Performance Of Individual Models Research Articles

How to enhance the applicability of a risk prediction model for term small-for-gestational-age neonates in clinical settings?

To construct a simple term small-for-gestational-age (SGA) neonate prediction model that is clinically practical. This analysis was based on the Born in Guangzhou Cohort Study (BIGCS). Mothers who had a singleton pregnancy, delivered a term neonate, and had an ultrasonography within 30 + 0 to 32 + 6 weeks of gestation were included. Term SGA was defined with customized population percentiles. Prediction models were constructed with backward selection logistic regression in a four-step approach, where model 1 contained fetal biometrics only, models 2 and 3 included maternal features and a time factor (weeks between ultrasonography and delivery), respectively; and model 4 contained all features mentioned. The prediction performance of individual models was evaluated based on area under the curve (AUC) and a calibration test was performed. The prevalence of SGA in the study population of 21 346 women was 11.5%. With a complete-case analysis approach, data of 19 954 women were used for model construction and validation. The AUC of the four models were 0.781, 0.793, 0.823, and 0.834, respectively, and all were well-calibrated. Model 3 consisted of fetal biometrics and corrected for time to delivery was chosen as the final model to build risk prediction graphs for clinical use. A prediction model derived from fetal biometrics in early third trimester is satisfactory to predict SGA.

Hybrid model of CT-fractional flow reserve, pericoronary fat attenuation index and radiomics for predicting the progression of WMH: a dual-center pilot study.

To develop and validate a hybrid model incorporating CT-fractional flow reserve (CT-FFR), pericoronary fat attenuation index (pFAI), and radiomics signatures for predicting progression of white matter hyperintensity (WMH). A total of 226 patients who received coronary computer tomography angiography (CCTA) and brain magnetic resonance imaging from two hospitals were divided into a training set (n = 116), an internal validation set (n = 30), and an external validation set (n = 80). Patients who experienced progression of WMH were identified from subsequent MRI results. We calculated CT-FFR and pFAI from CCTA images using semi-automated software, and segmented the pericoronary adipose tissue (PCAT) and myocardial ROI. A total of 1,073 features were extracted from each ROI, and were then refined by Elastic Net Regression. Firstly, different machine learning algorithms (Logistic Regression [LR], Support Vector Machine [SVM], Random Forest [RF], k-nearest neighbor [KNN] and eXtreme Gradient Gradient Boosting Machine [XGBoost]) were used to evaluate the effectiveness of radiomics signatures for predicting WMH progression. Then, the optimal machine learning algorithm was used to compare the predictive performance of individual and hybrid models based on independent risk factors of WMH progression. Receiver operating characteristic (ROC) curve analysis, calibration and decision curve analysis were used to evaluate predictive performance and clinical value of the different models. CT-FFR, pFAI, and radiomics signatures were independent predictors of WMH progression. Based on the machine learning algorithms, the PCAT signatures led to slightly better predictions than the myocardial signatures and showed the highest AUC value in the XGBoost algorithm for predicting WMH progression (AUC: 0.731 [95% CI: 0.603-0.838] vs.0.711 [95% CI: 0.584-0.822]). In addition, pFAI provided better predictions than CT-FFR (AUC: 0.762 [95% CI: 0.651-0.863] vs. 0.682 [95% CI: 0.547-0.799]). A hybrid model that combined CT-FFR, pFAI, and two radiomics signatures provided the best predictions of WMH progression [AUC: 0.893 (95%CI: 0.815-0.956)]. pFAI was more effective than CT-FFR, and PCAT signatures were more effective than myocardial signatures in predicting WMH progression. A hybrid model that combines pFAI, CT-FFR, and two radiomics signatures has potential use for identifying WMH progression.

Assessment of Arctic sea ice and surface climate conditions in nine CMIP6 climate models

ABSTRACT The observed retreat and anticipated further decline in Arctic sea ice holds strong climate, environmental, and societal implications. In predicting climate evolution, ensembles of coupled climate models have demonstrated appreciable accuracy in simulating sea-ice area trends throughout the historical period, yet individual climate models still show significant differences in accurately representing the sea-ice thickness distribution. To better understand individual model performance in sea-ice simulation, nine climate models were evaluated in comparison with Arctic satellite and reanalysis-derived sea-ice thickness data, sea-ice area records, and atmospheric reanalysis data of surface wind and air temperature. This assessment found that the simulated spatial distribution of historical sea-ice thickness varies greatly between models and that several key limitations persist among models. Primarily, most models do not capture the thickest regimes of multiyear ice present in the Wandel and Lincoln seas; those that do often possess erroneous positive bias in other regions such as the Laptev Sea or along the Eurasian Arctic Shelf. This analysis provides enhanced understanding of individual model historical simulation performance, which is critical in informing the selection of coupled climate model projections for dependent future modeling efforts.

