Well-performing Models Research Articles

Low concentration cell painting images enable the identification of highly potent compounds

Image-based models that use features extracted from cell microscopy images can estimate the activity of small molecules in various biological assays. Typically, models are trained on images stained by an optimized protocol (e.g. Cell Painting) after exposure to a fairly high small molecule concentration (referred to as ’image concentration’) of 10μM\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$10\\; \\upmu {\ ext{M}}$$\\end{document} or higher. Low concentration images (e.g. 0.16\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$0.16$$\\end{document} μM, 0.8\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$0.8$$\\end{document} μM, 4\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$4$$\\end{document} μM) tend to yield models with worse performance. In this work, we nevertheless report a practical use for low image concentration data. We propose the combination of well-performing models trained at higher image concentrations, with lower image concentration for inference to identify more potent compounds. We show that this approach improves on the conventional method (directly training a high-potency model) in 65%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document} of assays investigated in terms of AUC-ROC, and 75%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document} of assays in terms of RIPtoP-corrected AUC-PR.

Quality evaluation parameter and classification model for effluents of wastewater treatment plant based on machine learning

With the growing consensus of emerging pollutants and biological toxicity risks in wastewater treatment plant (WWTP) effluents, traditional water quality management based on general chemical parameters no longer meets the new challenges. Here, a first-hand dataset containing 9 conventional parameters, 22 mental and inorganic ions, 25 biotoxicity parameters, and 54 emerging pollutants from effluents of 176 municipal WWTPs across China were measured. Four clustering algorithms and five classification algorithms were applied to 65 well-performing models to determine a novel evaluation parameter system. A total of 14 parameters were selected by semi-supervised machine learning, including TN, TP, NH4+-N, NO2--N, Se, SO42-, Caenorhabditis elegans body width, 72 hpf zebrafish embryo hatching rate, tetracycline, acetaminophen, gemfibrozil, PFBA, PFHxA, and HFPO-DA. These parameters were then used to construct a Healthy Effluent Quality Index model (HEQi). The application efficiency of HEQi was compared with other common methods such as the Water Quality Index (WQI), Fuzzy Synthesized Evaluation (FSE), and Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) in classifying 176 effluents. Results implicated that under the new evaluation criteria, the major task in North and Northeast China remains to reduce the conventional parameters, especially NO2--N. However, it is necessary to strengthen the removal of biotoxicity and emerging pollutants in parts of Central and Eastern China. This study offers new methodological tools and scientific insights for improving water quality assessment and safe discharge of wastewater.

Optimization of sampling design for soil total organic carbon assessment in the precision agriculture framework: Impact of different variogram models and potentiality of ground penetrating radar (GPR) covariate information

Assessing soil organic carbon (SOC) at the field scale is crucial for efficient environmental and agronomic management, especially within the precision agriculture framework. This enables the implementation of crop and soil management strategies to enhance soil quality, increase carbon sequestration, and improve crop yields. However, the process of sampling and assessing SOC is resource-intensive, demanding both in time and labour. This aspect is particularly relevant in agronomic research, where the need to assess the spatial correlation structure of the investigated variables requires the collection of large datasets with georeferenced data and often measurements need to be repeated over time as in the case long-term field experiments − LTE. Methods for minimizing information loss while reducing the sampling scheme, such as spatial simulated annealing (SSA), are particularly valuable in the scope of SOC assessment, especially when faced with budget and time/labour constraints. Within the structure of the SSA method, two critical components can be identified: i) the inclusion of highly informative covariates for the primary variable (SOC); ii) the selection of the most appropriate variogram model for spatial variability assessment. Covariates strongly correlated with SOC, such as those obtained from ground penetrating radar (GPR) that can be collected at a higher spatial density compared to SOC data, along with a well-performing model, can significantly enhance the efficiency of the sampling scheme reduction process. We conducted a study using data from an agronomic field experiment, which included 71 georeferenced sampling locations and, through an iterative downsizing process utilizing SSA, we progressively reduced the number of sampling points by removing 10, 15, and 20 observations. The sampling scheme was refined according to two distinct variogram models: the spherical and Gaussian-Matérn models, both for the primary variable (SOC) and the covariate GPR variable. This process allowed us to identify the optimal variogram model, which has a key role in maximizing the reduction of redundant points while preserving those with valuable information, and to assess the role of the covariate variable both in improving the optimization of the sampling scheme for SOC and in replacing the primary variable to optimize the sampling scheme. Finally, to assess the impact of sampling scheme reduction, a validation process was performed by estimating the dropped points by means of the remaining points using ordinary kriging and regression kriging and analysing the accuracy of such estimation. Our analysis demonstrated that it was possible to reduce the original sampling scheme by approximately 20%, equivalent to eliminating 15 sampling points out of 71, without compromising its predictive capability.

