Performance Of Model Research Articles

Enhanced diabetic retinopathy detection and classification using fundus images with ResNet50 and CLAHE-GAN

Diabetic retinopathy (DR), a progressive eye disorder, can lead to irreversible vision impairment ranging from no DR to severe DR, necessitating precise identification for early treatment. This study introduces an innovative deep learning (DL) approach, surpassing traditional methods in detecting DR stages. It evaluated two scenarios for training DL models on balanced datasets. The first employed image enhancement via contrast limited adaptive histogram equalization (CLAHE) and a generative adversarial network (GAN), while the second did not involve any image enhancement. Tested on the Asia pacific tele-ophthalmology society 2019 blindness detection (APTOS-2019 BD) dataset, the enhanced model (scenario 1) reached 98% accuracy and a 99% Cohen kappa score (CKS), with the non-enhanced model (scenario 2) achieving 95.4% accuracy and a 90.5% CKS. The combination of CLAHE and GAN, termed CLANG, significantly boosted the model's performance and generalizability. This advancement is pivotal for early DR detection and intervention, offering a new pathway to prevent irreversible vision loss in diabetic patients.

A data-driven analysis to determine the optimal number of topics 'K' for latent Dirichlet allocation model

Topic modeling is an unsupervised machine learning technique successfully used to classify and retrieve textual data. However, the performance of topic models is sensitive to selecting optimal hyperparameters, the number of topics 'K' and Dirichlet priors 'α' and 'β.' This data-driven analysis aims to determine the optimum number of topics, 'K,' within the latent Dirichlet allocation (LDA) model. This work utilizes three datasets, namely 20-Newsgroups news articles, Wikipedia articles, and Web of Science containing science articles, to assess and compare various 'K' values through the grid search approach. The grid search approach finds the best combination of hyperparameter values by trying all possible combinations to see which performs best. This research seeks to identify the 'K' that optimizes topic relevance, coherence, and model performance by leveraging statistical metrics, such as coherence scores, perplexity, and topic distribution quality. Through empirical analysis and rigorous evaluation, this work provides valuable insights for determining the ideal 'K' for LDA models.

Multi-layer perceptron hyperparameter optimization using Jaya algorithm for disease classification

This study introduces an innovative hyperparameter optimization approach for enhancing multilayer perceptrons (MLP) using the Jaya algorithm. Addressing the crucial role of hyperparameter tuning in MLP’s performance, the Jaya algorithm, inspired by social behavior, emerges as a promising optimization technique without algorithm-specific parameters. Systematic application of Jaya dynamically adjusts hyperparameter values, leading to notable improvements in convergence speeds and model generalization. Quantitatively, the Jaya algorithm consistently achieves convergences at first iteration, faster convergence compared to conventional methods, resulting in 7% higher accuracy levels on several datasets. This research contributes to hyperparameter optimization, offering a practical and effective solution for optimizing MLP in diverse applications, with implications for improved computational efficiency and model performance.

A Blockchain-Based Model for Traceability and Quality Management of Raw Materials in Clothing Products

In this study, we employ blockchain-based traceability technology to construct a comprehensive model for raw material traceability and quality management in clothing products. Additionally, we investigate engineered water nanostructures (EWNS) for their antibacterial, hydrophilic, and anti-pilling effects on textiles. The EWNS can serve as an effective tool for monitoring quality, enabling precise tracking and verification of nanomaterials during clothing production. The research validated the traceability performance of the blockchain-based model for raw material traceability and quality management in clothing products, achieving a remarkable traceability accuracy of 99.92% and an exceptionally low traceability response time of merely 1.83 seconds. Moreover, after being treated with EWNS, antibacterial rates against E. coli and S. aureus in clothing antibacterial quality management reached 99.9% and 99.0%, respectively. In the hydrophilicity test, EWNS was applied to clothing for a duration of 1 to 7 hours, resulting in a significant reduction in the wetting time of clothing from 10.8 seconds to 1.2 seconds. In the anti-pilling tests, there was a notable enhancement in the effectiveness of anti-pilling properties after EWNS treatment. In summary, our model has achieved remarkable outcomes in material traceability and quality management, offering consumers and the clothing industry a reliable method to ensure product quality and traceability, while simultaneously improving performance aspects such as antibacterial efficacy, hydrophilicity, and resistance against pilling.

