Test Set Research Articles

Machine learning assisted designing of Y-series small molecule acceptors: Library generation and property prediction

The designing of new small molecule acceptors (SMAs) for organic solar cells has been a prominent area of research for many decades. It is challenging to find unique materials due to expensive experimentation. Machine learning emerges as a promising tool for rapid and cost-effective prediction of properties. In the current study, electron affinity is predicted using machine learning models. Molecular descriptors are calculated for model training. Four machine learning models are used. A reasonably high predictive capability is obtained for random forest model (r-squared values of 0.92 and 0.82 for training and test set, respectively). Furthermore, the chemical database of small molecule acceptors is generated. Generated virtual space is anticipated by exploiting the t-distributed stochastic neighbor embedding (t-SNE) method. Structure Activity Landscape Index (SALI) analysis is conducted to examine how properties change with structural variations. A minimal change in electron affinity resulted from structural modifications. Additionally, clustering analysis is performed on selected small molecule acceptors for structural grouping of SMAs. Five SMAs with lowest synthetic accessibility (SA) score are selected for density functional theory (DFT) calculations. The introduced multiple dimensional framework has potential screened the materials in short-time and efficient way.

A two-stage approach for aircraft detection with convolutional neural network

Over the past few years, object detection has experienced remarkable advancements, primarily attributable to significant progress in deep learning architectures. Nonetheless, the task of identifying aircraft targets within remote sensing images remains a challenging and actively explored area. Presently, there are two main approaches employed for this task: one utilizing convolutional neural network (CNN) techniques and the other relying on conventional methods. In this work, a CNN based architecture is proposed to recognize aircraft types using remote sensing images. The experiments performed on multi-type aircraft remote sensing images (MTARSI) dataset show that the proposed architecture achieves 97.07%, 94.81%, and 94.44% accuracy rates for training, validation and testing sets. The results approve that, the architecture outperforms state of the art models.

Development of Chromatin Regulator-related Molecular Subtypes and a Signature to Predict Prognosis and Immunotherapeutic Response in Head and Neck Squamous Cell Carcinoma.

Chromatin regulators (CRs) serve as indispensable factors in tumor biological processes by influencing tumorigenesis and the immune microenvironment and have been identified in head and neck squamous cell carcinoma (HNSCC). Hence, CR-related genes (CRRGs) are considered potential biomarkers for predicting prognosis and immune infiltration in HNSCC. In this study, we established a novel signature for predicting the prognosis and immunotherapeutic response of HSNCC. A total of 870 CRRGs were obtained according to previous studies. Subsequently, patients in the TCGA-HNSC cohort were divided into different clusters based on the expression of prognostic CRRGs. Kaplan‒Meier (K‒M) survival analysis was conducted to compare the prognosis in clusters, and the CIBERSORT and ssGSEA methods assessed the immune infiltration status. In addition, the differences in immunotherapeutic responses were determined based on the TICA database. Furthermore, the differentially expressed CRRGs between clusters were identified, and the predictive signature was established according to the results of univariate Cox, least absolute shrinkage and selection operator regression analysis, and multivariate Cox. The predictive effects of the risk model were evaluated according to the area under the receiver operating characteristic (ROC) curve (AUC) in both the training and external test cohorts. A nomogram was established, and survival comparisons, functional enrichment analyses, and immune infiltration status and clinical treatment assessments were performed. In addition, the hub gene network and related analysis were conducted with the Cytohubba application. Based on the expression of prognostic CRRGs, patients were divided into two clusters, in which Cluster 1 exhibited a better prognosis, more enriched immune infiltration, and a better immunotherapeutic response but exhibited chemotherapy sensitivity. The AUC values of the 1-, 3- and 5- year ROC curves for the risk model were 0.673, 0.732, and 0.692, respectively, as well as 0.645, 0.608, and 0.623 for the test set. In addition, patients in the low-risk group exhibited more immune cell enrichment and immune function activation, as well as a better immunotherapy response. The hub gene network indicated ACTN2 as the core gene differentially expressed between the two risk groups. We identified molecular subtypes and established a novel predictive signature based on CRRGs. This effective CRRS system can possibly provide a novel research direction for exploring the correlation between CRs and HNSCC and requires further experimental validation.