Proteomic Assessment of the Risk of Secondary Cardiovascular Events among Individuals with CKD.

Cardiovascular risk models have been developed primarily for incident events. Well-performing models are lacking to predict secondary cardiovascular events among people with a history of coronary heart disease, stroke, or heart failure who also have chronic kidney disease (CKD). We sought to develop a proteomics-based risk score for cardiovascular events in individuals with CKD and a history of cardiovascular disease. We measured 4638 plasma proteins among 1067 participants from the Chronic Renal Insufficiency Cohort (CRIC) and 536 individuals from the Atherosclerosis Risk in Communities Cohort (ARIC). All had non-dialysis-dependent CKD and coronary heart disease, heart failure, or stroke at study baseline. A proteomic risk model for secondary cardiovascular events was derived by elastic net regression in CRIC, validated in ARIC, and compared to clinical models. Biologic mechanisms of secondary events were characterized through proteomic pathway analysis. A 16-protein risk model was superior to the Framingham risk score for secondary events, including a modified score that included estimated glomerular filtration rate (eGFR). In CRIC, the annualized area under the receiver operating characteristic (AUC) within 1 to 5 years ranged between 0.77 and 0.80 for the protein model and 0.57 and 0.72 for the clinical models. These findings were replicated in the ARIC validation cohort. Biologic pathway analysis identified pathways and proteins for cardiac remodeling and fibrosis, vascular disease, and thrombosis. The proteomic risk model for secondary cardiovascular events outperformed clinical models based on traditional risk factors and eGFR.

Neural architecture search for image super-resolution: A review on the emerging state-of-the-art

Nowadays, complex and expensive neural architectures are seen by many as a way to improve the performance of existing models in image recognition, voice recognition, translation, and other tasks. Such a perspective caused an increased interest in expert architecture engineering within Deep Learning. Fueled by this interest, neural architecture search originated as a promising way to automate the tedious process of constructing a deep neural network by hand. Over the last five years, we have seen an increasing number of works focusing all efforts on studying the impact of automating deep neural network design. The spotlight has recently turned from automatically discovering classification models to other more complex tasks. Motivated by a desire for high-resolution images in real-world user-centered and expert computer vision applications, architecture search for super-resolution image restoration centers in approaches capable of automatically finding efficient and well-performing models. Here, we present a survey that, beyond delving into an overview of modern approaches to automatic neural network design, focuses on the recollection and study of neural architecture search approaches that have directed their efforts at the super-resolution image restoration tasks and future lines of research found within this emerging area of study.

Disagreement serves an explainable ensemble model based on Dempster–Shafer evidence-fusion for an improved skin lesion classification

The ever-growing development of deep learning models for skin lesion classification yields consistently to exceptional performance. These models revealed remarkable accuracy and robustness in the diagnosis of various skin lesion classes. Although this advancement enables significant aids in early detection and diagnosis, selecting the most suitable model for a task is challenging especially when most models demonstrate high proficiency. Consequently, small performance variations among well-performing models become critical factors. These variations present an additional challenge especially when considering the black-box nature of the model, which may result in a lack of evidence and certainty when making critical decisions. In this study, we tackle the problem of selecting the most appropriate models for skin lesion classification among various well-performing, yet unexplainable deep learning models. In fact, by proposing an ensemble approach that leverages Dempster–Shafer theory and feature importance visualization, we aim to create a trustworthy and interpretable model for improved skin lesion classification. The proposed Dempster–Shafer evidence-fusion ensemble model, provides insights into evidence-based integration of multiple models allowing for enhanced trustworthiness in the decision-making process. This is accomplished through comparing the high performances between different deep learning models. Hence, metrics are extracted by performing statistical tests and explainable methods besides numerical assessment, from the failure cases to profile the disagreement from different perspectives. With these metrics, the classifiers’ complementary behaviors is then identified, accordingly making contribution in proposing an explainable ensemble model and leveraging the complementary strengths of the individual models. The proposed ensemble model has been exhaustively evaluated using the challenging HAM10000 dataset. The results obtained indicate that the proposed skin lesion classifier can provide useful decision support while significantly improving classification performance by 6% in terms of accuracy compared with the best-performing individual model, and achieving 92% accuracy.