Enhancing machine learning algorithm performance through feature selection for driver behavior classification

Machine learning (ML) techniques empower computers to learn from data and make predictions or decisions in various domains, while preprocessing methods assist in cleaning and transforming data before it can be effectively utilized by ML. Feature selection in ML is a critical process that significantly influences the performance and effectiveness of models. By carefully choosing the most relevant and informative attributes from the dataset, feature selection enhances model accuracy, reduces overfitting, and minimizes computational complexity. In this study, we leverage the UAH-DriveSet dataset to classify driver behavior, employing Filter, embedded, and wrapper methods encompassing 10 distinct feature selection techniques. Through the utilization of diverse ML algorithms, we effectively categorize driver behavior into normal, drowsy, and aggressive classes. The second objective is to employ feature selection techniques to pinpoint the most influential features impacting driver behavior. As a results, random forest emerges as the top-performing classifier, achieving an impressive accuracy of 96.4% and an F1-score of 96.36% using backward feature selection in 7.43 s, while K-nearest neighbour (K-NN) attains an accuracy of 96.29% with forward feature selection in 0.05 s. Following our comprehensive results, we deduce that the primary influential features for studying driver behavior include speed (km/h), course, yaw, impact time, road width, distance to the ahead vehicle, vehicle position, and number of detected vehicles.

Hybrid machine learning approach to prediction of the compressive and flexural strengths of UHPC and parametric analysis Shapley additive explanations algorithm

Ultra-high-performance concrete (UHPC) is a sustainable construction material; it can be applied as a substitute for cement concrete. Artificial intelligence methods have been used to evaluate concrete composites to reduce time and money in the construction industries. So, this study applied machine learning (ML) and hybrid ML approaches to predict the compressive and flexural strength of UHPC. A dataset of 626 compressive strength and 317 flexural strength data points was collected from the published research articles, where fourteen important variables were selected as input parameters for the analysis of hybrid ML and ML algorithms. This research used XGBoost, LightGBM, and hybrid XGBoost- LightGBM algorithms to predict UHPC materials. Grid search (GS) techniques were used to adjust model hyper-parameters in search of improved high accuracy and efficiency. ML and hybrid ML models were train, and the test stage utilized statistical assessments such as R-square, root mean square error (RMSE), mean absolute error (MAE), and coefficient of efficiency (CE). The results presented hybrid ML algorithm was superior to the XGBoost and LightGBM algorithms in terms of R-square and RMSE values for both compressive strength and flexural strength prediction. A hybrid ML model and two ML models showed CS considerable R-square values above 0.94 at the testing stages and just over 0.97 at the training phase. Hybrid ML model performance accuracy for CS prediction was R-square value found that almost 0.996 for training and 0.963 for testing phases. At the same time, the FS prediction result showed that the R-square value of the Hybrid ML model and two traditional ML models were founded at almost 0.95 for the training phase and around 0.81 for the testing phase. But among them, the hybrid XGB-LGB model prediction performance was high accuracy and lowest error for CS and FS of UHPC trained and its hyperparameters optimized. Additionally, the SHAP investigation reveals the impact and relationship of the input variables with the output variables. SHAP analysis outcome reveals that curing age and steel fiber content input parameter had the highest positive impact on compressive strength and flexural strength of UHPC.

Analyzing election trends incorporating memory effect through a fractional-order mathematical modeling

This study proposes a new mathematical model to analyze and predict the results of a political election. In general, we predict or analyze the results using statistical methods; however, to minimize the effort of the study, we propose a fractional-order modeling approach. This study proposes a model to analyze and predict general election result trends in India, focusing on the state of West Bengal. To incorporate memory into the model, we consider the Caputo fractional derivative. The model solution’s positivity, boundedness, existence, and uniqueness were tested analytically. Numerical simulations were carried out to investigate the impact of the parameters and evaluate the model’s performance by incorporating the implications of the previous election for realistic situations. Following this, a qualitative analysis of the performance of political parties is discussed, and a prediction of the electoral victory is obtained.