Enhancing Breast Cancer Detection: Leveraging Convolutional Neural Networks

Abstract: Breast cancer continues to be a critical health concern, necessitating early detection and accurate classification for effective treatment. This study presents a comparative analysis between a custom-designed Convolutional Neural Network (CNN) and the pre-trained DenseNet121 model for breast cancer detection and classification. We compiled a comprehensive dataset of breast cancer images and applied appropriate preprocessing techniques to optimize the input for the models. The dataset was divided into training, validation, and testing sets to evaluate the models' performance. The CNN model comprises multiple convolutional and pooling layers followed by fully connected layers for classification, while the DenseNet121 model is fine-tuned specifically for breast cancer detection. The models were evaluated based on metrics such as accuracy, precision, recall, F1 score, and AUC-ROC. The DenseNet121 model outperformed the custom CNN model, achieving higher accuracy and reliability. For wider accessibility, we integrated the superior DenseNet121 model into a user-friendly web-based interface using Python Flask, enabling real-time breast cancer predictions. Ethical considerations were paramount, ensuring data privacy, security, and transparency in all model predictions. This study highlights the effectiveness of the DenseNet121 model and contributes to improved breast cancer diagnosis and patient care.

Influencing factors and risk prediction model for emergence agitation after general anesthesia for primary liver cancer

BACKGROUND General anesthesia is commonly used in the surgical management of gastrointestinal tumors; however, it can lead to emergence agitation (EA). EA is a common complication associated with general anesthesia, often characterized by behaviors, such as crying, struggling, and involuntary limb movements in patients. If treatment is delayed, there is a risk of incision cracking and bleeding, which can significantly affect surgical outcomes. Therefore, having a proper understanding of the factors influencing the occurrence of EA and implementing early preventive measures may reduce the incidence of agitation during the recovery phase from general anesthesia, which is beneficial for improving patient prognosis. AIM To analyze influencing factors and develop a risk prediction model for EA occurrence following general anesthesia for primary liver cancer. METHODS Retrospective analysis of clinical data from 200 patients who underwent hepatoma resection under general anesthesia at Wenzhou Central Hospital (January 2020 to December 2023) was conducted. Post-surgery, the Richmond Agitation-Sedation Scale was used to evaluate EA presence, noting EA incidence after general anesthesia. Patients were categorized by EA presence postoperatively, and the influencing factors were analyzed using logistic regression. A nomogram-based risk prediction model was constructed and evaluated for differentiation and fit using receiver operating characteristics and calibration curves. RESULTS EA occurred in 51 (25.5%) patients. Multivariate analysis identified advanced age, American Society of Anesthesiologists (ASA) grade III, indwelling catheter use, and postoperative pain as risk factors for EA (P < 0.05). Conversely, postoperative analgesia was a protective factor against EA (P < 0.05). The area under the curve of the nomogram was 0.972 [95% confidence interval (CI): 0.947-0.997] for the training set and 0.979 (95%CI: 0.951-1.000) for the test set. Hosmer-Lemeshow test showed a good fit (χ 2 = 5.483, P = 0.705), and calibration curves showed agreement between predicted and actual EA incidence. CONCLUSION Age, ASA grade, catheter use, postoperative pain, and analgesia significantly influence EA occurrence. A nomogram constructed using these factors demonstrates strong predictive accuracy.

Interpretable Machine Learning-Based Risk Score for Predicting Ten-Year Corneal Graft Survival After Penetrating Keratoplasty and Deep Anterior Lamellar Keratoplasty in Asian Eyes.