A realistic model extraction attack against graph neural networks

Model extraction attacks are considered to be a significant avenue of vulnerability in machine learning. In model extraction attacks, the attacker repeatedly queries a victim model so as to train a surrogate model that can mimic the output of the victim model. Graph neural networks (GNNs), which are designed to process graph data, were previously thought to be less sensitive to such attacks. This is because, in black-box settings, attackers only have limited access to the victim model. Also, the number of queries any one user can make within a given time window is usually restricted and within this finite number of responses some of the information may contain errors. Yet training a useful surrogate model not only requires a substantial number of queries, but incorrect node labels appearing in the victim GNN’s responses is highly problematic. However, in this paper, we demonstrate that GNNs may have a similar vulnerability to model extraction attacks as a normal machine learning model. Our proposed method of extraction addresses the issue of incorrect node labels while also significantly reducing the number of required queries required to train a well-performing model. With this method, GNN extraction attacks are actually highly practical in the real world. Specifically, our proposed methodology incorporates an edge prediction module that introduces potential edges into the original graph data. This module links incorrectly labeled nodes with more accurately labeled ones, thereby mitigating the impact of incorrect labels. And, by increasing the number of possible edges, our approach enables the surrogate model to better leverage the graph’s structure, enhancing the contribution of the labeled nodes and allowing the model extraction attack to be executed with fewer queries. Our experiments demonstrate a significant performance improvement over existing approaches, especially in a black-box setting. As such, this research shows that GNNs are also vulnerable to model extraction attacks in real-world scenarios.

Electricity Price Forecasting in the Irish Balancing Market

Short-term electricity markets are becoming more relevant due to less-predictable renewable energy sources, attracting considerable attention from the industry. The balancing market is the closest to real-time and the most volatile among them. Its price forecasting literature is limited, inconsistent and outdated, with few deep learning attempts and no public dataset. This work applies to the Irish balancing market a variety of price prediction techniques proven successful in the widely studied day-ahead market. We compare statistical, machine learning, and deep learning models using a framework that investigates the impact of different training sizes. The framework defines hyperparameters and calibration settings; the dataset and models are made public to ensure reproducibility and to be used as benchmarks for future works. An extensive numerical study shows that well-performing models in the day-ahead market do not perform well in the balancing one, highlighting that these markets are fundamentally different constructs. The best model is LEAR, a statistical approach based on LASSO, achieving a mean absolute error of 32.82 €/MWh, surpassing more complex and computationally demanding approaches with errors ranging from 33.71 €/MWh to 44.55 €/MWh.

#1085 Prediction models for hospitalization in patients with advanced chronic kidney disease: systematic review, external validation and development study

