Performance Of Model Research Articles

Exploring the performance of large language models on hepatitis B infection-related questions: A comparative study

BACKGROUND Patients with hepatitis B virus (HBV) infection require chronic and personalized care to improve outcomes. Large language models (LLMs) can potentially provide medical information for patients. AIM To examine the performance of three LLMs, ChatGPT-3.5, ChatGPT-4.0, and Google Gemini, in answering HBV-related questions. METHODS LLMs’ responses to HBV-related questions were independently graded by two medical professionals using a four-point accuracy scale, and disagreements were resolved by a third reviewer. Each question was run three times using three LLMs. Readability was assessed via the Gunning Fog index and Flesch-Kincaid grade level. RESULTS Overall, all three LLM chatbots achieved high average accuracy scores for subjective questions (ChatGPT-3.5: 3.50; ChatGPT-4.0: 3.69; Google Gemini: 3.53, out of a maximum score of 4). With respect to objective questions, ChatGPT-4.0 achieved an 80.8% accuracy rate, compared with 62.9% for ChatGPT-3.5 and 73.1% for Google Gemini. Across the six domains, ChatGPT-4.0 performed better in terms of diagnosis, whereas Google Gemini demonstrated excellent clinical manifestations. Notably, in the readability analysis, the mean Gunning Fog index and Flesch-Kincaid grade level scores of the three LLM chatbots were significantly higher than the standard level eight, far exceeding the reading level of the normal population. CONCLUSION Our results highlight the potential of LLMs, especially ChatGPT-4.0, for delivering responses to HBV-related questions. LLMs may be an adjunctive informational tool for patients and physicians to improve outcomes. Nevertheless, current LLMs should not replace personalized treatment recommendations from physicians in the management of HBV infection.

A multi-constraint representation learning model for identification of ovarian cancer with missing laboratory indicators.

To evaluate the performance of a multi-constraint representation learning classification model for identifying ovarian cancer with missing laboratory indicators. Tabular data with missing laboratory indicators were collected from 393 patients with ovarian cancer and 1951 control patients. The missing ovarian cancer laboratory indicator features were projected to the latent space to obtain a classification model using the representational learning classification model based on discriminative learning and mutual information coupled with feature projection significance score consistency and missing location estimation. The proposed constraint term was ablated experimentally to assess the feasibility and validity of the constraint term by accuracy, area under the ROC curve (AUC), sensitivity, and specificity. Cross-validation methods and accuracy, AUC, sensitivity and specificity were also used to evaluate the discriminative performance of this classification model in comparison with other interpolation methods for processing of the missing data. The results of the ablation experiments showed good compatibility among the constraints, and each constraint had good robustness. The cross-validation experiment showed that for identification of ovarian cancer with missing laboratory indicators, the AUC, accuracy, sensitivity and specificity of the proposed multi-constraints representation-based learning classification model was 0.915, 0.888, 0.774, and 0.910, respectively, and its AUC and sensitivity were superior to those of other interpolation methods. The proposed model has excellent discriminatory ability with better performance than other missing data interpolation methods for identification of ovarian cancer with missing laboratory indicators.

Development and Validation of KCPREDICT: A Deep Learning Model for Early Detection of Coronary Artery Lesions in Kawasaki Disease Patients.

Kawasaki disease (KD) is a febrile vasculitis disorder, with coronary artery lesions (CALs) being the most severe complication. Early detection of CALs is challenging due to limitations in echocardiographic equipment (UCG). This study aimed to develop and validate an artificial intelligence algorithm to distinguish CALs in KD patients and support diagnostic decision-making at admission. A deep learning algorithm named KCPREDICT was developed using 24 features, including basic patient information, five classic KD clinical signs, and 14 laboratory measurements. Data were collected from patients diagnosed with KD between February 2017 and May 2023 at Shanghai Children's Medical Center. Patients were split into training and internal validation cohorts at an 80:20 ratio, and fivefold cross-validation was employed to assess model performance. Among the 1474 KD cases, the decision tree model performed best during the full feature experiment, achieving an accuracy of 95.42%, a precision of 98.83%, a recall of 93.58%, an F1 score of 96.14%, and an area under the receiver operating characteristic curve (AUROC) of 96.00%. The KCPREDICT algorithm can aid frontline clinicians in distinguishing KD patients with and without CALs, facilitating timely treatment and prevention of severe complications. The use of the complete set of 24 diagnostic features is the optimal choice for predicting CALs in children with KD.

