA ECG-DL base model, with ECGs within one day and with age and sex data, had an area under the receiver operating curve (AUROC) of 0.774 during hold-out testing for discrimination of subsequent SCA. With sensitivity set at 95%, the base model specificity of the SCA ECG-DL model was 31%. Gradient-weighted class activation mapping showed that the SCA ECG-DL model relied most heavily on the QRS complex to make predictions. The performance of the SCA ECG-DL model was similar removing age and sex data (AUROC 0.775) and less reliable when predicting SCA within one month (AUROC 0.734) or between one month and one year (AUROC 0.574). Conclusion: DL interpretation of ECG data is a promising means of screening for SCA, and this method explains differences in SCA’s due to age and sex. Model performance improved when ECGs were obtained at shorter intervals to the incident SCA event, and SCA prediction was more dependent upon QRS complex data than other temporal segments of the ECG.

An Improved Transformer Model for Sap Flow Prediction that Efficiently Utilizes Environmental Information

AbstractAs an important reference for assessing plant water consumption and estimating plant transpiration, it is of great significance to achieve accurate prediction of plant sap flow. A number of deep learning models were established and compared using approximately 3 years of continuous eucalyptus flow time series data collected from the SAPFLUXNET open dataset and 6 environmental factors, including shortwave solar incident radiation, air temperature, air relative humidity, net radiation, vapor pressure deficit, and photosynthetic photon flux density. The experimental results show that the improved Transformer model, with the introduction of a two-step self-attention mechanism and simplified design, maintains significant predictive performance advantages compared to the original Transformer model, long short-term memory, gated recurrent unit, and temporal convolutional neural network models. In the shorter 1-h forecast, the mean squared error and coefficient of determination (R2) of the improved Transformer model are 0.0191 and 0.965, respectively. Compared to the suboptimal typical Transformer model, the MSE is reduced by 22.9%, and R2 is increased by 1.0%. Additionally, the improved model maintains stable predictive performance advantages in long-term plant flow prediction. In the longest 8-h advance prediction, the MSE is reduced by 14.9% compared to the suboptimal Transformer model, and R2 increases by 3.0% compared to the Transformer model. The comprehensive experimental results show that the improved Transformer model makes more effective use of environmental information to achieve more accurate and long-term plant flow prediction. This study emphasizes the basic principle and validity of the two-step self-attention network structure and provides a valuable basis for developing more effective methods for predicting plant sap flow.

Abstract 4114276: Identification of the heterogeneity in the risk of heart failure in those with left bundle branch block pattern using artificial intelligence

Introduction: The association between left bundle branch block (LBBB) and heart failure is well-established. However, the lack of a practical risk assessment tool for patients with LBBB poses a clinical challenge to the management despite the heterogeneity of the LBBB population. Hypothesis: Artificial intelligence (AI) can extract features to stratify the risk of heart failure (HF) within the heterogeneity of LBBB. Methods: All 12-lead ECGs diagnosed with LBBB recorded from 1/1/2015 to 12/31/2019 were identified in our institution. ECG data were analyzed as 2D matrices with a shape of 12×2500. The Uniform Manifold Approximation and Projection (UMAP) was used to visualize the heterogeneity of LBBB ECG. Additionally, a 2-dimensional convolutional neural network model was trained to detect LBBB ECGs associated with past HF diagnosis. The model was then applied to ECGs from an external cohort of patients with LBBB but without a history of HF. Cumulative incidences of HF admission were compared by stratifying according to tertiles of the model output (low, intermediate and high AI score groups) Results: We identified 15,124 LBBB ECGs in our institution (training dataset) and 2,463 individuals with LBBB ECGs in the external cohort (external test dataset). The UMAP projection of the ECGs revealed distinct heterogeneity in the data ( Fig A ). This heterogeneity was not fully explained by patient demographics, ECG features, or institutions. When the external cohort was divided into 3 groups according to the tertile of AI prediction, patients with high AI scores had a higher risk of HF admission (high vs low AI score group: hazard ratio (HR), 2.10; 95% confidence interval (CI), 1.66-2.65; intermediate vs low AI score group: HR, 1.39; 95%CI, 1.08-1.78: log-rank p <0.0001; Fig 1B ). This finding was unchanged after adjusting for multiple clinical characteristics and ECG parameters, including QRS duration. Conclusion: We observed significant heterogeneity in ECGs presenting LBBB, and the AI model excellently stratified the risk of heart failure admission in patients with LBBB from a single recording of 12-lead ECG, suggesting great potential for clinical utility in managing patients with LBBB.