Which pooling method is better: Max, Avg, or Concat (Max, Avg)

Pooling is a non-linear operation that aggregates the results of a given region to a single value. This method effectively removes extraneous details in feature maps while keeping the overall information. As a result, the size of feature maps is reduced, which decreases computing costs and prevents overfitting by eliminating irrelevant data. In CNN models, the max pooling and average pooling methods are commonly utilized. The max pooling selects the highest value within the pooling area and aids in preserving essential features of the image. However, it ignores the other values inside the pooling region, resulting in a significant loss of information. The average pooling computes the average values within the pooling area, which reduces data loss. However, by failing to emphasize critical pixels in the image, it may result in the loss of significant features. To examine the performance of pooling methods, this study comprised the experimental analysis of multiple models, i.e. shallow and deep, datasets, i.e. Cifar10, Cifar100, and SVHN, and pool sizes, e.g. $2x2$, $3x3$, $10x10$. Furthermore, the study investigated the effectiveness of combining two approaches, namely Concat (Max, Avg), to minimize information loss. The findings of this work provide an important guideline for selecting pooling methods in the design of CNNs. The experimental results demonstrate that pooling methods have a considerable impact on model performance. Moreover, there are variances based on the model and pool size.

YOLO Network-based URL Detection in Varied Conditions with Small-Sample Insights

Object detection is a pivotal aspect of computer vision, essential for diverse recognition tasks. This study centers on exploring deep learning methodologies for object detection, specifically targeting the identification of URLs in images captured by mobile phones. We conduct a comparative analysis of three models from the YOLO family – YOLOv3, YOLOv4, and YOLOv5 – recognized for their efficacy in object detection. Our research addresses the unique challenge of detecting URLs in images, particularly considering the limited availability of URL-labeled dataset. Through rigorous experimentation and evaluation, we demonstrate the generalization capabilities of YOLOv3, YOLOv4, and YOLOv5, as measured by average precision scores. Furthermore, we highlight the resilience of the YOLOv4 model against various image-related challenges. Our findings contribute significantly to the advancement of computer vision, specifically in the domain of object detection for real-world applications. By evaluating the performance of cutting-edge deep learning models, we provide valuable insights into their effectiveness for URL detection, thereby enriching our understanding of their practical utility. This research serves as a foundation for future investigations aimed at leveraging deep learning techniques to enhance object detection accuracy across diverse contexts.

An enhanced meta-learning network with sensitivity penalty for cross-domain few-shot fault diagnosis

Abstract Big data-driven rotating machine intelligent diagnostic technology has gained widespread applications. In practice, however, fault data are limited as well as inconsistencies in fault categories among different domains are widespread. These make developing robust intelligent diagnostic models a challenge. To this end, this paper develops an enhanced meta-learning network with a sensitivity penalization mechanism (EMLN-SP) for few-shot fault diagnosis in severe domain bias. First, lightweight channel attention is introduced to establish an enhanced feature encoder under meta-learning framework, which elevates the key feature expression to facilitate the extraction of generalized diagnostic knowledge within limited samples. Second, a boundary-enhanced loss calculation method is designed, which boosts the focus for decision boundary information to prevent the model from the overfitting dilemma in the case of few-shot. Finally, a sensitivity penalty mechanism is constructed to adjust the optimization direction, which prevents the model from falling into a local optimum, to boost the generalization of the model performance. The effectiveness of EMLN-SP is validated by three cross-domain diagnostic cases with diverse domain offsets.