Abstract Background and Aims Hospitalization is a common complication in patients with advanced chronic kidney disease (CKD). Predicting hospitalization would be beneficial for clinicians and patients. Clinicians may be assisted in adequate healthcare planning by being able to anticipate hospitalizations. Additionally, patients have indicated wanting more information about their risk of hospitalization. By providing patients a personal indication of their risk of hospitalization, they better know what to expect of the future which can foster a feeling of control. Especially for incident dialysis patients this might help, as they enter a new and to them still largely unknown stage of their disease. Moreover, if a patient knows they have a high risk of hospitalization, they may better cope with the hospitalization if it occurs, as it is less unexpected. Therefore, we aimed to 1) identify and give an overview of current prediction models for hospitalization in CKD patients, 2) externally validate these models, and 3) develop a new model if no well-performing models were identified. Method We performed a systematic review in PubMed to identify prediction models for hospitalization, length of stay (LOS), and readmission. Identified models were appraised using the Prediction model Risk of Bias Assessment Tool. We targeted three external validations in hemodialysis (HD), peritoneal dialysis (PD), and pre-dialysis patients in a Dutch dialysis and European pre-dialysis cohort. Prediction models were externally validated on discrimination (C-statistic) and calibration (calibration plot, slope, and calibration-in-the-large). We developed a Fine-Gray model for hospitalization within one year in patients initiating HD, taking into account the competing risk of death. Informed by model performance, we also investigated Cox and logistic regression as modeling strategies. We performed a sensitivity analysis, redefining the outcome to be predicted as the 1-year risk of hospitalization of 3 days or longer. Results From the systematic search, we identified 17 models in 6 studies. All models were developed in the USA or in China. Two models predicted hospitalization rate in incident PD patients and 3 models predicted 30-day readmission in maintenance HD patients and in CKD stage 3-5 patients not on dialysis discharged after a heart failure hospitalization. The risk of 1 and 6 hospitalizations within one year in prevalent HD patients were predicted by one model each. Prolonged LOS in PD patients was predicted by ten models. All models were of low quality with a high risk of bias. The most common shortcomings were using future information to make predictions, inadequate handling of missing data, and model overfitting. Due to underreporting and unavailability of predictors, we could only externally validate 3 models. These models were poorly calibrated and had poor discrimination. We developed a new model in 1,020 incident HD patients predicting 1-year risk of hospitalization. We used age, cardiovascular disease, diabetes mellitus, malignancy, estimated glomerular filtration rate, and body mass index as predictors, selected based on literature and clinical expertise. The model discriminated poorly (C-statistic, 95% confidence interval [CI]; 0.55, 0.53-0.58), also in targeted external validations in 620 PD patients (0.55, 0.52-0.58) and 1,718 pre-dialysis patients (0.52, 0.52-0.54). Calibration was poor for all models. Using different modelling strategies, model performance did not improve. In the sensitivity analysis, discrimination improved but was still low (HD population: 0.58, 0.55-0.62; PD population: 0.59, 0.54-0.63; pre-dialysis: 0.55, 0.52-0.57). Conclusion Hospitalizations are important but difficult to predict for patients with advanced CKD. Current models are insufficient to predict hospitalizations. We offer two main suggestions to improve future predictions of hospitalizations. First, the outcome of hospitalization should be more specific (e.g., cardiovascular hospitalizations) to reduce outcome heterogeneity. Second, the selection of predictors should be extended with more patient-centered factors, such as patient-reported outcomes.

Image Classification of Room Tidiness Using VGGNet with Data Augmentation

Tidiness becomes an essential aspect that everyone should maintain. Tidiness encompasses various elements, and one of the aspects closely related to it is the tidiness of a room. The tidiness of a room creates a comfortable and clean environment. The tidiness of a room is particularly crucial for individuals involved in businesses such as the hospitality industry. Therefore, a solution is needed to address this issue, and one of the approaches is to utilize Deep Learning for automatic room tidiness classification. One popular deep learning method for implementing image classification of room tidiness is the convolutional neural network (CNN), which creates a well-performing model for image classification with data augmentation. This research aims to develop an image classification model using CNN with the VGGNet architecture and data augmentation. This study is a reference for further development, with potential applications in the hospitality industry. The research results in a model that achieves an accuracy of 98.44% with a data proportion of 90% for training and validation, while the remaining 10% is used for testing purposes. The conclusion drawn from this study is that the CNN method, combined with data augmentation, can be utilized for image classification of room tidiness.

Development and Internal Validation of Nomograms for Survival of Advanced Epithelial Ovarian Cancer Based on Established Prognostic Factors and Hematologic Parameters.