To predict 10-year graft survival after deep anterior lamellar keratoplasty (DALK) and penetrating keratoplasty (PK) using a machine learning (ML)-based interpretable risk score. Singapore Corneal Transplant Registry patients (n = 1687) who underwent DALK (n = 524) or PK (n = 1163) for optical indications (excluding endothelial diseases) were followed up for 10 years. Variable importance scores from random survival forests were used to identify variables associated with graft survival. Parsimonious analysis using nested Cox models selected the top factors. An ML-based clinical score generator (AutoScore) converted identified variables into an interpretable risk score. Predictive performance was evaluated using Kaplan-Meier (KM) curves and time-integrated AUC (iAUC) on an independent testing set. Mean recipient age was 51.8 years, 54.1% were male, and majority were Chinese (60.0%). Surgical indications included corneal scar (46.5%), keratoconus (18.3%), and regraft (16.2%). Five-year and ten-year KM survival was 93.4% and 92.3% for DALK, compared with 67.6% and 56.6% for PK (log-rank P < 0.001). Five factors were identified by ML algorithm as predictors of 10-year graft survival: recipient sex, preoperative visual acuity, choice of procedure, surgical indication, and active inflammation. AutoScore stratified participants into low-risk and high-risk groups-with KM survival of 73.6% and 39.0%, respectively (log-rank P < 0.001). ML analysis outperformed traditional Cox regression in predicting graft survival beyond 5 years (iAUC 0.75 vs. 0.69). A combination of ML and traditional techniques identified factors associated with graft failure to derive a clinically interpretable risk score to stratify PK and DALK patients-a technique that may be replicated in other corneal transplant programs.

Deep learning-based fully automatic Risser stage assessment model using abdominal radiographs.

Artificial intelligence has been increasingly used in medical imaging and has demonstrated expert level performance in image classification tasks. To develop a fully automatic approach for determining the Risser stage using deep learning on abdominal radiographs. In this multicenter study, 1,681 supine abdominal radiographs (age range, 9-18years, 50% female) obtained between January 2019 and April 2022 were collected retrospectively from three medical institutions and graded manually using the United States Risser staging system. A total of 1,577 images from Hospitals 1 and 2 were used for development, and 104 images from Hospital 3 for external validation. From each radiograph, right and left iliac crest patch images were extracted using the pelvic bone segmentation model DeepLabv3 + with the EfficientNet-B0 encoder trained with 90 digitally reconstructed radiographs from pelvic computed tomography scans with a pelvic bone mask. Using these patch images, ConvNeXt-B was trained to grade according to the Risser classification. The model's performance was evaluated using accuracy, area under the receiver operating characteristic curve (AUROC), and mean absolute error. The fully automatic Risser stage assessment model showed an accuracy of 0.87 and 0.75, mean absolute error of 0.13 and 0.26, and AUROC of 0.99 and 0.95 on internal and external test sets, respectively. We developed a deep learning-based, fully automatic segmentation and classification model for Risser stage assessment using abdominal radiographs.

Building Dual AI Models and Nomograms Using Noninvasive Parameters for Aiding Male Bladder Outlet Obstruction Diagnosis and Minimizing the Need for Invasive Video-Urodynamic Studies: Development and Validation Study.