A Deep Learning-Based Ensemble Method for Early Diagnosis of Alzheimer’s Disease using MRI Images

Recently, the early diagnosis of Alzheimer’s disease has gained major attention due to the growing prevalence of the disease and the resulting costs imposed on individuals and society. The main objective of this study was to propose an ensemble method based on deep learning for the early diagnosis of AD using MRI images. The methodology of this study consisted of collecting the dataset, preprocessing, creating the individual and ensemble models, evaluating the models based on ADNI data, and validating the trained model based on the local dataset. The proposed method was an ensemble approach selected through a comparative analysis of various ensemble scenarios. Finally, the six best individual CNN-based classifiers were selected to combine and constitute the ensemble model. The evaluation showed an accuracy rate of 98.57, 96.37, 94.22, 99.83, 93.88, and 93.92 for NC/AD, NC/EMCI, EMCI/LMCI, LMCI/AD, four-way and three-way classification groups, respectively. The validation results on the local dataset revealed an accuracy of 88.46 for three-way classification. Our performance results were higher than most reviewed studies and comparable with others. Although comparative analysis showed superior results of ensemble methods against individual architectures, there were no significant differences among various ensemble approaches. The validation results revealed the low performance of individual models in practice. In contrast, the ensemble method showed promising results. However, further studies on various and larger datasets are required to validate the generalizability of the model.

Explainable End-to-End Supervised Learning Identifies Myelodysplastic Neoplasms in Bone Marrow Smears

Myelodysplastic neoplasms (MDS) encompass clonal myeloid malignancies that are characterized by ineffective hematopoiesis, cytopenia, myelodysplasia, and recurrent genetic events. Apart from genetic abnormalities and the assessment of peripheral blood for cytopenia, cytomorphologic evaluation of bone marrow morphology is key in initial diagnosis and response assessment as well as to detect disease transformation to acute myeloid leukemia (AML). However, the identification of aberrant morphologies in bone marrow cells is challenging and prone to inter-observer variability even for seasoned morphologists resulting in discrepancies between local site and central review. After the introduction of deep learning (DL) models for computer vision tasks, a wide variety of applications in imaging diagnostics has been addressed using convolutional neural networks including the assessment of bone marrow cell types and prediction of hematological malignancies. Bone marrow smear (BMS) images of manually selected regions of interest (ROI) from 483 MDS patients treated at the University Hospital Dresden were captured at 40-fold magnification with a resolution of 2560*1920 pixels and pre-processed. In parallel, BMS images from two control cohorts were taken: (i) 1226 AML patients registered in the German Study Alliance Leukemia (SAL) registry, and (ii) 236 healthy bone marrow donors. In addition, an external validation cohort of 50 MDS patients was obtained from the Munich Leukemia Laboratory (MLL). Image-level labels were provided denoting the corresponding diagnosis: “MDS”, “AML” and “healthy donor”. To balance different sample sizes between the datasets, simple image augmentation techniques such as random sized cropping, color shifting, and linear transformations were used. Training and testing were performed with 5-fold cross-validation, i.e. multiple train-test-splits of 80:20. Importantly, cell-level labels, e.g. specifically annotated cell types or features like distinct dysplastic morphologies were not used to train our model as we intended to abstract visual information exclusively through our neural networks instead of using expert-guided feature engineering. Six different recently introduced convolutional neural network (CNN) architectures were compared (ResNet-34/50/101/152, ResNeXt-50/101, ShuffleNet V2, DenseNet-121/169/201, Wide ResNet-50/101, and SqueezeNet 1.1). To evaluate individual model performance, we used accuracy, recall (=sensitivity), precision (=positive predictive value), and area-under-the-curve (AUC) of both the receiver-operating-characteristic (ROC) and precision-recall-curve (PRC). Binary classification tasks showed high accuracy of 0.956 and 0.981 for distinguishing MDS from healthy donors and AML, respectively (Table I). Results were validated in the external MDS cohort reaching an accuracy of 0.916 for MDS vs. donors and 0.968 for MDS vs. AML (Table I). For explainability, we used occlusion sensitivity mapping which iteratively blocks image areas from being evaluated by the CNNs, measures consecutive drops in prediction performance, and thus identifies image areas of high importance (highlighted in green). By reviewing images, we found the CNNs to target nuclei not only of dysplastic cells but rather of all available cells in a ROI (Figure 1). Using supervised end-to-end deep learning, we developed a software framework to distinguish between MDS, AML, and healthy controls with very high accuracy. Importantly, models do not require laborious feature engineering, i.e. manually drafted, cell-level labels alleviating the need for time-consuming and cost-ineffective manual labeling and thereby bypassing a key bottleneck for current standards of computer vision in microscopy.