Chest Decompressions - The Driver of CPR Efficacy: Exploring the Relationship Between Compression Rate, Depth, Recoil Velocity, and End-Tidal CO2

Objective Cardiopulmonary arrest survival is dependent on optimization of perfusion via high quality cardiopulmonary resuscitation (CPR), defined by a complex dynamic between rate, depth, and recoil velocity. Here we explore the interaction between these metrics and create a model that explores the impact of these variables on compression efficacy. Methods This study was performed in a large urban/suburban fire-based emergency medical services (EMS) system over a nine-month period from 2019 to 2020. Manual chest compression parameters [rate/depth/recoil velocity] from a cohort of out-of-hospital cardiac arrest (OOHCA) victims were abstracted from monitor defibrillators (ZOLL X-series) and end-tidal carbon dioxide (ETCO2) from sensors. The mean values of these parameters were modeled against each other using multiple regression and structural equation modeling with ETCO2 as a dependent variable. Results Data from a total of 335 patients were analyzed. Strong linear relationships were observed between compression depth/recoil velocity (r = .87, p < .001), ETCO2/depth (r = .23, p < .001) and ETCO2/recoil velocity (r = .61, p < .001). Parabolic relationships were observed between rate/depth (r = .39, p < .001), rate/recoil velocity (r = .26, p < .001), and ETCO2/rate (r = .20, p = .003). Rate, depth, and recoil velocity were modeled as independent variables and ETCO2 as a dependent variable with excellence model performance suggesting the primary driver of stroke volume to be recoil velocity rather than compression depth. Conclusions We used manual CPR metrics from out of hospital cardiac arrests to model the relationship between CPR metrics. These results consistently support the importance of chest recoil on CPR hemodynamics, suggesting that guidelines for optimal CPR should emphasize the importance of maximum chest recoil.

Robust and transferable end-to-end navigation against disturbances and external attacks: an adversarial training approach

Purpose This paper aims to facilitate the research and development of resilient navigation approaches, explore the robustness of adversarial training to different interferences and promote their practical applications in real complex environments. Design/methodology/approach In this paper, the authors first summarize the real accidents of self-driving cars and develop a set of methods to simulate challenging scenarios by introducing simulated disturbances and attacks into the input sensor data. Then a robust and transferable adversarial training approach is proposed to improve the performance and resilience of current navigation models, followed by a multi-modality fusion-based end-to-end navigation network to demonstrate real-world performance of the methods. In addition, an augmented self-driving simulator with designed evaluation metrics is built to evaluate navigation models. Findings Synthetical experiments in simulator demonstrate the robustness and transferability of the proposed adversarial training strategy. The simulation function flow can also be used for promoting any robust perception or navigation researches. Then a multi-modality fusion-based navigation framework is proposed as a light-weight model to evaluate the adversarial training method in real-world. Originality/value The adversarial training approach provides a transferable and robust enhancement for navigation models both in simulation and real-world.

Performance evaluation of different regression models: application in a breast cancer patient data

This paper provides a comprehensive analysis of linear regression models, focusing on addressing multicollinearity challenges in breast cancer patient data. Linear regression methodologies, including GAM, Beta, GAM Beta, Ridge, and Beta Ridge, are compared using two statistical criteria. The study, conducted with R software, showcases the Beta regression model’s exceptional performance, achieving a BIC of − 5520.416. Furthermore, the Ridge regression model demonstrates remarkable results with the best AIC at − 8002.647. The findings underscore the practical application of these models in real-world scenarios and emphasize the Beta regression model’s superior ability to handle multicollinearity challenges. The preference for AIC over BIC in Generalized Additive Models (GAMs) is rooted in the AIC’s calculation framework, highlighting its effectiveness in capturing the complexity and flexibility inherent in GAMs.

Rapid quantitative analysis of coal composition using laser-induced breakdown spectroscopy coupled with random forest algorithm.