Objective: To assess the association between pretreatment thrombocytosis, anemia, and leukocytosis and overall survival (OS) of advanced-stage EOC. Furthermore, to develop nomograms using established prognostic factors and pretreatment hematologic parameters to predict the OS of advanced EOC patients. Methods: Advanced-stage EOC patients treated between January 1996 and January 2010 in eastern Netherlands were included. Survival outcomes were compared between patients with and without pretreatment thrombocytosis (≥450,000 platelets/µL), anemia (hemoglobin level of <7.5 mmol/L), or leukocytosis (≥11.0 × 109 leukocytes/L). Three nomograms (for ≤3-, ≥5-, and ≥10-year OS) were developed. Candidate predictors were fitted into multivariable logistic regression models. Multiple imputation was conducted. Model performance was assessed on calibration, discrimination, and Brier scores. Bootstrap validation was used to correct for model optimism. Results: A total of 773 advanced-stage (i.e., FIGO stages IIB-IV) EOC patients were included. The median [interquartile range, IQR] OS was 2.3 [1.3-4.2] and 3.0 [1.4-7.0] years for patients with and without pretreatment thrombocytosis (p < 0.01). The median OS was not notably different for patients with and without pretreatment leukocytosis (p = 0.58) or patients with and without pretreatment anemia (p = 0.07). The final nomograms comprised established predictors with either pretreatment leukocyte or platelet count. The ≥5- and ≥10-year OS models demonstrated good calibration and adequate discrimination with optimism-corrected c-indices [95%-CI] of 0.76 [0.72-0.80] and 0.78 [0.73-0.83], respectively. The ≤3-year OS model demonstrated suboptimal performance with an optimism-corrected c-index of 0.71 [0.66-0.75]. Conclusions: Pretreatment thrombocytosis is associated with poorer EOC survival. Two well-performing models predictive of ≥5-year and ≥10-year OS in advanced-stage EOC were developed and internally validated.

Application of 18F-FDG PET/CT imaging radiomics in the differential diagnosis of single-nodule pulmonary metastases and second primary lung cancer in patients with colorectal cancer.

It is crucially essential to differentially diagnose single-nodule pulmonary metastases (SNPMs) and second primary lung cancer (SPLC) in patients with colorectal cancer (CRC), which has important clinical implications for treatment strategies. In this study, we aimed to establish a feasible differential diagnosis model by combining 18F-fluorodeoxyglucose positron-emission tomography (18F-FDG PET) radiomics, computed tomography (CT) radiomics, and clinical features. CRC patients with SNPM or SPLC who underwent 18F-FDG PET/CT from January 2013 to July 2022 were enrolled in this retrospective study. The radiomic features were extracted by manually outlining the lesions on PET/CT images, and the radiomic modeling was realized by various screening methods and classifiers. In addition, clinical features were analyzed by univariate analysis and logistic regression (LR) analysis to be included in the combined model. Finally, the diagnostic performances of these models were illustrated by the receiver operating characteristic (ROC) curves and the area under the curve (AUC). We studied data from 61 patients, including 36 SNPMs and 25 SPLCs, with an average age of 65.56 ± 10.355 years. Spicule sign and ground-glass opacity (GGO) were significant independent predictors of clinical features (P = 0.012 and P < 0.001, respectively) to build the clinical model. We achieved a PET radiomic model (AUC = 0.789), a CT radiomic model (AUC = 0.818), and a PET/CT radiomic model (AUC = 0.900). The PET/CT radiomic models were combined with the clinical model, and a well-performing model was established by LR analysis (AUC = 0.940). For CRC patients, the radiomic models we developed had good performance for the differential diagnosis of SNPM and SPLC. The combination of radiomic and clinical features had better diagnostic value than a single model.

Development of risk models of incident hypertension using machine learning on the HUNT study data

In this study, we aimed to create an 11-year hypertension risk prediction model using data from the Trøndelag Health (HUNT) Study in Norway, involving 17 852 individuals (20–85 years; 38% male; 24% incidence rate) with blood pressure (BP) below the hypertension threshold at baseline (1995–1997). We assessed 18 clinical, behavioral, and socioeconomic features, employing machine learning models such as eXtreme Gradient Boosting (XGBoost), Elastic regression, K-Nearest Neighbor, Support Vector Machines (SVM) and Random Forest. For comparison, we used logistic regression and a decision rule as reference models and validated six external models, with focus on the Framingham risk model. The top-performing models consistently included XGBoost, Elastic regression and SVM. These models efficiently identified hypertension risk, even among individuals with optimal baseline BP (< 120/80 mmHg), although improvement over reference models was modest. The recalibrated Framingham risk model outperformed the reference models, approaching the best-performing ML models. Important features included age, systolic and diastolic BP, body mass index, height, and family history of hypertension. In conclusion, our study demonstrated that linear effects sufficed for a well-performing model. The best models efficiently predicted hypertension risk, even among those with optimal or normal baseline BP, using few features. The recalibrated Framingham risk model proved effective in our cohort.