A privacy-enhanced framework with deep learning for botnet detection

A botnet is a group of hijacked devices that conduct various cyberattacks, which is one of the most dangerous threats on the internet. Organizations or individuals use network traffic to mine botnet communication behavior features. Network traffic often contains individual users’ private information, such as website passwords, personally identifiable information, and communication content. Among the existing botnet detection methods, whether they extract deterministic traffic interaction features, use DNS traffic, or methods based on raw traffic bytes, these methods focus on the detection performance of the detection model and ignore possible privacy leaks. And most methods are combined with machine learning and deep learning technologies, which require a large amount of training data to obtain high-precision detection models. Therefore, preventing malicious persons from stealing data to infer privacy during the botnet detection process has become an issue worth pondering. Based on this problem, this article proposes a privacy-enhanced framework with deep learning for botnet detection. The goal of this framework is to learn a feature extractor. It can hide the private information that the attack model tries to infer from the intermediate anonymity features, while maximally retaining the interactive behavior features contained in the original traffic for botnet detection. We design a privacy confrontation algorithm based on a mutual information calculation mechanism. This algorithm simulates the game between the attacker trying to infer private information through the attack model and the data processor retaining the original content of the traffic to the maximum extent. In order to further ensure the privacy protection of the feature extractor during the training process, we train the feature extractor in the federated learning training mode. We extensively evaluate our approach, validating it on two public datasets and comparing it with existing methods. The results show that our method can effectively ensure detection accuracy on the basis of removing private information. For the CTU-13 dataset, the detection framework achieves the best detection performance; for the ISCX-2014 dataset, the accuracy of the framework is less than 1% lower than the best effect.

Prevalence and Risk Factors of Deep Vein Thrombosis with Acute Stroke: A Cross-sectional Study

Background: Deep vein thrombosis (DVT) is a serious complication in stroke patients and may lead to the devastating consequences of a pulmonary embolism. The risk factors for DVT in acute stroke are advanced age, high National Institute of Health Stroke Scale (NIHSS) score, hemiparesis, immobility, female gender, atrial fibrillation. The present study is aimed at identifying the prevalence and risk factors for DVT in acute stroke patients. Methods: This was a prospective cross sectional study conducted in of Dhaka Medical College and Hospital from January 2022 to December 2023. Data was collected in a preformed questionnaire. Statistical analysis was performed using IBM SPSS version 26.0. Continuous variables were analyzed in the form of the means with standard deviations (Mean ± SD). Categorical variables were shown as numbers and proportions. Continuous data were analyzed using the independent samples t-test. Categorical data were analyzed using the chi-squared test. Univariate variables with a p-value of <0.05 were retained in a binary logistic regression model. Receiver operating characteristic (ROC) curve analysis was performed by identifying the area under the ROC curve (AUC) to evaluate the performance of the logistic regression model. Statistical significance was defined as a p-value of < 0.05. Results: A total of 106 patients with acute stroke (AS) were included in the study. The included patients were divided into two groups: DVT and non-DVT group. Out of 106 patients, 34(32.07%) patients presented with DVT. Patients with DVT were older than those without DVT (62.7 ± 10.7 vs. 56.5 ± 11.9 years, p < 0.05). Females were more affected by DVT (58.88% vs. 26.85%, p<0.05). In the DVT group, the level of D-dimer was significantly higher than that in the non- DVT group (2.2 ± 1.7 vs. 0.45 ± 1.1, p < 0.05). Elevated CRP and low serum albumin level were also observed in DVT group which was statistically significant. (p<0.05). Multiple logistic regression analysis showed that age>60years, NIHSS score>25 and a D-dimer level of ≥ 1.52 μg/mL and caprine score >6 were independent factors for the presence of DVT at the time of stroke admission. Odds ratio for age, Caprini RAM score, NIHSS score and elevated d-dimer were 1.043(1.000–1.075, p=0.029),1.43(0.85-0.1.96, p=0.001),1.266(1.168-1.284, p=0.008) and1.446 (1.130–1.849, p=0.003) respectively. Conclusion: The present study has showed that advanced age, high NIHSS and caprini score and raised serum D-dimer levels are independent predictors of DVT in stroke patients. A further large multicenter study should be done to provide a better management to the patients of DVT in stroke patients.