Development of a PAT platform for the prediction of granule tableting properties

In this work, the feasibility of implementing a process analytical technology (PAT) platform consisting of Near Infrared Spectroscopy (NIR) and particle size distribution (PSD) analysis was evaluated for the prediction of granule downstream processability. A Design of Experiments-based calibration set was prepared using a fluid bed melt granulation process by varying the binder content, granulation time, and granulation temperature. The granule samples were characterized using PAT tools and a compaction simulator in the 100-500kg load range. Comparing the systematic variability in NIR and PSD data, their complementarity was demonstrated by identifying joint and unique sources of variation. These particularities of the data explained some differences in the performance of individual models. Regarding the fusion of data sources, the input data structure for partial least squares (PLS) based models did not significantly impact the predictive performance, as the root mean squared error of prediction (RMSEP) values were similar. Comparing PLS and artificial neural network (ANN) models, it was observed that the ANNs systematically provided superior model performance. For example, the best tensile strength, ejection stress, and detachment stress prediction with ANN resulted in an RMSEP of 0.119, 0.256, and 0.293 as opposed to the 0.180, 0.395, and 0.430 RMSEPs of the PLS models, respectively. Finally, the robustness of the developed models was assessed.

House Price Prediction: Comparative Analysis of Regression-Based Machine Learning Algorithms

Abstract: Advancement in technology has revolutionized the ways of doing things in contemporary time. One of such is Artificial intelligence which has given birth to myriads of methods/techniques employed in solving real life problems. Many Machine learning techniques can be used in predicting house prices with several factors in consideration. House prices rise on annual basis, triggering the need for house price prediction models. Predictive models enable families to acquire a house of their choice when accurately developed. There have been a significant number of articles that adopt traditional machine learning algorithms to successfully estimate house prices, but they rarely compare the performance of individual models. This study will extensively test and compare numerous machine learning technique and present an optimistic model that will be used in developing a house pricing prediction system. Several models were developed and compared using Linear Regression (LR), Least Absolute Shrinkage & Selection Operator Regression (LASSO-R), Ridge Regression (RR), K-Nearest Neighbours Regression (KNN-R), Decision Tree Regression (DTR) and Extra Trees Regression (ETR) algorithms. Implemented using Python programming language. Amongst them ETR outperformed the others with MSE (16233.4), RMSE (128.7), MAE (49.6) and R2 (0.63) while the least performed is KNN-R with MSE (45763.3), RMSE (213.9), MAE (99.2) and R2 (-0.04).

Development of an integrated modelling framework to evaluate impacts of pressures on habitat conditions and riverine biota

AbstractIt is becoming extremely challenging to protect the structural and functional integrity of freshwater ecosystems, as many ecological systems face pressures from climate change, population growth and other environmental impacts. However, existing approaches of assessing impacts of external stressors to habitat conditions often neglect the long‐term hydrological changes. This study aimed to develop an integrated modelling framework to assess impacts of different stressors on reach‐scale habitat conditions for benthic macroinvertebrates in a temperate lowland river in central Poland. We coupled hydrological (SWAT+), hydraulic (HEC‐RAS 1D and SRH‐2D) and habitat (CASiMiR) models to predict the response of habitat suitability for the functional group of filter feeders to changes in the catchment for a 135‐m section of the Jeziorka river. Setting up and calibrating individual models required extensive field data collection (e.g., bathymetry field surveys, gauging water levels, discharge and flow velocity measurements and biological sampling). The performance of individual models in simulating discharge, water levels and depth/flow velocity was assessed as satisfactory according to literature. The modelling chain was then applied to compare the effects of three flow scenarios (average, low and high flows). The results showed an overall high value of hydraulic habitat suitability index for filter feeders, with moderately and severely lower values for low and high flow scenario, respectively. The implemented modelling framework performed well could be applied to examine the impacts of climate change and other anthropogenic stressors on reach‐scale habitat conditions. Its practical implementation could be beneficial towards sustainable water resources and environmental flow regime development.