ADGRE5-centered Tsurv model in T cells recognizes responders to neoadjuvant cancer immunotherapy.

Neoadjuvant immunotherapy with anti-programmed death-1 (neo-antiPD1) has revolutionized perioperative methods for improvement of overall survival (OS), while approaches for major pathologic response patients' (MPR) recognition along with methods for overcoming non-MPR resistance are still in urgent need. We utilized and integrated publicly-available immune checkpoint inhibitors regimens (ICIs) single-cell (sc) data as the discovery datasets, and innovatively developed a cell-communication analysis pipeline, along with a VIPER-based-SCENIC process, to thoroughly dissect MPR-responding subsets. Besides, we further employed our own non-small cell lung cancer (NSCLC) ICIs cohort's sc data for validation in-silico. Afterward, we resorted to ICIs-resistant murine models developed by us with multimodal investigation, including bulk-RNA-sequencing, Chip-sequencing and high-dimensional cytometry by time of flight (CYTOF) to consolidate our findings in-vivo. To comprehensively explore mechanisms, we adopted 3D ex-vivo hydrogel models for analysis. Furthermore, we constructed an ADGRE5-centered Tsurv model from our discovery dataset by machine learning (ML) algorithms for a wide range of tumor types (NSCLC, melanoma, urothelial cancer, etc.) and verified it in peripheral blood mononuclear cells (PBMCs) sc datasets. Through a meta-analysis of multimodal sequential sc sequencing data from pre-ICIs and post-ICIs, we identified an MPR-expanding T cells meta-cluster (MPR-E) in the tumor microenvironment (TME), characterized by a stem-like CD8+ T cluster (survT) with STAT5-ADGRE5 axis enhancement compared to non-MPR or pre-ICIs TME. Through multi-omics analysis of murine TME, we further confirmed the existence of survT with silenced function and immune checkpoints (ICs) in MPR-E. After verification of the STAT5-ADGRE5 axis of survT in independent ICIs cohorts, an ADGRE5-centered Tsurv model was then developed through ML for identification of MPR patients pre-ICIs and post-ICIs, both in TME and PBMCs, which was further verified in pan-cancer immunotherapy cohorts. Mechanistically, we unveiled ICIs stimulated ADGRE5 upregulation in a STAT5-IL32 dependent manner in a 3D ex-vivo system (3D-HYGTIC) developed by us previously, which marked Tsurv with better survival flexibility, enhanced stemness and potential cytotoxicity within TME. Our research provides insights into mechanisms underlying MPR in neo-antiPD1 and a well-performed model for the identification of non-MPR.

Projecting Wintertime Newly Formed Arctic Sea Ice through Weighting CMIP6 Model Performance and Independence

Precipitous Arctic sea-ice decline and the corresponding increase in Arctic open-water areas in summer months give more space for sea-ice growth in the subsequent cold seasons. Compared to the decline of the entire Arctic multiyear sea ice, changes in newly formed sea ice indicate more thermodynamic and dynamic information on Arctic atmosphere–ocean–ice interaction and northern mid–high latitude atmospheric teleconnections. Here, we use a large multimodel ensemble from phase 6 of the Coupled Model Intercomparison Project (CMIP6) to investigate future changes in wintertime newly formed Arctic sea ice. The commonly used model-democracy approach that gives equal weight to each model essentially assumes that all models are independent and equally plausible, which contradicts with the fact that there are large interdependencies in the ensemble and discrepancies in models’ performances in reproducing observations. Therefore, instead of using the arithmetic mean of well-performing models or all available models for projections like in previous studies, we employ a newly developed model weighting scheme that weights all models in the ensemble with consideration of their performance and independence to provide more reliable projections. Model democracy leads to evident bias and large intermodel spread in CMIP6 projections of newly formed Arctic sea ice. However, we show that both the bias and the intermodel spread can be effectively reduced by the weighting scheme. Projections from the weighted models indicate that wintertime newly formed Arctic sea ice is likely to increase dramatically until the middle of this century regardless of the emissions scenario. Thereafter, it may decrease (or remain stable) if the Arctic warming crosses a threshold (or is extensively constrained).