Advancements in Frank’s sign Identification using deep learning on 3D brain MRI

Frank’s sign (FS) is a diagnostic marker associated with aging and various health conditions. Despite its clinical significance, there lacks a standardized method for its identification. This study aimed to develop a deep learning model for automated FS detection in 3D facial images derived from MRI scans. Four deep learning architectures were evaluated for FS segmentation on a dataset of 400 brain MRI scans. The optimal model was subsequently validated on two external datasets, comprising 300 brain MRI scans each with varying FS presence. Dice similarity coefficient (DSC) and receiver operating characteristic (ROC) analysis were employed to assess model performance. The U-net architecture demonstrated superior performance in terms of accuracy and efficiency. On the validation datasets, the model achieved a DSC of 0.734, an intra-class correlation coefficient of 0.865, and an area under the ROC curve greater than 0.9 for FS detection. Additionally, the model identified optimal voxel thresholds for accurate FS classification, resulting in high sensitivity, specificity, and accuracy metrics. This study successfully developed a deep learning model for automated FS segmentation in MRI scans. This tool has the potential to enhance FS identification in clinical practice and contribute to further research on FS and its associated health implications.

Conserving saproxylic flagship species by complementing 150 years of natural history with citizen science data—the case of the stag beetles (Lucanidae, Coleoptera) of Portugal

Abstract The use of flagship species in conservation, such as the European Stag Beetle (Lucanus cervus), can positively affect the conservation of other taxonomic or functionally related species. This is especially true for taxonomic groups where the knowledge regarding their distribution and ecology is generally limited, as is the case of insects. This work represents the most accurate and up-to-date publication on the distribution, ecology and environmental niche of the four lucanids found in Portugal. Moreover, it unravels how citizen science initiatives can positively impact conservation planning for flagship species and their taxonomic and functionally related species. Compared to non-citizen science sources, citizen science data increased the known distribution area for all four species (1.6 to 7.2 fold) and confirmed most of the current known species distribution (50.0% to 95.5%). It also expanded the known environmental niche breadth for all species (1.1 to 4.2 fold) and the species suitable modelled habitat (1.1 to 2.1 fold) with very good model performances (AUC > 0.8 and TSS > 0.6). Our results suggest that to improve the conservation of these species, it is paramount to value native forests in Portugal. Management actions to accomplish this might include preserving current native forests and small natural features (such as large old trees) throughout the landscape while simultaneously increasing native forest development. In addition, climate change will affect the species' distribution since all are highly susceptible to specific (bio-)climatic conditions that are expected to change in the next decades. These findings not only enhance our understanding of these species' current distribution but also pave the way for targeted and integrative conservation efforts throughout the country, especially in areas where the (now assessed) ecological niche requirements and habitat availability (either in terms of landscape composition or structure) are met. Guidelines and priorities are discussed on how to implement conservation measures in Portugal following the recently approved European Nature Restoration Law.

Quantitative Pupillometry Predicts Neurologic Deterioration in Patients with Large Middle Cerebral Artery Stroke.

This study assesses whether longitudinal quantitative pupillometry predicts neurological deterioration after large middle cerebral artery (MCA) stroke and determines how early changes are detectable. This prospective, single-center observational cohort study included patients with large MCA stroke admitted to Boston Medical Center's intensive care unit (2019-2024). Associations between time-to-neurologic deterioration and quantitative pupillometry, including Neurological Pupil Index (NPi), were assessed using Cox proportional hazards models with time-dependent covariates adjusted for age, sex, and Alberta Stroke Program Early CT Score. Models using dilation velocity were compared with partial likelihood ratio tests. Pupillometric changes over 2-h intervals in the 12 h preceding deterioration were analyzed with linear mixed-effects modeling and Tukey's test. Matched referents (age, sex, stroke side, follow-up duration) were used for comparison. Optimal thresholds were identified using the Youden Index. Among 71 patients (mean age 66.5 years; 59.2% women), 32 (45.1%) experienced deterioration. A 1-unit decrease in NPi was associated with a higher hazard of deterioration (hazard ratio 2.46; 95% confidence interval 1.68-3.61). Dilation velocity improved model performance compared to NPi alone. NPi was significantly lower at 0-2 h (3.81 vs. 4.38, p = 0.001) and 2-4 h (3.71 vs. 4.38, p < 0.001) before deterioration compared to 10-12 h prior. Optimal thresholds were 4.01 for NPi, 0.49 mm/s for dilation velocity, and -0.15 change in NPi over 12 h. Quantitative pupillometry predicts neurological deterioration in MCA stroke, with declines detectable up to 12 h prior. Dilation velocity shows promise as a novel biomarker. ANN NEUROL 2025.