Diagnosing underlying causes of nonneurogenic male lower urinary tract symptoms associated with bladder outlet obstruction (BOO) is challenging. Video-urodynamic studies (VUDS) and pressure-flow studies (PFS) are both invasive diagnostic methods for BOO. VUDS can more precisely differentiate etiologies of male BOO, such as benign prostatic obstruction, primary bladder neck obstruction, and dysfunctional voiding, potentially outperforming PFS. These examinations' invasive nature highlights the need for developing noninvasive predictive models to facilitate BOO diagnosis and reduce the necessity for invasive procedures. We conducted a retrospective study with a cohort of men with medication-refractory, nonneurogenic lower urinary tract symptoms suspected of BOO who underwent VUDS from 2001 to 2022. In total, 2 BOO predictive models were developed-1 based on the International Continence Society's definition (International Continence Society-defined bladder outlet obstruction; ICS-BOO) and the other on video-urodynamic studies-diagnosed bladder outlet obstruction (VBOO). The patient cohort was randomly split into training and test sets for analysis. A total of 6 machine learning algorithms, including logistic regression, were used for model development. During model development, we first performed development validation using repeated 5-fold cross-validation on the training set and then test validation to assess the model's performance on an independent test set. Both models were implemented as paper-based nomograms and integrated into a web-based artificial intelligence prediction tool to aid clinical decision-making. Among 307 patients, 26.7% (n=82) met the ICS-BOO criteria, while 82.1% (n=252) were diagnosed with VBOO. The ICS-BOO prediction model had a mean area under the receiver operating characteristic curve (AUC) of 0.74 (SD 0.09) and mean accuracy of 0.76 (SD 0.04) in development validation and AUC and accuracy of 0.86 and 0.77, respectively, in test validation. The VBOO prediction model yielded a mean AUC of 0.71 (SD 0.06) and mean accuracy of 0.77 (SD 0.06) internally, with AUC and accuracy of 0.72 and 0.76, respectively, externally. When both models' predictions are applied to the same patient, their combined insights can significantly enhance clinical decision-making and simplify the diagnostic pathway. By the dual-model prediction approach, if both models positively predict BOO, suggesting all cases actually resulted from medication-refractory primary bladder neck obstruction or benign prostatic obstruction, surgical intervention may be considered. Thus, VUDS might be unnecessary for 100 (32.6%) patients. Conversely, when ICS-BOO predictions are negative but VBOO predictions are positive, indicating varied etiology, VUDS rather than PFS is advised for precise diagnosis and guiding subsequent therapy, accurately identifying 51.1% (47/92) of patients for VUDS. The 2 machine learning models predicting ICS-BOO and VBOO, based on 6 noninvasive clinical parameters, demonstrate commendable discrimination performance. Using the dual-model prediction approach, when both models predict positively, VUDS may be avoided, assisting in male BOO diagnosis and reducing the need for such invasive procedures.

Large-Scale Proteomics Improve Prediction of Chronic Kidney Disease in People With Diabetes.

To develop and validate a protein risk score for predicting chronic kidney disease (CKD) in patients with diabetes and compare its predictive performance with a validated clinical risk model (CKD Prediction Consortium [CKD-PC]) and CKD polygenic risk score. This cohort study included 2,094 patients with diabetes who had proteomics and genetic information and no history of CKD at baseline from the UK Biobank Pharma Proteomics Project. Based on nearly 3,000 plasma proteins, a CKD protein risk score including 11 proteins was constructed in the training set (including 1,047 participants; 117 CKD events). The median follow-up duration was 12.1 years. In the test set (including 1,047 participants; 112 CKD events), the CKD protein risk score was positively associated with incident CKD (per SD increment; hazard ratio 1.78; 95% CI 1.44, 2.20). Compared with the basic model (age + sex + race, C-index, 0.627; 95% CI 0.578, 0.675), the CKD protein risk score (C-index increase 0.122; 95% CI 0.071, 0.177), and the CKD-PC risk factors (C-index increase 0.175; 95% CI 0.126, 0.217) significantly improved the prediction performance of incident CKD, but the CKD polygenic risk score (C-index increase 0.007; 95% CI -0.016, 0.025) had no significant improvement. Adding the CKD protein risk score into the CKD-PC risk factors had the largest C-index of 0.825 (C-index from 0.802 to 0.825; difference 0.023; 95% CI 0.006, 0.044), and significantly improved the continuous 10-year net reclassification (0.199; 95% CI 0.059, 0.299) and 10-year integrated discrimination index (0.041; 95% CI 0.007, 0.083). Adding the CKD protein risk score to a validated clinical risk model significantly improved the discrimination and reclassification of CKD risk in patients with diabetes.

Machine learning-based estimation of patient body weight from radiation dose metrics in computed tomography.

Currently, precise patient body weight (BW) at the time of diagnostic imaging cannot always be used for radiation dose management. Various methods have been explored to address this issue, including the application of deep learning to medical imaging and BW estimation using scan parameters. This study develops and evaluates machine learning-based BW prediction models using 11 features related to radiation dose obtained from computed tomography (CT) scans. A dataset was obtained from 3996 patients who underwent positron emission tomography CT scans, and training and test sets were established. Dose metrics and descriptive data were automatically calculated from the CT images or obtained from Digital Imaging and Communications in Medicine metadata. Seven machine-learning models and three simple regression models were employed to predict BW using features such as effective diameter (ED), water equivalent diameter (WED), and mean milliampere-seconds. The mean absolute error (MAE) and correlation coefficient between the estimated BW and the actual BW obtained from each BW prediction model were calculated. Our results found that the highest accuracy was obtained using a light gradient-boosting machine model, which had an MAE of 1.99kg and a strong positive correlation between estimated and actual BW (ρ=0.972). The model demonstrated significant predictive power, with 73% of patients falling within a±5% error range. WED emerged as the most relevant dose metric for BW estimation, followed by ED and sex. The proposed machine-learning approach is superior to existing methods, with high accuracy and applicability to radiation dose management. The model's reliance on universal dose metrics that are accessible through radiation dose management software enhances its practicality. In conclusion, this study presents a robust approach for BW estimation based on CT imaging that can potentially improve radiation dose management practices in clinical settings.

