Variable Subset Selection Research Articles

Assessment of Patient Self-awareness and Related Neural Correlates in Frontotemporal Dementia and Corticobasal Syndrome.

We compared two different methods of assessing self-awareness (clinician-rated vs. self- and caregiver report) in participants with neurodegenerative conditions. Additionally, we examined the contribution of memory dysfunction to assessment of self-awareness. Sixty-seven participants with various neurodegenerative disorders participated in this study. Data were collected on brain volume, neurocognitive function, demographic characteristics, and two measures of patient self-awareness, defined as (1) the discrepancy between patient and caregiver ratings of dysexecutive syndrome and (2) clinician-observed rating of patient insight. Penalized regression with best subset variable selection and 10-fold cross-validation was used to evaluate three neurocognitive frameworks: self-regulation, language, and perspective-taking, each predicting the results from the two methods of self-awareness measurement. The self-regulation framework was more robustly predictive for both the clinician rating and discrepancy method than language or perspective-taking. Frameworks in which the clinician rating was the criterion were more robust than those with the discrepancy method as criterion. When a measure of memory functioning was added to the framework, there was no appreciable improvement in the prediction of self-awareness. A self-regulation neurocognitive framework, consisting of regions of interest and neuropsychological test scores, was more effective in understanding patient self-awareness than perspective-taking or language frameworks. Compared to the discrepancy method, a clinician rating of self-awareness was more robustly associated with relevant clinical variables of regional brain volume and neuropsychological performance, suggesting it may be a useful measure to aid clinical diagnosis.

The conformation-independent QSPR approach for predicting the oxidation rate constant of water micropollutants

In advanced water treatment processes, the degradation efficiency of contaminants depends on the reactivity of the hydroxyl radical toward a target micropollutant. The present study predicts the hydroxyl radical rate constant in water (k OH) for 118 emerging micropollutants, by means of quantitative structure-property relationships (QSPR). The conformation-independent QSPR approach is employed, together with a large number of 15,251 molecular descriptors derived with the PaDEL, Epi Suite, and Mold2 freewares. The best multivariable linear regression (MLR) models are found with the replacement method variable subset selection technique. The proposed five-descriptor model has the following statistics for the training set: [Formula: see text], RMS train = 0.21, while for the test set is [Formula: see text], RMS test = 0.11. This QSPR serves as a rational guide for predicting oxidation processes of micropollutants.

Predicting the bioconcentration factor through a conformation-independent QSPR study

The ANTARES dataset is a large collection of known and verified experimental bioconcentration factor data, involving 851 highly heterogeneous compounds from which 159 are pesticides. The BCF ANTARES data were used to derive a conformation-independent QSPR model. A large set of 27,017 molecular descriptors was explored, with the main intention of capturing the most relevant structural characteristics affecting the studied property. The structural descriptors were derived with different freeware tools, such as PaDEL, Epi Suite, CORAL, Mold2, RECON, and QuBiLs-MAS, and so it was interesting to find out the way that the different descriptor tools complemented each other in order to improve the statistical quality of the established QSPR. The best multivariable linear regression models were found with the Replacement Method variable sub-set selection technique. The proposed QSPR model improves previous reported models of the bioconcentration factor in the present dataset.

The quantitative structure–insecticidal activity relationships from plant derived compounds against chikungunya and zika Aedes aegypti (Diptera:Culicidae) vector

The insecticidal activity of a series of 62 plant derived molecules against the chikungunya, dengue and zika vector, the Aedes aegypti (Diptera:Culicidae) mosquito, is subjected to a Quantitative Structure–Activity Relationships (QSAR) analysis. The Replacement Method (RM) variable subset selection technique based on Multivariable Linear Regression (MLR) proves to be successful for exploring 4885 molecular descriptors calculated with Dragon 6. The predictive capability of the obtained models is confirmed through an external test set of compounds, Leave-One-Out (LOO) cross-validation and Y-Randomization. The present study constitutes a first necessary computational step for designing less toxic insecticides.