Dependence of rainfall–runoff model performance on calibration conditions under changing climatic conditions

ABSTRACT Calibration period selection is crucial for rainfall–runoff (RR) models in stationary and non-stationary climates. In this study, we assessed the impact of four calibration conditions on RR model performance: calibration duration, temporal lag between calibration and simulation periods, variation in rainfall volume, and rainfall patterns similarity. To ensure the generalizability of our findings, we used five RR models (GR4J, Simhyd, WASMOD, HBV, and XAJ) to simulate streamflow across two catchments in Iran with stationary and non-stationary climate conditions. Our results show that catchments exhibiting non-stationary RR relationships necessitate longer calibration periods to obtain stable simulations. Model performance gradually drops as the temporal lag between the calibration and simulation periods increases. This drop in performance is even more pronounced in catchments experiencing non-stationary climate conditions. Similarity in rainfall volume is more important than rainfall pattern similarity for predicting future streamflow in non-stationary RR relationships.

Developing and implementing AI-driven models for demand forecasting in US supply chains: A comprehensive approach to enhancing predictive accuracy

Demand forecasting has long been a critical challenge in the US supply chain operations, plagued by disruptions, fluctuating demand, and price volatility. Developing and implementing AI models that can accurately predict demand is essential in response to these issues. This study aimed to investigate the feasibility of applying machine learning techniques to demand forecasting, particularly in supply chain operations. A comprehensive analysis was conducted using historical data from a logistics company in the USA, which was used to train five traditional demand forecasting methods: Linear Regression, ElasticNet, Random Forest, MLPRegressorn, and XGBoost. Additionally, feature selection, data normalization, and dimensionality reduction techniques were employed to improve the accuracy of these models. Strategic metrics were used to evaluate the model's performance: Random Mean Squared Error, Mean Absolute Error, and R-squared score. The results of this study indicate that AI models have shown significant promise in predicting target sales in supply chains with Linear Regression emerging as the most effective model with the lowest RMSE, MAE, and an R-squared score close to 1. Practical implications of implementing such AI models in the US supply chain include optimized inventory management, reduced costs, and enhanced customer satisfaction. This research contributes to the existing body of knowledge on AI applications in demand forecasting, highlights the importance of traditional methods being supplemented by machine learning techniques, and provides practical recommendations for businesses seeking to improve their supply chain operations.

Machine learning to predict periprosthetic joint infections following primary total hip arthroplasty using a national database.

Periprosthetic joint infection (PJI) following total hip arthroplasty (THA) remains a devastating complication for patients and surgeons. Given the implications of these infections and the current paucity of risk calculators utilizing machine learning (ML), this study aimed to develop an ML algorithm that could accurately identify risk factors for developing a PJI following primary THA using a national database. A total of 51,053 patients who underwent primary THA between 2013 and 2020 were identified using the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) database. Demographic, preoperative, intraoperative, and immediate postoperative outcomes were collected. Five ML models were created. The receiver operating characteristic curves, the area under the curve (AUC), calibration plots, slopes, intercepts, and Brier scores were evaluated. The histogram-based gradient boosting (HGB) model demonstrated good PJI discriminatory ability with an AUC of 0.88. The test-specific metrics supported the model's performance and validation in predicting PJI (calibration curve slope: 0.79; intercept: 0.32; Brier score: 0.007). The top five predictors of PJI were the length of stay (>3 days), patient weight at the time of surgery (>94.3 kg), an American Society of Anesthesiologists (ASA) class of 4 or higher, preoperative platelet count (<249,890/mm3), and preoperative sodium (<139.5 mEq/L). This study developed a highly specific ML model that could predict patient-specific PJI development following primary THA. Considering the feature importance of the top predictors of infection, surgeons should counsel at-risk patients to optimize resource utilization and potentially improve surgical outcomes.