Application of machine learning (individual vs stacking) models on MERRA-2 data to predict surface PM2.5 concentrations over India

The spatial coverage of PM2.5 monitoring is non-uniform across India due to the limited number of ground monitoring stations. Alternatively, Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2), is an atmospheric reanalysis data used for estimating PM2.5. MERRA-2 does not explicitly measure PM2.5 but rather follows an empirical model. MERRA-2 data were spatiotemporally collocated with ground observation for validation across India. Significant underestimation in MERRA-2 prediction of PM2.5 was observed over many monitoring stations ranging from −20 to 60 μg m−3. The utility of Machine Learning (ML) models to overcome this challenge was assessed. MERRA-2 aerosol and meteorological parameters were the input features used to train and test the individual ML models and compare them with the stacking technique. Initially, with 10% of randomly selected data, individual model performance was assessed to identify the best model. XGBoost (XGB) was the best model (r2 = 0.73) compared to Random Forest (RF) and LightGBM (LGBM). Stacking was then applied by keeping XGB as a meta-regressor. Stacked model results (r2 = 0.77) outperformed the best standalone estimate of XGB. Stacking technique was used to predict hourly and daily PM2.5 in different regions across India and each monitoring station. The eastern region exhibited the best hourly prediction (r2 = 0.80) and substantial reduction in Mean Bias (MB = −0.03 μg m−3), followed by the northern region (r2 = 0.63 and MB = −0.10 μg m−3), which showed better output due to the frequent observation of PM2.5 >100 μg m−3. Due to sparse data availability to train the ML models, the lowest performance was for the central region (r2 = 0.46 and MB = −0.60 μg m−3). Overall, India's PM2.5 prediction was good on an hourly basis compared to a daily basis using the ML stacking technique.

Deep-Learning-Based Classification of Bangladeshi Medicinal Plants Using Neural Ensemble Models

This research addresses the lack of publicly available datasets for Bangladeshi medicinal plants by presenting a comprehensive dataset comprising 5000 images of ten species collected under controlled conditions. To improve performance, several preprocessing techniques were employed, such as image selection, background removal, unsharp masking, contrast-limited adaptive histogram equalization, and morphological gradient. Then, we applied five state-of-the-art deep learning models to achieve benchmark performance on the dataset: VGG16, ResNet50, DenseNet201, InceptionV3, and Xception. Among these models, DenseNet201 demonstrated the highest accuracy of 85.28%. In addition to benchmarking the deep learning models, three novel neural network architectures were developed: dense-residual–dense (DRD), dense-residual–ConvLSTM-dense (DRCD), and inception-residual–ConvLSTM-dense (IRCD). The DRCD model achieved the highest accuracy of 97%, surpassing the benchmark performances of individual models. This highlights the effectiveness of the proposed architectures in capturing complex patterns and dependencies within the data. To further enhance classification accuracy, an ensemble approach was adopted, employing both hard ensemble and soft ensemble techniques. The hard ensemble achieved an accuracy of 98%, while the soft ensemble achieved the highest accuracy of 99%. These results demonstrate the effectiveness of ensembling techniques in boosting overall classification performance. The outcomes of this study have significant implications for the accurate identification and classification of Bangladeshi medicinal plants. This research provides valuable resources for traditional medicine, drug discovery, and biodiversity conservation efforts. The developed models and ensemble techniques can aid researchers, botanists, and practitioners in accurately identifying medicinal plant species, thereby facilitating the utilization of their therapeutic potential and contributing to the preservation of biodiversity.