Coal is the primary energy source in China, widely used in energy production, industrial processes, and chemical engineering. Due to the complexity and diversity of coal quality, there is an urgent need for new technologies to achieve rapid and accurate detection and analysis of coal, aiming to improve coal resource utilization and reduce pollutant emissions. This study proposes a rapid quantitative analysis of coal using laser-induced breakdown spectroscopy combined with the random forest algorithm. Firstly, a Q-switched Nd: YAG laser at 1064nm was employed to ablate coal samples, generating plasma, and spectral data were collected using a spectrometer. Secondly, the study explores the impact of different parameters in the preprocessing method (wavelet transform) on the predictive performance of the random forest model. It identifies elements related to coal ash content and calorific value along with their spectral information. Subsequently, to further validate the predictive performance of the model, a comparison is made with models established using support vector machine, artificial neural network, and partial least squares. Finally, under optimal parameters for spectral information preprocessing (wavelet transform with Db4 as the base function and 3 decomposition levels), a model combining wavelet transform with Random Forest is established to predict and analyze the ash content and calorific value of coal. The results demonstrate that the Wavelet Transform-Random Forest model exhibits excellent predictive performance (coal ash content: R2 = 0.9470, RMSECV = 4.8594, RMSEP = 4.8450; coal calorific value: R2 = 0.9485, RMSECV = 1.5996, RMSEP = 1.5949). Therefore, laser-induced breakdown spectroscopy combined with the random forest algorithm is an effective method for rapid and accurate detection and analysis of coal. The predicted coal composition values show high accuracy, providing insights and methods for coal composition monitoring and analysis.

Resolving spatial response heterogeneity in glioblastoma.

Spatial intratumoral heterogeneity poses a significant challenge for accurate response assessment in glioblastoma. Multimodal imaging coupled with advanced image analysis has the potential to unravel this response heterogeneity. Based on automated tumor segmentation and longitudinal registration with follow-up imaging, we categorized contrast-enhancing voxels of 61 patients with suspected recurrence of glioblastoma into either true tumor progression (TP) or pseudoprogression (PsP). To allow the unbiased analysis of semantically related image regions, adjacent voxels with similar values of cerebral blood volume (CBV), FET-PET, and contrast-enhanced T1w were automatically grouped into supervoxels. We then extracted first-order statistics as well as texture features from each supervoxel. With these features, a Random Forest classifier was trained and validated employing a 10-fold cross-validation scheme. For model evaluation, the area under the receiver operating curve, as well as classification performance metrics were calculated. Our image analysis pipeline enabled reliable spatial assessment of tumor response. The predictive model reached an accuracy of 80.0% and a macro-weighted AUC of 0.875, which takes class imbalance into account, in the hold-out samples from cross-validation on supervoxel level. Analysis of feature importances confirmed the significant role of FET-PET-derived features. Accordingly, TP- and PsP-labeled supervoxels differed significantly in their 10th and 90th percentile, as well as the median of tumor-to-background normalized FET-PET. However, CBV- and T1c-related features also relevantly contributed to the model's performance. Disentangling the intratumoral heterogeneity in glioblastoma holds immense promise for advancing precise local response evaluation and thereby also informing more personalized and localized treatment strategies in the future.

Computed tomography radiomics in predicting patient satisfaction after robotic-assisted total knee arthroplasty.

After robotic-assisted total knee arthroplasty (RA-TKA) surgery, some patients still experience joint discomfort. We aimed to establish an effective machine learning model that integrates radiomic features extracted from computed tomography (CT) scans and relevant clinical information to predict patient satisfaction three months postoperatively following RA-TKA. After careful selection, data from 142 patients were randomly divided into a training set (n = 99) and a test set (n = 43), approximately in a 7:3 ratio. A total of 1329 radiomic features were extracted from the regions of interest delineated in CT scans. The features were standardized using normalization algorithms, and the least absolute shrinkage and selection operator regression model was employed to select radiomic features with ICC > 0.75 and P < 0.05, generating the Rad-score as feature markers. Univariate and multivariate logistic regression was then used to screen clinical information (age, body mass index, operation time, gender, surgical side, comorbidities, preoperative KSS score, preoperative range of motion (ROM), preoperative and postoperative HKA angle, preoperative and postoperative VAS score) as potential predictive factors. The satisfaction scale ≥ 20 indicates patient satisfaction. Finally, three prediction models were established, focusing on radiomic features, clinical features, and their fusion. Model performance was evaluated using Receiver Operating Characteristic curves and decision curve analysis. In the training set, the area under the curve (AUC) of the clinical model was 0.793 (95% CI 0.681-0.906), the radiomic model was 0.854 (95% CI 0.743-0.964), and the combined radiomic-clinical model was 0.899 (95% CI 0.804-0.995). In the test set, the AUC of the clinical model was 0.908 (95% CI 0.814-1.000), the radiomic model was 0.709 (95% CI 0.541-0.878), and the combined radiomic-clinical model was 0.928 (95% CI 0.842-1.000). The AUC of the radiomic-clinical model was significantly higher than the other two models. The decision curve analysis indicated its clinical application value. We developed a radiomic-based nomogram model using CT imaging to predict the satisfaction of RA-TKA patients at 3 months postoperatively. This model integrated clinical and radiomic features and demonstrated good predictive performance and excellent clinical application potential.