A three-stage variable selection method for supersaturated designs

ABSTRACTA supersaturated design (SSD) is a design whose run size is not enough for estimating all main effects. Such a design is commonly used in screening experiments to screen active effects based on the effect sparsity principle. Traditional approaches, such as the ordinary stepwise regression and the best subset variable selection, may not be appropriate in this situation. In this article, a new variable selection method is proposed based on the idea of staged dimensionality reduction. Simulations and several real data studies indicate that the newly proposed method is more effective than the existing data analysis methods.

Conformation-Independent QSPR Approach for the Soil Sorption Coefficient of Heterogeneous Compounds.

We predict the soil sorption coefficient for a heterogeneous set of 643 organic non-ionic compounds by means of Quantitative Structure-Property Relationships (QSPR). A conformation-independent representation of the chemical structure is established. The 17,538 molecular descriptors derived with PaDEL and EPI Suite softwares are simultaneously analyzed through linear regressions obtained with the Replacement Method variable subset selection technique. The best predictive three-descriptors QSPR is developed on a reduced training set of 93 chemicals, having an acceptable predictive capability on 550 test set compounds. We also establish a model with a single optimal descriptor derived from CORAL freeware. The present approach compares fairly well with a previously reported one that uses Dragon descriptors.

A new structure-based model for estimation of true critical volume of multi-component mixtures

In this study, a new quantitative structure–property relationship (QSPR) has been proposed to estimate true vapor–liquid critical volume of multi-component mixtures. In developing this model, 598 experimental data on true critical volume for 110 different binary mixtures were applied. The mixture molecular descriptors were computed according to the molecular descriptors of pure components involved in the mixture and their molar fractions. Enhanced replacement method (ERM), as an effective tool for subset variable selection, was utilized. The proposed model is simple multivariate linear equations with six variables. The prediction capability of the proposed model for different families of compounds was critically analyzed. Besides, the capability of the proposed model has been tested by predicting true critical volume of 7 different multi-component mixtures containing 40 experimental data points. The average absolute relative deviation of the proposed QSPR model over all experimental data is 9.7% and 8.7% for binary mixtures and multi-component mixtures, respectively.

Chemometric modeling to predict aquatic toxicity of benzene derivatives in Pimephales Promelas

Purpose – The purpose of this paper is to predict the aquatic toxicity (LC50) of 92 substituted benzenes derivatives in Pimephales promelas. Design/methodology/approach – Quantitative structure-activity relationship analysis was performed on a series of 92 substituted benzenes derivatives using multiple linear regression (MLR), artificial neural network (ANN) and support vector machines (SVM) methods, which correlate aquatic toxicity (LC50) values of these chemicals to their structural descriptors. At first, the entire data set was split according to Kennard and Stone algorithm into a training set (74 chemicals) and a test set (18 chemical) for statistical external validation. Findings – Models with six descriptors were developed using as independent variables theoretical descriptors derived from Dragon software when applying genetic algorithm – variable subset selection procedure. Originality/value – The values of Q2 and RMSE in internal validation for MLR, SVM, and ANN model were: (0.8829; 0.225), (0.8882; 0.222); (0.8980; 0.214), respectively and also for external validation were: (0.9538; 0.141); (0.947; 0.146); (0.9564; 0.146). The statistical parameters obtained for the three approaches are very similar, which confirm that our six parameters model is stable, robust and significant.

Comparison of Bayesian predictive methods for model selection

The goal of this paper is to compare several widely used Bayesian model selection methods in practical model selection problems, highlight their differences and give recommendations about the preferred approaches. We focus on the variable subset selection for regression and classification and perform several numerical experiments using both simulated and real world data. The results show that the optimization of a utility estimate such as the cross-validation (CV) score is liable to finding overfitted models due to relatively high variance in the utility estimates when the data is scarce. This can also lead to substantial selection induced bias and optimism in the performance evaluation for the selected model. From a predictive viewpoint, best results are obtained by accounting for model uncertainty by forming the full encompassing model, such as the Bayesian model averaging solution over the candidate models. If the encompassing model is too complex, it can be robustly simplified by the projection method, in which the information of the full model is projected onto the submodels. This approach is substantially less prone to overfitting than selection based on CV-score. Overall, the projection method appears to outperform also the maximum a posteriori model and the selection of the most probable variables. The study also demonstrates that the model selection can greatly benefit from using cross-validation outside the searching process both for guiding the model size selection and assessing the predictive performance of the finally selected model.