Evaluation of CMIP6 models in reproducing observed rainfall over Ethiopia

Abstract Ethiopia is highly susceptible to the effects of climate change and variability. This study evaluated the performances of 37 CMIP6 models against a gridded rainfall product of Ethiopia known as Enhancing National Climate Services (ENACTS) in simulating the observed rainfall from 1981 to 2014. Taylor Skill Score was used for ranking the performance of individual models for mean monthly, June–September, and February–May seasonal rainfall. Comprehensive rating metrics (RM) were used to derive the overall ranks of the models. Results show that the performances of the models were not consistent in reproducing rainfall distributions at different statistical metrics and timeframes. More than 20 models simulated the largest dry bias on high topographic and rainfall-receiving areas of the country during the June–September season. The RM-based overall ranks of CMIP6 models showed that GFDL-CM4 is the best-performing model followed by GFDL-ESM4, NorESM2-MM, and CESM2 in simulating rainfall over Ethiopia. The ensemble of these four Global Climate Models showed the best performance in representing the spatiotemporal patterns of the observed rainfall relative to the ensembles of all models. Generally, this study highlighted the existence of dry bias in climate model projections for Ethiopia, which requires bias adjustment of the models, for impact assessment.

Deep learning framework for rapid and accurate respiratory COVID-19 prediction using chest X-ray images

COVID-19 is a contagious disease that affects the human respiratory system. Infected individuals may develop serious illnesses, and complications may result in death. Using medical images to detect COVID-19 from essentially identical thoracic anomalies is challenging because it is time-consuming, laborious, and prone to human error. This study proposes an end-to-end deep-learning framework based on deep feature concatenation and a Multi-head Self-attention network. Feature concatenation involves fine-tuning the pre-trained backbone models of DenseNet, VGG-16, and InceptionV3, which are trained on a large-scale ImageNet, whereas a Multi-head Self-attention network is adopted for performance gain. End-to-end training and evaluation procedures are conducted using the COVID-19_Radiography_Dataset for binary and multi-classification scenarios. The proposed model achieved overall accuracies (96.33% and 98.67%) and F1_scores (92.68% and 98.67%) for multi and binary classification scenarios, respectively. In addition, this study highlights the difference in accuracy (98.0% vs. 96.33%) and F_1 score (97.34% vs. 95.10%) when compared with feature concatenation against the highest individual model performance. Furthermore, a virtual representation of the saliency maps of the employed attention mechanism focusing on the abnormal regions is presented using explainable artificial intelligence (XAI) technology. The proposed framework provided better COVID-19 prediction results outperforming other recent deep learning models using the same dataset.

Multiple model averaging methods for predicting regional rice yield

AbstractUsing multiple crop models in an ensemble can generate more accurate crop growth predictions than only using an individual crop model. However, few studies have investigated the performances of individual crop models and multiple model averaging (MMA) methods for predicting regional rice (Oryza sativa L.) yield with limited inputs and observations. This study assessed the performances of three individual crop models, that is, AquaCrop, WOFOST, and Oryza version 3 (OryzaV3), and five MMA methods for predicting regional rice yield. Results showed that the AquaCrop model achieved better performances than the WOFOST and OryzaV3 models for predicting regional rice yield, especially for the early‐season rice yield, which implied that a crop model with a simple structure was powerful for making accurate regional rice yield predictions. The MMA methods achieved better performances than the AquaCrop model for predicting the late‐season rice yield; however, the AquaCrop model and three information criterion (IC) methods, that is, Akaike IC (AIC), Bayesian IC (BIC), and AICc (bias‐corrected), achieved more acceptable prediction accuracies of the early‐season rice yield than the other MMA methods. Overall, not all the MMA methods could necessarily produce more accurate regional rice yield prediction than individual crop models. This study suggested using the simple‐structure crop models and the AIC, AICc, and BIC methods for making accurate regional crop yield prediction when limited data are available for model parameters calibration.

Prediction of Cereal Rye Cover Crop Biomass and Nutrient Accumulation Using Multi-Temporal Unmanned Aerial Vehicle Based Visible-Spectrum Vegetation Indices