Optimal risk and diagnosis assessment strategies in perinatal depression: A machine learning approach from the life-ON study cohort

This study aimed to assess the concordance of various psychometric scales in detecting Perinatal Depression (PND) risk and diagnosis. A cohort of 432 women was assessed at 10–15th and 23–25th gestational weeks, 33–40 days and 180–195 days after delivery using the Edinburgh Postnatal Depression Scale (EPDS), Visual Analogue Scale (VAS), Hamilton Depression Rating Scale (HDRS), Montgomery-Åsberg Depression Rating Scale (MADRS), and Mini International Neuropsychiatric Interview (MINI). Spearman's rank correlation coefficient was used to assess agreement across instruments, and multivariable classification models were developed to predict the values of a binary scale using the other scales. Moderate agreement was shown between the EPDS and VAS and between the HDRS and MADRS throughout the perinatal period. However, agreement between the EPDS and HDRS decreased postpartum. A well-performing model for the estimation of current depression risk (EPDS > 9) was obtained with the VAS and MADRS, and a less robust one for the estimation of current major depressive episode (MDE) diagnosis (MINI) with the VAS and HDRS. When the EPDS is not feasible, the VAS may be used for rapid and comprehensive postpartum screening with reliability. However, a thorough structured interview or clinical examination remains necessary to diagnose a MDE.

Evaluation of CMIP6 model performance and extreme precipitation prediction in the Awash basin

Extreme rainfall and its accompanying hydrological extremes are happening more frequently as a result of global warming's alteration of regional and local weather patterns. This poses a serious risk to ecosystem, environment and the community livelihoods. The Awash basin in Ethiopia is especially vulnerable to these events, posing significant threats to the region. There are, however, limited information's available that could be used to characterize the condition of extreme precipitation in the basin. Therefore, this study aims to evaluate the performance of CMIP6 models in simulating extreme precipitation in the Awash basin. Additionally, the study calculated extreme precipitation using best-fit probability distribution functions (PDFs) for the period from 1985 to 2014. The Climate Hazards Group Infrared Precipitation with station data (CHIRPS) were used to evaluate the global climate models. Simulated data were interpolated using bilinear techniques. Four statistical indices (percentage of bias, root mean square error, mean absolute error, and Pearson correlation) assessed GCM performance in simulating precipitation extremes. Graphical approaches, numerical methods, and empirical distribution functions were employed to evaluate the performance of various probability distribution functions (PDFs). The study identified MIROC6, CESM2-WACCM, and Ensemble as well-performing models with PBIAS and RMSE of 6.6 %, −10.2 %, −17.2 %, and 11.5, 10, 9.7 respectively, while MPI-ESM1-2-HR and EC-Earth3 struggled with extreme rainfall simulation. The generalized extreme values distribution was found to be a good fit for extreme rainfall estimation. GFDL-ESM4 and BCC-CSM2-MR models estimated the highest extreme rainfall of 90 mm/day and 80 mm/day, respectively, however these models overestimated the return period. Conversely, MRI-ESM2-0, NorESM2-MM, ACCESS ESM1-5, and CMCC-ESM2 models underestimated the return periods. Spatially, GFDL-ESM4 and ACCESS-ESM1-5 models exhibited uniform peak rainfall values over a large area. Overall, the study suggests that employing the generalized extreme value distribution could effectively inform risk assessment and management of extreme events in the Awash basin.