QSPR Study of the Water Solubility of a Diverse Set of Agrochemicals : Hybrid ( GA / MLR ) Approach = Etude QSPR de la Solubilité Aqueuse d'un Ensemble de Divers Produits Agrochimiques : Approche Hybride ( AG / RLM )

Pesticides | Aqueous Solubility | QSPR | Molecular Descriptors | Multiple Linear Regression | Pesticides | Solubilité Aqueuse | QSPR | Descripteurs Moléculaires | Régression Linéaire Multiple | المبيدات | الانحلالية | QSPR | الموصفات الجزيئية | التراجع المتعدد الخطي

Computational Model for Chromatographic Relative Retention Time of Polychlorinated Biphenyls Using Sub-structural Molecular Fragments

Quantitative structure-retention relationship (QSRR) analysis is a useful technique capable of relating chromato- graphic retention time to the chemical structure of a solute. Using the sub-structural molecular fragments (SMF) derived directly from the molecular structures, the gas chromatographic relative retention times (RRTs) of 209 polychlorinated biphenyls (PCBs) on the SE-54 stationary phase were calculated. An eight-variable regression equation with the correlation coefficient of 0.9945 and the root mean square errors of 0.0134 was developed. Forward and backward stepwise regression variable selection and multi-linear regression analysis (MLRA) are combined to describe the effect of molecular structure on the RRT of PCB according to the QSRR method. To quantitatively relate RRT with the molecular structure MLR analysis is performed on the set of 163 sub-structural molecular fragments (SMF) provided by the ISIDA software. The eight fragments selected by variable subset selection, all belonging to the sub-fragments, adequately represent the structural factors influencing the affinity of PCB to SE-54 stationary phase in the separation process. Finally, a QSRR model is selected based on leave-one-out cross-validation and its prediction ability is further tested on 42 representative compounds excluded from model calibration. The prediction results from the MLR model are in good agreement with the experimental values. By applying the MLR method we can predict the test set with squared cross validated correlation coefficient (Q 2) of 0.9913 and root mean square error (RMSE) of 0.0169.

A New QSPR Study on Relative Sweetness

The aim of this work was to develop predictive structure-property relationships (QSPR) of natural and synthetic sweeteners in order to predict and model relative sweetness (RS). The data set was composed of 233 sweeteners collected from diverse sources in the literature, which was divided into training (163) and test (70) molecules according to a procedure based on k-means cluster analysis. A total of 3763 non-conformational Dragon molecular descriptors were calculated which were simultaneously analyzed through multivariable linear regression analysis coupled with the replacement method variable subset selection technique. The established six-parameter model was validated through the cross-validation techniques, together with Y-randomization and applicability domain analysis. The results for the training set and the test set showed that the non-conformational descriptors offer relevant information for modeling the RS of a compound. Thus, this model can be used to predict the sweetness of both un-evaluated and un-synthesized sweeteners.

Best Subsets Variable Selection in Nonnormal Regression Models

We present a new program, gvselect, that helps users perform variable selection in regression. Best subsets variable selection is performed and provides the user with the best combinations of predictors for each level of model complexity. The leaps-and-bounds (Furnival and Wilson, 1974, Technometrics 16: 499–511) algorithm is applied using the log likelihoods of candidate models. This allows the user to perform variable selection on a wide variety of normal and non-normal regression models. Our method is described in Lawless and Singhal (1978, Biometrics 34: 318–327).