EP1 External validation of HAS model in predicting mortality after emergency laparotomy: A retrospective cohort study

Abstract Aims The HAS (Hajibandeh Index, American Society of Anaesthesiologists status, and sarcopenia) model has demonstrated superior predictive ability compared to the NELA (National Emergency Laparotomy Audit) score for mortality after emergency laparotomy. However, external validation of this model is lacking. This study aimed to validate the HAS model's performance in predicting mortality after emergency laparotomy and compare its predictive accuracy with the NELA score. Methods This retrospective cohort study included all adult patients (&amp;gt;17 years) who underwent emergency laparotomy between January 2022 and June 2023 at an urban district general hospital. Receiver operating characteristic curve analysis was used to compare the HAS model and NELA score in predicting 30-day mortality. Subgroup analysis was conducted for age groups: ≥50, ≥60, ≥70, and ≥80. Results A total of 117 patients were included. The area under the curve (AUC) for the HAS model in predicting 30-day mortality was 0.90 (95% CI 0.83-0.95). There was no statistically significant difference in AUC between the HAS and NELA scores for all patients (0.90 vs 0.80, p = 0.097). However, the AUC of the HAS model was significantly superior to that of the NELA score in patients aged ≥50 (0.89 vs 0.75, p = 0.040), ≥60 (0.87 vs 0.69, p = 0.020), ≥70 (0.85 vs 0.67, p = 0.030), and ≥80 (0.90 vs 0.66, p &amp;lt; 0.001). Conclusions This study provides external validation of the HAS model for predicting 30-day mortality after emergency laparotomy. Prospective studies with larger sample sizes are warranted to confirm these findings.

From Prediction to Explanation: Using Explainable AI to Understand Satellite-Based Riot Forecasting Models

This study investigates the application of explainable AI (XAI) techniques to understand the deep learning models used for predicting urban conflict from satellite imagery. First, a ResNet18 convolutional neural network achieved 89% accuracy in distinguishing riot and non-riot urban areas. Using the Score-CAM technique, regions critical to the model’s predictions were identified, and masking these areas caused a 20.9% drop in the classification accuracy, highlighting their importance. However, Score-CAM’s ability to consistently localize key features was found to be limited, particularly in complex, multi-object urban environments. Analysis revealed minimal alignment between the model-identified features and traditional land use metrics, suggesting that deep learning captures unique patterns not represented in existing GIS datasets. These findings underscore the potential of deep learning to uncover previously unrecognized socio-spatial dynamics while revealing the need for improved interpretability methods. This work sets the stage for future research to enhance explainable AI techniques, bridging the gap between model performance and interpretability and advancing our understanding of urban conflict drivers.

PLASMA: Partial LeAst Squares for Multiomics Analysis

Background/Objectives: Recent growth in the number and applications of high-throughput “omics” technologies has created a need for better methods to integrate multiomics data. Much progress has been made in developing unsupervised methods, but supervised methods have lagged behind. Methods: Here we present the first algorithm, PLASMA, that can learn to predict time-to-event outcomes from multiomics data sets, even when some samples have only been assayed on a subset of the omics data sets. PLASMA uses two layers of existing partial least squares algorithms to first select components that covary with the outcome and then construct a joint Cox proportional hazards model. Results: We apply PLASMA to the stomach adenocarcinoma (STAD) data from The Cancer Genome Atlas. We validate the model both by splitting the STAD data into training and test sets and by applying them to the subset of esophageal cancer (ESCA) containing adenocarcinomas. We use the other half of the ESCA data, which contains squamous cell carcinomas dissimilar to STAD, as a negative comparison. Our model successfully separates both the STAD test set (p = 2.73 × 10−8) and the independent ESCA adenocarcinoma data (p = 0.025) into high-risk and low-risk patients. It does not separate the negative comparison data set (ESCA squamous cell carcinomas, p = 0.57). The performance of the unified multiomics model is superior to that of individually trained models and is also superior to an unsupervised method (Multi-Omics Factor Analysis; MOFA), which finds latent factors to be used as putative predictors in a post hoc survival analysis. Conclusions: Many of the factors that contribute strongly to the PLASMA model can be justified from the biological literature.