Survival after Resection of Malignant Peripheral Nerve Sheath Tumors: introducing and validating a novel type-specific prognostic model

Abstract Background This study aimed to assess the performance of currently available risk calculators in a cohort of patients with malignant peripheral nerve sheath tumors (MPNST) and to create an MPNST-specific prognostic model including type-specific predictors for overall survival (OS). Methods This is a retrospective multicenter cohort study of patients with MPNST from eleven secondary or tertiary centers in The Netherlands, Italy and the USA. All patients diagnosed with primary MPNST who underwent macroscopically complete surgical resection from 2000-2019 were included in this study. A multivariable Cox proportional hazard model for OS was estimated with pre-specified predictors (age, grade, size, NF-1 status, triton status, depth, tumor location and surgical margin). Model performance was assessed for the Sarculator and PERSARC calculators by examining discrimination (C-index) and calibration (calibration plots and observed-expected statistic; O/E-statistic). Internal-external cross-validation by different regions was performed to evaluate the generalizability of the model. Results A total of 507 patients with primary MPNSTs were included from 11 centers in 7 regions. During follow-up (median 8.7 years), 211 patients died. The C-index was 0.60 (95% CI 0.53-0.67) for both Sarculator and PERSARC. The MPNST-specific model had a pooled C-index of 0.69 (95%CI 0.65-0.73) at validation, with adequate discrimination and calibration across regions. Conclusions The MPNST-specific MONACO model can be used to predict 3-, 5-, and 10-year OS in patients with primary MPNST who underwent macroscopically complete surgical resection. Further validation may refine the model to inform patients and physicians on prognosis and support them in shared decision-making.

Predicting treatment response in multicenter non-small cell lung cancer patients based on federated learning

BackgroundMulticenter non-small cell lung cancer (NSCLC) patient data is information-rich. However, its direct integration becomes exceptionally challenging due to constraints involving different healthcare organizations and regulations. Traditional centralized machine learning methods require centralizing these sensitive medical data for training, posing risks of patient privacy leakage and data security issues. In this context, federated learning (FL) has attracted much attention as a distributed machine learning framework. It effectively addresses this contradiction by preserving data locally, conducting local model training, and aggregating model parameters. This approach enables the utilization of multicenter data with maximum benefit while ensuring privacy safeguards. Based on pre-radiotherapy planning target volume images of NSCLC patients, a multicenter treatment response prediction model is designed by FL for predicting the probability of remission of NSCLC patients. This approach ensures medical data privacy, high prediction accuracy and computing efficiency, offering valuable insights for clinical decision-making.MethodsWe retrospectively collected CT images from 245 NSCLC patients undergoing chemotherapy and radiotherapy (CRT) in four Chinese hospitals. In a simulation environment, we compared the performance of the centralized deep learning (DL) model with that of the FL model using data from two sites. Additionally, due to the unavailability of data from one hospital, we established a real-world FL model using data from three sites. Assessments were conducted using measures such as accuracy, receiver operating characteristic curve, and confusion matrices.ResultsThe model’s prediction performance obtained using FL methods outperforms that of traditional centralized learning methods. In the comparative experiment, the DL model achieves an AUC of 0.718/0.695, while the FL model demonstrates an AUC of 0.725/0.689, with real-world FL model achieving an AUC of 0.698/0.672.ConclusionsWe demonstrate that the performance of a FL predictive model, developed by combining convolutional neural networks (CNNs) with data from multiple medical centers, is comparable to that of a traditional DL model obtained through centralized training. It can efficiently predict CRT treatment response in NSCLC patients while preserving privacy.