Validation and Selection of a Representative Subset from the Ensemble of EURO-CORDEX EUR11 Regional Climate Model Outputs for the Czech Republic

To better understand the impact of climate change at a given location, it is crucial to consider a wide range of climate models that are representative of the area. In this study, we emphasize the importance of the careful validation and selection of climate models most suitable for a particular region. This step is critical to enhance the relevance of climate change impact studies and consequently design appropriate and robust adaptation measures, particularly in agriculture, forestry and water resources management. We propose validation and selection methods for regional climate models that can help identify a smaller group of well-performing models using the Central European area and Czech Republic as examples. In the validation process, 7 out of 19 regional climate models performed poorly. Of the 12 well-performing models, a subset of 7 models was selected to represent the uncertainty in the entire ensemble, which could be used in subsequent studies. The methodology is sufficiently general and may be applied to other climate model ensembles.

393. A RADIOMICS STRATEGY BASED ON CT INTRA-TUMORAL AND PERITUMORAL REGIONS FOR PREOPERATIVE PREDICTION OF NEOADJUVANT CHEMORADIOTHERAPY FOR ESOPHAGEAL CANCER

Abstract Background The standard treatment for esophageal cancer patients is neoadjuvant chemoradiotherapy followed by surgery. However, some of these patients do not achieve pathological complete response with this therapy, resulting in poor outcomes. The objective of this study is to develop a method for selecting patients who can achieve pathological complete response through pre-neoadjuvant therapy chest-enhanced CT scans. Methods The study enrolled 201 patients with esophageal cancer and divided them into a training set and a testing set in a 7:3 ratio. Radiomics features of intra-tumoral and peritumoral images were extracted from preoperative chest-enhanced CT scans of these patients. The features underwent dimensionality reduction in two steps, using Student’s t-test and least absolute shrinkage and selection operator. The selected intra-tumoral and peritumoral (including marginal and adjacent ROI) features were used to build models with four machine learning classifiers. The models with satisfactory accuracy and stability levels were considered to perform well. Finally, the performance of these well-performing models on the testing set was displayed using ROC curves. Results Among the 16 models, the best-performing models were the integrated (intra-tumoral and peritumoral features) -XGBoost and integrated-random forest models. In the training set, the two models had average ROC AUCs of 0.906 and 0.918 respectively, with relative standard deviations (RSDs) of 6.26 and 6.89. In the testing set, the AUCs were 0.845 and 0.871, respectively. There was no significant difference in the ROC curves between the two models. Conclusion The addition of peritumoral radiomics features to the radiomics analysis may improve the predictive performance of pathological response for esophageal cancer patients to neoadjuvant chemoradiotherapy. The integrated (intra-tumoral and peritumoral features) -XGBoost and integrated-random forest models developed in this study show potential for predicting pathological complete response in esophageal cancer patients and may help in selecting patients for neoadjuvant therapy.

Domain adversarial graph neural network with cross-city graph structure learning for traffic prediction

Deep learning models have emerged as a promising way for traffic prediction. However, the requirement for large amounts of training data remains a significant issue for achieving well-performing models. Data scarcity in real-world scenarios, caused by costly collection or privacy policies, can severely impede the performance of existing deep learning models. Transfer learning aims to leverage knowledge learned from data-sufficient cities to improve prediction performance in data-scarce cities. Unfortunately, most existing methods solely focus on transferring knowledge at the city level, neglecting fine-grained node-level correlations and distribution discrepancies between cities. In this paper, we propose DAGN, a domain adversarial graph neural network that mines inter-city spatial–temporal correlations and alleviates domain distribution discrepancies to address the data scarcity problem in traffic prediction. Specifically, DAGN comprises three key modules: (1) A cross-city graph structure learning module is developed to capture node-pair adjacent relationships across cities, enabling the dynamic aggregation of inter-city spatial–temporal information. Additionally, a graph reconstruction loss is proposed to enforce structural consistency between the learned and priori graphs. (2) A domain adversarial strategy is integrated with a spatial–temporal module, which jointly extracts domain-invariant spatial and temporal features to reduce the distribution discrepancies between cities. (3) To adaptively extract transferable knowledge from a global perspective, a global spatial–temporal attention module is designed. Extensive experiments on six traffic flow and traffic speed prediction benchmarks demonstrate that DAGN consistently outperforms state-of-the-art methods.