Predictors of new persistent opioid use after surgery in adults

PurposePersistent opioid use is one of the most common post-operative complications. Identification of at-risk patients pre-operatively is key to reducing post-operative opioid use. We sought to develop a predictive model for persistent post-operative opioid used and to determine if geographic factors from community databases improve model prediction based solely on electronic health records (EHRs) and claims data.MethodsEHR and claims data for 4,116 opioid-naïve surgical patients older than 18 in North Carolina were linked with census tract-level unemployment data from the American Community Survey and Centers for Disease Control and Prevention data on opioid prescriptions and deaths attributed to drug poisoning. Primary outcome was new persistent opioid use and covariates included patient factors from EHR, claims data, and geographic factors. Multivariable logistic regression models of potential risk factors were evaluated.Results6.0% of patients developed new persistent opioid use. Associated risk factors based on multivariable logistic regressions include age (adjusted odds ratio [AOR] 1.08; 95% confidence interval [CI] 1.00, 1.16), back and neck pain (1.82; 1.39, 2.39), joint disorders (1.58; 1.18, 2.11), mood disorders (1.71; 1.28, 2.28), opioid retail prescription (1.04; 1.00, 1.07) and drug poisoning rates (1.33; 1.09, 1.62). On Monte-Carlo cross-validation, the addition of geographic factors to EHRs and claims may modestly improve prediction performance (area under the curve, AUC) of logistic regression models compared to those based on EHRs and claims data (AUC 0.667 (95% CI 0.619, 0.717) vs AUC 0.653 (0.600, 0.706)).ConclusionsCo-morbidities and area-based factors are predictive of new persistent post-operative opioid use. As the addition of geographic-based factors did not significantly improve performance of multivariable logistic regression, larger samples are needed to fully differentiate models.

Mitigating bias in AI mortality predictions for minority populations: a transfer learning approach

BackgroundThe COVID-19 pandemic has highlighted the crucial role of artificial intelligence (AI) in predicting mortality and guiding healthcare decisions. However, AI models may perpetuate or exacerbate existing health disparities due to demographic biases, particularly affecting racial and ethnic minorities. The objective of this study is to investigate the demographic biases in AI models predicting COVID-19 mortality and to assess the effectiveness of transfer learning in improving model fairness across diverse demographic groups.MethodsThis retrospective cohort study used a population-based dataset of COVID-19 cases from the Centers for Disease Control and Prevention (CDC), spanning the years 2020–2024. The study analyzed AI model performance across different racial and ethnic groups and employed transfer learning techniques to improve model fairness by adapting pre-trained models to the specific demographic and clinical characteristics of the population.ResultsDecision Tree (DT) and Random Forest (RF) models consistently showed improvements in accuracy, precision, and ROC-AUC scores for Non-Hispanic Black, Hispanic/Latino, and Asian populations. The most significant precision improvement was observed in the DT model for Hispanic/Latino individuals, which increased from 0.3805 to 0.5265. The precision for Asians or Pacific Islanders in the DT model increased from 0.4727 to 0.6071, and for Non-Hispanic Blacks, it rose from 0.5492 to 0.6657. Gradient Boosting Machines (GBM) produced mixed results, showing accuracy and precision improvements for Non-Hispanic Black and Asian groups, but declines for the Hispanic/Latino and American Indian groups, with the most significant decline in precision, which dropped from 0.4612 to 0.2406 in the American Indian group. Logistic Regression (LR) demonstrated minimal changes across all metrics and groups. For the Non-Hispanic American Indian group, most models showed limited benefits, with several performance metrics either remaining stable or declining.ConclusionsThis study demonstrates the potential of AI in predicting COVID-19 mortality while also underscoring the critical need to address demographic biases. The application of transfer learning significantly improved the predictive performance of models across various racial and ethnic groups, suggesting these techniques are effective in mitigating biases and promoting fairness in AI models.

Transformer-based neural marked spatio temporal point process model for analyzing football match events

Predictive modeling plays a crucial role in machine learning, data analysis, and statistics. In sports, predictive modeling methods have emerged to provide insights and evaluate performances based on key performance metrics. However, most existing models tend to focus on predicting only partial aspects of an event, such as the outcome, action type, or location, while neglecting the temporal factors involved. To address this gap, this study introduces the Transformer-Based Neural Marked Spatio-Temporal Point Process (NMSTPP) model, specifically designed for football event data. The NMSTPP model predicts a comprehensive set of future event components, including inter-event time, zone, and action. Additionally, it features a dependent prediction layers architecture to enhance model performance. The Holistic Possession Utilization Score (HPUS) metric is also proposed to evaluate the effectiveness and efficiency of possession periods in football based on the NMSTPP model. With open-source football event data, the NMSTPP model successfully predicted the aforementioned three components of future events, with an improvement of up to 4% overall and 9% for individual components compared to baseline models. The HPUS demonstrated a 0.9 correlation with existing performance metrics, highlighting its utility in performance evaluation. The NMSTPP and HPUS were applied to the Premier League to demonstrate their practical feasibility.