Groundwater salinity modeling and mapping using machine learning approaches: a case study in Sidi Okba region, Algeria.

The groundwater salinization process complexity and the lack of data on its controlling factors are the main challenges for accurate predictions and mapping of aquifer salinity. For this purpose, effective machine learning (ML) methodologies are employed for effective modeling and mapping of groundwater salinity (GWS) in the Mio-Pliocene aquifer in the Sidi Okba region, Algeria, based on limited dataset of electrical conductivity (EC) measurements and readily available digital elevation model (DEM) derivatives. The dataset was randomly split into training (70%) and testing (30%) sets, and three wrapper selection methods, recursive feature elimination (RFE), forward feature selection (FFS), and backward feature selection (BFS) are applied to train the data. The resulting combinations are used as inputs for five ML models, namely random forest (RF), hybrid neuro-fuzzy inference system (HyFIS), K-nearest neighbors (KNN), cubist regression model (CRM), and support vector machine (SVM). The best-performing model is identified and applied to predict and map GWS across the entire study area. It is highlighted that the applied methods yield input variation combinations as critical factors that are often overlocked by many researchers, which substantially impacts the models' accuracy. Among different alternatives the RF model emerged as the most effective for predicting and mapping GWS in the study area, which led to the high performance in both the training (RMSE = 1.016, R = 0.854, and MAE = 0.759) and testing (RMSE = 1.069, R = 0.831, and MAE = 0.921) phases. The generated digital map highlighted the alarming situation regarding excessive GWS levels in the study area, particularly in zones of low elevations and far from the Foum Elgherza dam and Elbiraz wadi. Overall, this study represents a significant advancement over previous approaches, offering enhanced predictive performance for GWS with the minimum number of input variables.

Multimodal deep learning using on-chip diffractive optics with in situ training capability

Multimodal deep learning plays a pivotal role in supporting the processing and learning of diverse data types within the realm of artificial intelligence generated content (AIGC). However, most photonic neuromorphic processors for deep learning can only handle a single data modality (either vision or audio) due to the lack of abundant parameter training in optical domain. Here, we propose and demonstrate a trainable diffractive optical neural network (TDONN) chip based on on-chip diffractive optics with massive tunable elements to address these constraints. The TDONN chip includes one input layer, five hidden layers, and one output layer, and only one forward propagation is required to obtain the inference results without frequent optical-electrical conversion. The customized stochastic gradient descent algorithm and the drop-out mechanism are developed for photonic neurons to realize in situ training and fast convergence in the optical domain. The TDONN chip achieves a potential throughput of 217.6 tera-operations per second (TOPS) with high computing density (447.7 TOPS/mm2), high system-level energy efficiency (7.28 TOPS/W), and low optical latency (30.2 ps). The TDONN chip has successfully implemented four-class classification in different modalities (vision, audio, and touch) and achieve 85.7% accuracy on multimodal test sets. Our work opens up a new avenue for multimodal deep learning with integrated photonic processors, providing a potential solution for low-power AI large models using photonic technology.