In general, remote sensing studies assessing cover crop growth are species nonspecific, use imagery from satellites or modified unmanned aerial vehicles (UAVs), and rely on multispectral vegetation indexes (VIs). However, using RGB imagery and visible-spectrum VIs from commercial off-the-shelf (COTS) UAVs to assess species specific cover crop growth is limited in the current scientific literature. Thus, this study evaluated RGB imagery and visible-spectrum VIs from COTS UAVs for suitability to estimate concentration (%) and content (kg ha−1) based cereal rye (CR) biomass, carbon (C), nitrogen (N), phosphorus (P), potassium (K), and sulfur (S). UAV surveys were conducted at two fields in Indiana and evaluated five visible-spectrum VIs—Visible Atmospherically Resistant Index (VARI), Green Leaf Index (GLI), Modified Green Red Vegetation Index (MGRVI), Red Green Blue Vegetation Index (RGBVI), and Excess of Greenness (ExG). This study utilized simple linear regression (VI only) and stepwise multiple regression (VI with weather and geographic data) to produce individual models for estimating CR biomass, C, N, P, K, and S concentration and content. The goodness-of-fit statistics were generated using repeated K-fold cross-validation to compare individual model performance. In general, the models developed using simple linear regression were inferior to those developed using the multiple stepwise regression method. Furthermore, for models developed using the multiple stepwise regression method all five VIs performed similarly when estimating concentration-based CR variables; however, when estimating content-based CR variables the models developed with GLI, MGRVI, and RGBVI performed similarly explaining 74–81% of the variation in CR data, and outperformed VARI and ExG. However, on an individual field basis, MGRVI consistently outperformed GLI and RGBVI for all CR characteristics. This study demonstrates the potential to utilize COTS UAVs for estimating in-field CR characteristics; however, the models generated in this study need further development to expand geographic scope and incorporate additional abiotic factors.

Forecasting and trading Bitcoin with machine learning techniques and a hybrid volatility/sentiment leverage

AbstractThis paper explores the use of machine learning algorithms and narrative sentiments when applied to the task of forecasting and trading Bitcoin. The forecasting framework starts from the selection among 295 individual prediction models. Three machine learning approaches, namely, neural networks, support vector machines, and gradient boosting approach, are used to further improve the forecasting performance of individual models. By taking data‐snooping bias into account, three different metrics are applied to examine the forecasting ability of each model. Our results suggest that the machine learning techniques always outperform the best individual model whereas the gradient boosting framework has the best performance among all the models. Finally, a time‐varying leverage trading strategy combined with narrative sentiments and volatility is proposed to enhance trading performance. This suggests that the hybrid leverage strategy provides the highest Bitcoin profits consistently among all trading exercises.

Re-considering the status quo: Improving calibration of land use change models through validation of transition potential predictions

The increasing complexity of the dynamics captured in Land Use and Land Cover (LULC) change modelling has made model behaviour less transparent and calibration more extensive. For cellular automata models in particular, this is compounded by the fact that validation is typically performed indirectly, using final simulated change maps; rather than directly considering the probabilistic predictions of transition potential. This study demonstrates that evaluating transition potential predictions provides detail into model behaviour and performance that cannot be obtained from simulated map comparison alone. This is illustrated by modelling LULC transitions in Switzerland using both Logistic Regression and Random Forests. The results emphasize the need for LULC modellers to explicitly consider the performance of individual transition models independently to ensure robust predictions. Additionally, this study highlights the potential for predictor variable selection as a means to improve transition model generalizability and parsimony, which is beneficial for simulating future LULC change.

A novel hybrid method for oil price forecasting with ensemble thought

As a globally hard currency, crude oil has influenced all aspects of the world, especially because volatility in its price seriously affect economies, politics and wars. The prediction of oil prices is challenging due to remarkable price volatility, uncertainty, elusiveness, and complexity. There have been many papers adopting traditional machine learning (ML), economic approaches or combinations of the two; however, these papers have rarely focused on the performance of individual models and ignored less popular but complex models. Considering the above difficulties and situations, in this paper, we design an advanced approach for valuable and robust forecasting of crude oil prices to investigate the differences among the three other popular methods. For this purpose, the Random Forest (RF), XGBoost (XGB), and LightGBM (LGBM) are employed in the proposed hybrid method. The merits of this hybridization lie in the fact that the ensemble model is capable of handling volatile features such as nonlinearity, noncyclicity, and market interrelationships existing in oil price time series; incorporating the strengths of the three single models; not easily overfitting; and finally achieving a better performance. The results reveal that the hybrid method achieves the highest Directional Accuracy (DA) and the lowest errors (0.9612, 13.7417, and 0.0368 of DA, MAE, and MAPE, respectively) compared with the results of three other methods-the RF (0.9569, 14.2699, and 0.0385, respectively), XGB (0.9526, 14.5111, and 0.0387, respectively), and the LGBM (0.9526, 14.4022, and 0.0384, respectively)-in experiments on real datasets.

Development of QSAR models to predict blood-brain barrier permeability.