Multiclass Synthetic Accessibility Prediction.

Evaluating synthetic accessibility of in silico molecules is an integral component of the drug discovery process. While the application of machine learning models to predict whether small molecules are easy or hard to synthesize has gained attention recently, predetermined thresholds and data set imbalances present challenges for these binary classification approaches. In this study, we introduce a novel multiclass fold-ensembled classification approach to predict the minimum number of steps needed to synthesize a small molecule. By ensembling the base models trained on multiple stratified subsampled folds, this approach effectively mitigates the impact of class imbalance through probability aggregation or voting aggregation strategies. Additionally, we propose fuzzy evaluation metrics that account for practical tolerances in predictions, providing a more flexible and realistic assessment of model performance. Through experimentation on two reaction benchmark data sets, we demonstrate the effectiveness of our model in a multiclass synthetic accessibility prediction task and the superiority of our proposed method over six existing models in binary synthetic accessibility prediction tasks.

A robust auto-contouring and data augmentation pipeline for adaptive MRI-guided radiotherapy of pancreatic cancer with a limited dataset.

This study aims to develop and evaluate a fast and robust deep learningbased auto-segmentation approach for organs at risk in MRI-guided radiotherapy of pancreatic cancer to overcome the problems of time-intensive manual contouring in online adaptive workflows. The research focuses on implementing novel data augmentation techniques to address the challenges posed by limited datasets.&#xD;Approach: This study was conducted in two phases. In phase I, we selected and customized the best-performing segmentation model among ResU-Net, SegResNet, and nnUNet, using 43 balanced 3DVane images from 10 patients with 5-fold cross-validation. Phase II focused on optimizing the chosen model through two advanced data augmentation approaches to improve performance and generalizability by increasing the effective input dataset: (1) a novel structure-guided deformation-based augmentation approach (sgDefAug) and (2) a generative adversarial network-based method using a cycleGAN (GANAug). These were compared with comprehensive conventional augmentations (ConvAug). The approaches were evaluated using geometric (Dice score, average surface distance (ASD)) and dosimetric (D2% and D50% from dose-volume histograms) criteria.&#xD;Main Results: The nnU-Net framework demonstrated superior performance (mean Dice: 0.78±0.10, mean ASD: 3.92±1.94 mm) compared to other models. The sgDefAug and GANAug approaches significantly improved model performance over ConvAug, with sgDefAug demonstrating slightly superior results (mean Dice: 0.84±0.09, mean ASD: 3.14±1.79 mm). The proposed methodology produced auto-contours in under 30 seconds, with 75% of organs showing less than 1% difference in D2% and D50% dose criteria compared to ground&#xD;truth.&#xD;Significance: The integration of the nnU-Net framework with our proposed novel augmentation technique effectively addresses the challenges of limited datasets and stringent time constraints in online adaptive radiotherapy for pancreatic cancer. Our approach offers a promising solution for streamlining online adaptive workflows and represents a substantial step forward in the practical application of auto-segmentation techniques in clinical radiotherapy settings.

Practical Implementation of Federated Learning for Detecting Backdoor Attacks in a Next-word Prediction Model

This article details the development of a next-word prediction model utilizing federated learning and introduces a mechanism for detecting backdoor attacks. Federated learning enables multiple devices to collaboratively train a shared model while retaining data locally. However, this decentralized approach is susceptible to manipulation by malicious actors who control a subset of participating devices, thereby biasing the model’s outputs on specific topics, such as a presidential election. The proposed detection mechanism aims to identify and exclude devices with anomalous datasets from the training process, thereby mitigating the influence of such attacks. By using the example of a presidential election, the study demonstrates a positive correlation between the proportion of compromised devices and the degree of bias in the model’s outputs. The findings indicate that the detection mechanism effectively reduces the impact of backdoor attacks, particularly when the number of compromised devices is relatively low. This research contributes to enhancing the robustness of federated learning systems against malicious manipulation, ensuring more reliable and unbiased model performance.