Multi-kernel support vector regression with improved moth-flame optimization algorithm for software effort estimation

In this paper, a novel Moth-Flame Optimization (MFO) algorithm, namely MFO algorithm enhanced by Multiple Improvement Strategies (MISMFO) is proposed for solving parameter optimization in Multi-Kernel Support Vector Regressor (MKSVR), and the MISMFO-MKSVR model is further employed to deal with the software effort estimation problems. In MISMFO, the logistic chaotic mapping is applied to increase initial population diversity, while the mutation and flame number phased reduction mechanisms are carried out to improve the search efficiency, as well the adaptive weight adjustment mechanism is used to accelerate convergence and balance exploration and exploitation. The MISMFO model is verified on fifteen benchmark functions and CEC 2020 test set. The results show that the MISMFO has advantages over other meta-heuristic algorithms and MFO variants in terms of convergence speed and accuracy. Additionally, the MISMFO-MKSVR model is tested by simulations on five software effort datasets and the results demonstrate that the proposed model has better performance in software effort estimation problem. The Matlab code of MISMFO can be found at https://github.com/loadstar1997/MISMFO.

Second opinion machine learning for fast-track pathway assignment in hip and knee replacement surgery: the use of patient-reported outcome measures

BackgroundThe frequency of hip and knee arthroplasty surgeries has been rising steadily in recent decades. This trend is attributed to an aging population, leading to increased demands on healthcare systems. Fast Track (FT) surgical protocols, perioperative procedures designed to expedite patient recovery and early mobilization, have demonstrated efficacy in reducing hospital stays, convalescence periods, and associated costs. However, the criteria for selecting patients for FT procedures have not fully capitalized on the available patient data, including patient-reported outcome measures (PROMs).MethodsOur study focused on developing machine learning (ML) models to support decision making in assigning patients to FT procedures, utilizing data from patients’ self-reported health status. These models are specifically designed to predict the potential health status improvement in patients initially selected for FT. Our approach focused on techniques inspired by the concept of controllable AI. This includes eXplainable AI (XAI), which aims to make the model’s recommendations comprehensible to clinicians, and cautious prediction, a method used to alert clinicians about potential control losses, thereby enhancing the models’ trustworthiness and reliability.ResultsOur models were trained and tested using a dataset comprising 899 records from individual patients admitted to the FT program at IRCCS Ospedale Galeazzi-Sant’Ambrogio. After training and selecting hyper-parameters, the models were assessed using a separate internal test set. The interpretable models demonstrated performance on par or even better than the most effective ‘black-box’ model (Random Forest). These models achieved sensitivity, specificity, and positive predictive value (PPV) exceeding 70%, with an area under the curve (AUC) greater than 80%. The cautious prediction models exhibited enhanced performance while maintaining satisfactory coverage (over 50%). Further, when externally validated on a separate cohort from the same hospital-comprising patients from a subsequent time period-the models showed no pragmatically notable decline in performance.ConclusionsOur results demonstrate the effectiveness of utilizing PROMs as basis to develop ML models for planning assignments to FT procedures. Notably, the application of controllable AI techniques, particularly those based on XAI and cautious prediction, emerges as a promising approach. These techniques provide reliable and interpretable support, essential for informed decision-making in clinical processes.

A comparative analysis of boosting and genetic programming techniques for predicting mechanical properties of soilcrete materials

This study aims to predict compressive strength (CS) and modulus of elasticity (E) of soilcrete mixes to foster their widespread use in the industry. Soilcfigrete has the potential to promote sustainable construction practices by making use of locally available raw materials. However, the accurate determination of mechanical properties of soilcrete mixes is inevitable to foster their widespread use. Thus, this study employs different machine learning algorithms including Extreme Gradient Boosting (XGB), Gene Expression Programming (GEP), AdaBoost, and Multi Expression Programming (MEP) for this purpose. The XGB and AdaBoost algorithms were implemented using python programming language while MEP and GEP were implemented using specialized softwares. The data used for model development was obtained from previously published literature containing five input parameters and two output parameters. This data was split into two sets named training and testing sets for training and testing of the algorithms respectively. The developed models for CS and E prediction were validated using several error metrices and residual comparison. The objective function value which should be closer to zero for an accurate model is the least for XGB model for prediction of both variables (0.0036 for CS and 0.00315 for E). Moreover, shapley analysis was carried out using XGB model to get insights into the underlying model framework. The results highlighted that water-to-binder ratio (W/B), metakaolin (MK), and ultrasonic pulse velocity (UV) are the most significant variables for predicting E and CS of soilcrete materials. These insights can be used practically to optimize the mixture composition of soilcrete mixes according to different site requirements.