Assessing drug permeability across the blood-brain barrier (BBB) is important when evaluating the abuse potential of new pharmaceuticals as well as developing novel therapeutics that target central nervous system disorders. One of the gold-standard in vivo methods for determining BBB permeability is rodent log BB; however, like most in vivo methods, it is time-consuming and expensive. In the present study, two statistical-based quantitative structure-activity relationship (QSAR) models were developed to predict BBB permeability of drugs based on their chemical structure. The in vivo BBB permeability data were harvested for 921 compounds from publicly available literature, non-proprietary drug approval packages, and University of Washington's Drug Interaction Database. The cross-validation performance statistics for the BBB models ranged from 82 to 85% in sensitivity and 80-83% in negative predictivity. Additionally, the performance of newly developed models was assessed using an external validation set comprised of 83 chemicals. Overall, performance of individual models ranged from 70 to 75% in sensitivity, 70-72% in negative predictivity, and 78-86% in coverage. The predictive performance was further improved to 93% in coverage by combining predictions across the two software programs. These new models can be rapidly deployed to predict blood brain barrier permeability of pharmaceutical candidates and reduce the use of experimental animals.

Microcirculatory alterations in critically ill COVID-19 patients analyzed using artificial intelligence

BackgroundThe sublingual microcirculation presumably exhibits disease-specific changes in function and morphology. Algorithm-based quantification of functional microcirculatory hemodynamic variables in handheld vital microscopy (HVM) has recently allowed identification of hemodynamic alterations in the microcirculation associated with COVID-19. In the present study we hypothesized that supervised deep machine learning could be used to identify previously unknown microcirculatory alterations, and combination with algorithmically quantified functional variables increases the model’s performance to differentiate critically ill COVID-19 patients from healthy volunteers.MethodsFour international, multi-central cohorts of critically ill COVID-19 patients and healthy volunteers (n = 59/n = 40) were used for neuronal network training and internal validation, alongside quantification of functional microcirculatory hemodynamic variables. Independent verification of the models was performed in a second cohort (n = 25/n = 33).ResultsSix thousand ninety-two image sequences in 157 individuals were included. Bootstrapped internal validation yielded AUROC(CI) for detection of COVID-19 status of 0.75 (0.69–0.79), 0.74 (0.69–0.79) and 0.84 (0.80–0.89) for the algorithm-based, deep learning-based and combined models. Individual model performance in external validation was 0.73 (0.71–0.76) and 0.61 (0.58–0.63). Combined neuronal network and algorithm-based identification yielded the highest externally validated AUROC of 0.75 (0.73–0.78) (P < 0.0001 versus internal validation and individual models).ConclusionsWe successfully trained a deep learning-based model to differentiate critically ill COVID-19 patients from heathy volunteers in sublingual HVM image sequences. Internally validated, deep learning was superior to the algorithmic approach. However, combining the deep learning method with an algorithm-based approach to quantify the functional state of the microcirculation markedly increased the sensitivity and specificity as compared to either approach alone, and enabled successful external validation of the identification of the presence of microcirculatory alterations associated with COVID-19 status.

Generalizable multi-task, multi-domain deep segmentation of sparse pediatric imaging datasets via multi-scale contrastive regularization and multi-joint anatomical priors.

Clinical diagnosis of the pediatric musculoskeletal system relies on the analysis of medical imaging examinations. In the medical image processing pipeline, semantic segmentation using deep learning algorithms enables an automatic generation of patient-specific three-dimensional anatomical models which are crucial for morphological evaluation. However, the scarcity of pediatric imaging resources may result in reduced accuracy and generalization performance of individual deep segmentation models. In this study, we propose to design a novel multi-task, multi-domain learning framework in which a single segmentation network is optimized over the union of multiple datasets arising from distinct parts of the anatomy. Unlike previous approaches, we simultaneously consider multiple intensity domains and segmentation tasks to overcome the inherent scarcity of pediatric data while leveraging shared features between imaging datasets. To further improve generalization capabilities, we employ a transfer learning scheme from natural image classification, along with a multi-scale contrastive regularization aimed at promoting domain-specific clusters in the shared representations, and multi-joint anatomical priors to enforce anatomically consistent predictions. We evaluate our contributions for performing bone segmentation using three scarce and pediatric imaging datasets of the ankle, knee, and shoulder joints. Our results demonstrate that the proposed approach outperforms individual, transfer, and shared segmentation schemes in Dice metric with statistically sufficient margins. The proposed model brings new perspectives towards intelligent use of imaging resources and better management of pediatric musculoskeletal disorders.