Fracture prediction in a Swiss cohort.

Fracture prediction is essential in managing patients with osteoporosis, and is an integral component of many fracture prevention guidelines. We aimed to identify the most relevant clinical fracture risk factors in contemporary populations by training and validating short- and long-term fracture risk prediction models in two cohorts. We used traditional and machine learning survival models to predict risks of vertebral, hip and any fractures on the basis of clinical risk factors, T-scores and treatment history among participants in a nationwide Swiss osteoporosis registry (N = 5944 postmenopausal women, median follow-up of 4.1years between January 2015 and October 2022; a total of 1190 fractures during follow-up). The independent validation cohort comprised 5474 postmenopausal women from the UK Biobank with 290 incident fractures during follow-up. Uno's C-index and the time-dependent area under the receiver operating characteristics curve were calculated to evaluate the performance of different machine learning models (Random survival forests and eXtreme Gradient Boosting). In the independent validation set, the C-index was 0.74 [0.58, 0.86] for vertebral fractures, 0.83 [0.7, 0.94] for hip fractures and 0.63 [0.58, 0.69] for any fractures at year 2, and these values further increased for longer estimations of up to 7years. In comparison, the 10-year fracture probability calculated with FRAX® Switzerland was 0.60 [0.55, 0.64] for major osteoporotic fractures and 0.62 [0.49, 0.74] for hip fractures. The most important variables identified with Shapley additive explanations (SHAP) values were age, T-scores and prior fractures, while number of falls was an important predictor of hip fractures. Performances of both traditional and machine learning models showed similar C-indices. We conclude that fracture risk can be improved by including the lumbar spine T-score, trabecular bone score, numbers of falls and recent fractures, and treatment information has a significant impact on fracture prediction.

Stitched vision transformer for age-related macular degeneration detection using retinal optical coherence tomography images.

Age-related macular degeneration (AMD) is an eye disease that leads to the deterioration of the central vision area of the eye and can gradually result in vision loss in elderly individuals. Early identification of this disease can significantly impact patient treatment outcomes. Furthermore, given the increasing elderly population globally, the importance of automated methods for rapidly monitoring at-risk individuals and accurately diagnosing AMD is growing daily. One standard method for diagnosing AMD is using optical coherence tomography (OCT) images as a non-invasive imaging technology. In recent years, numerous deep neural networks have been proposed for the classification of OCT images. Utilizing pre-trained neural networks can speed up model deployment in related tasks without compromising accuracy. However, most previous methods overlook the feasibility of leveraging pre-existing trained networks to search for an optimal architecture for AMD staging on a new target dataset. In this study, our objective was to achieve an optimal architecture in the efficiency-accuracy trade-off for classifying retinal OCT images. To this end, we employed pre-trained medical vision transformer (MedViT) models. MedViT combines convolutional and transformer neural networks, explicitly designed for medical image classification. Our approach involved pre-training two distinct MedViT models on a source dataset with labels identical to those in the target dataset. This pre-training was conducted in a supervised manner. Subsequently, we evaluated the performance of the pre-trained MedViT models for classifying retinal OCT images from the target Noor Eye Hospital (NEH) dataset into the normal, drusen, and choroidal neovascularization (CNV) classes in zero-shot settings and through five-fold cross-validation. Then, we proposed a stitching approach to search for an optimal model from two MedViT family models. The proposed stitching method is an efficient architecture search algorithm known as stitchable neural networks. Stitchable neural networks create a candidate model in search space for each pair of stitchable layers by inserting a linear layer between them. A pair of stitchable layers consists of layers, each selected from one input model. While stitchable neural networks had previously been tested on more extensive and general datasets, this study demonstrated that stitching networks could also be helpful in smaller medical datasets. The results of this approach indicate that when pre-trained models were available for OCT images from another dataset, it was possible to achieve a model in 100 epochs with an accuracy of over 94.9% in classifying images from the NEH dataset. The results of this study demonstrate the efficacy of stitchable neural networks as a fine-tuning method for OCT image classification. This approach not only leads to higher accuracy but also considers architecture optimization at a reasonable computational cost.