Deep learning classification of pediatric spinal radiographs for use in large scale imaging registries.

The purpose of this study is to develop and apply an algorithm that automatically classifies spine radiographs of pediatric scoliosis patients. Anterior-posterior (AP) and lateral spine radiographs were extracted from the institutional picture archive for patients with scoliosis. Overall, there were 7777 AP images and 5621 lateral images. Radiographs were manually classified into ten categories: two preoperative and three postoperative categories each for AP and lateral images. The images were split into training, validation, and testing sets (70:15:15 proportional split). A deep learning classifier using the EfficientNet B6 architecture was trained on the spine training set. Hyperparameters and model architecture were tuned against the performance of the models in the validation set. The trained classifiers had an overall accuracy on the test set of 1.00 on 1166 AP images and 1.00 on 843 lateral images. Precision ranged from 0.98 to 1.00 in the AP images, and from 0.91 to 1.00 on the lateral images. Lower performance was observed on classes with fewer than 100 images in the dataset. Final performance metrics were calculated on the assigned test set, including accuracy, precision, recall, and F1 score (the harmonic mean of precision and recall). A deep learning convolutional neural network classifier was trained to a high degree of accuracy to distinguish between 10 categories pre- and postoperative spine radiographs of patients with scoliosis. Observed performance was higher in more prevalent categories. These models represent an important step in developing an automatic system for data ingestion into large, labeled imaging registries.

MPOCSR: optical chemical structure recognition based on multi-path Vision Transformer

AbstractOptical chemical structure recognition (OCSR) is a fundamental and crucial task in the field of chemistry, which aims at transforming intricate chemical structure images into machine-readable formats. Current deep learning-based OCSR methods typically use image feature extractors to extract visual features and employ encoder-decoder architectures for chemical structure recognition. However, the performance of these methods is limited by their image feature extractors and the class imbalance of elements in chemical structure representation. This paper proposes MPOCSR (multi-path optical chemical structure recognition), which introduces the multi-path Vision Transformer (MPViT) and the class-balanced (CB) loss function to address these two challenges. MPOCSR uses MPViT as an image feature extractor, combining the advantages of convolutional neural networks and Vision Transformers. This strategy enables the provision of richer visual information for subsequent decoding processes. Furthermore, MPOCSR incorporates CB loss function to rebalance the loss weights among different categories. For training and validation of our method, we constructed a dataset that includes both Markush and non-Markush structures. Experimental results show that MPOCSR achieves an accuracy of 90.95% on the test set, surpassing other existing methods.

Towards Exploring the Limitations of Test Selection Techniques on Graph Neural Networks: An Empirical Study

Graph Neural Networks (GNNs) have gained prominence in various domains, such as social network analysis, recommendation systems, and drug discovery, due to their ability to model complex relationships in graph-structured data. GNNs can exhibit incorrect behavior, resulting in severe consequences. Therefore, testing is necessary and pivotal. However, labeling all test inputs for GNNs can be prohibitively costly and time-consuming, especially when dealing with large and complex graphs. In response to these challenges, test selection has emerged as a strategic approach to alleviate labeling expenses. The objective of test selection is to select a subset of tests from the complete test set. While various test selection techniques have been proposed for traditional deep neural networks (DNNs), their adaptation to GNNs presents unique challenges due to the distinctions between DNN and GNN test data. Specifically, DNN test inputs are independent of each other, whereas GNN test inputs (nodes) exhibit intricate interdependencies. Therefore, it remains unclear whether DNN test selection approaches can perform effectively on GNNs. To fill the gap, we conduct an empirical study that systematically evaluates the effectiveness of various test selection methods in the context of GNNs, focusing on three critical aspects: 1) Misclassification detection: selecting test inputs that are more likely to be misclassified; 2) Accuracy estimation: selecting a small set of tests to precisely estimate the accuracy of the whole testing set; 3) Performance enhancement: selecting retraining inputs to improve the GNN accuracy. Our empirical study encompasses 7 graph datasets and 8 GNN models, evaluating 22 test selection approaches. Our study includes not only node classification datasets but also graph classification datasets. Our findings reveal that: 1) In GNN misclassification detection, confidence-based test selection methods, which perform well in DNNs, do not demonstrate the same level of effectiveness; 2) In terms of GNN accuracy estimation, clustering-based methods, while consistently performing better than random selection, provide only slight improvements; 3) Regarding selecting inputs for GNN performance improvement, test selection methods, such as confidence-based and clustering-based test selection methods, demonstrate only slight effectiveness; 4) Concerning performance enhancement, node importance-based test selection methods are not suitable, and in many cases, they even perform worse than random selection.