Grade prediction of lesions in cerebral white matter using a convolutional neural network.

We established a diagnostic method for cerebral white matter lesions using MRI images and examined the relationship between the MRI images and the medical checkup data. There were approximately 25 MRI images for each patient's head, from the top of the head to near the eyes. To order these images, we defined the unit of axial for convenience. We varied conditions, such as the location and extent of the images to be loaded, into a convolutional neural network model and verified the changes in discrimination performance on the test data. Co-occurrence network diagrams were also used to determine the relationship between the grade of cerebral white matter lesions and the biochemical test items, which were treated as categorical variables, the progression of cerebral white matter lesions, and patient health status. The convolutional neural network showed the highest discrimination performance when the images were loaded into the model with 80 pixels per side, axial from 9 to 15, along with FLAIR and T1-weighted images. The area under the curve for each grade was 0.9814 for grade 0, 0.9800 for grade 1, 0.9905 for grade 2, 0.9977 for grade 3, and 0.9998 for grade 4. In the co-occurrence network diagram, patients with no or mild cerebral white matter lesions, such as grade 0 and grade 1, had near normal blood pressure, whereas grade 2 patients were closer to (isolated) systolic hypertension. This indicates that patients with higher-grade cerebral white matter lesions tend to experience more severe hypertension.

Predicting malignancy in breast lesions: enhancing accuracy with fine-tuned convolutional neural network models.

This study aims to explore the accuracy of Convolutional Neural Network (CNN) models in predicting malignancy in Dynamic Contrast-Enhanced Breast Magnetic Resonance Imaging (DCE-BMRI). A total of 273 benign lesions (benign group) and 274 malignant lesions (malignant group) were collected and randomly divided into a training set (246 benign and 245 malignant lesions) and a testing set (28 benign and 28 malignant lesions) in a 9:1 ratio. An additional 53 lesions from 53 patients were designated as the validation set. Five models-VGG16, VGG19, DenseNet201, ResNet50, and MobileNetV2-were evaluated. Model performance was assessed using accuracy (Ac) in the training and testing sets, and precision (Pr), recall (Rc), F1 score (F1), and area under the receiver operating characteristic curve (AUC) in the validation set. The accuracy of VGG19 on the test set (0.96) is higher than that of VGG16 (0.91), DenseNet201 (0.91), ResNet50 (0.67), and MobileNetV2 (0.88). For the validation set, VGG19 achieved higher performance metrics (Pr 0.75, Rc 0.76, F1 0.73, AUC 0.76) compared to the other models, specifically VGG16 (Pr 0.73, Rc 0.75, F1 0.70, AUC 0.73), DenseNet201 (Pr 0.71, Rc 0.74, F1 0.69, AUC 0.71), ResNet50 (Pr 0.65, Rc 0.68, F1 0.60, AUC 0.65), and MobileNetV2 (Pr 0.73, Rc 0.75, F1 0.71, AUC 0.73). S4 model achieved higher performance metrics (Pr 0.89, Rc 0.88, F1 0.87, AUC 0.89) compared to the other four fine-tuned models, specifically S1 (Pr 0.75, Rc 0.76, F1 0.74, AUC 0.75), S2 (Pr 0.77, Rc 0.79, F1 0.75, AUC 0.77), S3 (Pr 0.76, Rc 0.76, F1 0.73, AUC 0.75), and S5 (Pr 0.77, Rc 0.79, F1 0.75, AUC 0.77). Additionally, S4 model showed the lowest loss value in the testing set. Notably, the AUC of S4 for BI-RADS 3 was 0.90 and for BI-RADS 4 was 0.86, both significantly higher than the 0.65 AUC for BI-RADS 5. The S4 model we propose has demonstrated superior performance in predicting the likelihood of malignancy in DCE-BMRI, making it a promising candidate for clinical application in patients with breast diseases. However, further validation is essential, highlighting the need for additional data to confirm its efficacy.