Porosity Prediction Based on Ensemble Learning for Feature Selection and an Optimized GRU Improved by the PSO Algorithm

Accurate and reliable prediction of porosity forms the foundational basis for evaluating reservoir quality, which is essential for the systematic deployment of oil and gas exploration and development plans. When data quality of samples is low, and critical model parameters are typically determined through subjective experience, resulting in diminished accuracy and reliability of porosity prediction methods utilizing gated recurrent units (GRU), a committee-voting ensemble learning (EL) method, and an enhanced particle swarm optimization (PSO) algorithm are proposed to optimize the GRU-based porosity prediction model. Initially, outliers are eliminated through box plots and the min–max normalization is applied to enhance data quality. To address issues related to model accuracy and high training costs arising from dimensional complexity, substantial noise, and redundant information in logging data, a committee-voting EL strategy based on four feature selection algorithms is introduced. Following data preprocessing, this approach is employed to identify logging parameters highly correlated with porosity, thereby furnishing the most pertinent data samples for the GRU model, mitigating constraints imposed by single-feature selection methods. Second, an improved PSO algorithm is suggested to tackle challenges associated with low convergence accuracy stemming from random population initialization, alongside the absence of global optimal solutions due to overly rapid particle movement during iteration. This algorithm uses a good-point set for population initialization and incorporates a compression factor to devise an adaptive velocity updating strategy, thereby enhancing search efficacy. The enhanced PSO algorithm’s superiority is substantiated through comparison with four alternative swarm intelligent algorithms across 10 benchmark test functions. Ultimately, optimal hyper-parameters for the GRU model are determined using the improved PSO algorithm, thereby minimizing the influence of human factors. Experimental findings based on approximately 15,000 logging data points from well A01 in an operational field validate that, relative to three other deep learning methodologies, the proposed model proficiently extracts spatiotemporal features from logging data, yielding enhanced accuracy in porosity prediction. The mean squared error on the test set was 7.19 × 10–6, the mean absolute error stood at 0.0082, and coefficient of determination reached 0.99, offering novel insights for predicting reservoir porosity.

Classification of subtask types and skill levels in robot-assisted surgery using EEG, eye-tracking, and machine learning.

Objective and standardized evaluation of surgical skills in robot-assisted surgery (RAS) holds critical importance for both surgical education and patient safety. This study introduces machine learning (ML) techniques using features derived from electroencephalogram (EEG) and eye-tracking data to identify surgical subtasks and classify skill levels. The efficacy of this approach was assessed using a comprehensive dataset encompassing nine distinct classes, each representing a unique combination of three surgical subtasks executed by surgeons while performing operations on pigs. Four ML models, logistic regression, random forest, gradient boosting, and extreme gradient boosting (XGB) were used for multi-class classification. To develop the models, 20% of data samples were randomly allocated to a test set, with the remaining 80% used for training and validation. Hyperparameters were optimized through grid search, using fivefold stratified cross-validation repeated five times. Model reliability was ensured by performing train-test split over 30 iterations, with average measurements reported. The findings revealed that the proposed approach outperformed existing methods for classifying RAS subtasks and skills; the XGB and random forest models yielded high accuracy rates (88.49% and 88.56%, respectively) that were not significantly different (two-sample t-test; P-value = 0.9). These results underscore the potential of ML models to augment the objectivity and precision of RAS subtask and skill evaluation. Future research should consider exploring ways to optimize these models, particularly focusing on the classes identified as challenging in this study. Ultimately, this study marks a significant step towards a more refined, objective, and standardized approach to RAS training and competency assessment.