Swarm Intelligence for Chemometrics

Together with the growing use of advanced analytical instruments [ranging from separation techniques such as gas chromatography to light-based techniques, e.g. near infrared (NIR) spectroscopy], interest in the use of multivariate calibration [e.g. partial least squares (PLS)] has emerged in many different industries (food, chemical, pharmaceutical etc). Although many different modelling techniques can be used, these techniques normally require either a variable selection before model building (e.g. multiple linear regression) or they use a dimension reduction method to create latent variables (e.g. partial least squares regression). In spite of the fact that there are many different variable selection techniques, most attention is paid to those based on swarm intelligence methodologies. The reason is that such methodologies are usually simulated based on animal and insect life behaviour to, for example, find the shortest path between a food source and a nest. From an optimisation point of view, a solution or decision is made by a crowd (e.g. a population of ants) which leads to a more robust and global optimised system (in this case a better subset selection of most relevant variables e.g. wavelengths from NIR spectra).

Quantitative structure–property relationship analysis for the retention index of fragrance-like compounds on a polar stationary phase

A quantitative structure–property relationship (QSPR) was developed for modeling the retention index of 1184 flavor and fragrance compounds measured using a Carbowax 20M glass capillary gas chromatography column. The 4885 molecular descriptors were calculated using Dragon software, and then were simultaneously analyzed through multivariable linear regression analysis using the replacement method (RM) variable subset selection technique. We proceeded in three steps, the first one by considering all descriptor blocks, the second one by excluding conformational descriptor blocks, and the last one by analyzing only 3D-descriptor families. The models were validated through an external test set of compounds. Cross-validation methods such as leave-one-out and leave-many-out were applied, together with Y-randomization and applicability domain analysis. The developed model was used to estimate the I of a set of 22 molecules. The results clearly suggest that 3D-descriptors do not offer relevant information for modeling the retention index, while a topological index such as the Randić-like index from reciprocal squared distance matrix has a high relevance for this purpose.

Variability in mortality after caesarean delivery, appendectomy, and groin hernia repair in low-income and middle-income countries: implications for expanding surgical services

BackgroundWhile surgical interventions occur at lower rates in resource-poor settings, rates of complication and death after surgery are substantial but have not been well quantified. A deeper understanding of outcomes is a crucial step to ensure that quality accompanies increased global access to surgical care. We aimed to assess mortality following surgery to assess the risks of such interventions in these environments. MethodsWe collected the most recent demographic, health, and economic data from WHO for 114 countries classified as low-income or lower-middle-income according to the World Bank in 2005. We searched OVID, MedLine, PubMed, and SCOPUS to identify studies in these countries reporting all-cause mortality after three commonly performed operations: caesarean delivery, appendectomy, and groin hernia repair. Reports from governmental and other agencies were also identified. We modelled surgical mortality rates for countries without reported data with the lasso technique that performs continuous variable subset selection to avoid model overfitting. We validated our model against known case fatality rates for caesarean delivery. We aggregated mortality results by subregion to account for variability in data availability. We then created collective surgical case fatality rates by WHO region. FindingsWe identified 42 countries with mortality data for at least one of the three procedures. Median reported mortality rates were 7·7 per 1000 operations for caesarean delivery (IQR 3–14), 4·0 per 1000 operations for appendectomy (IQR 0–17), and 4·7 per 1000 operations for hernia groin (IQR 0–13); all recorded deaths occurred during the same admission to hospital as the operation. Based on our model, case fatality rate estimates by subregion ranged from 0·7 (central Europe) to 13·9 (central sub-Saharan Africa) per 1000 caesarean deliveries, 5·6 (central Asia) to 6·4 (central sub-Saharan Africa) per 1000 appendectomies, and 3·5 (tropical Latin America) to 33·9 (central sub-Saharan Africa) per 1000 hernia repairs. InterpretationAll-cause postoperative mortality rates are exceedingly variable within resource-constrained environments, and substantially higher than those in middle-income and high-income settings. Efforts to expand surgical access and provision of services must include a strong commitment to improve the safety and quality of care. FundingNone.

Abstract MP38: Local Population Income, Geographic Space and Interactions Predict Increased Presence of Physical Activity Facilities in New York City Metropolitan Area Census Tracts, 1990-2010

Introduction: Physical activity is associated with improved health and is supported, in part, by the presence of facilities that provide space and equipment to pursue a variety of physical activities. We assessed the hypothesis that socio-geographic characteristics predict increased local availability of commercial physical activity facilities over time. Longitudinal examination of physical activity facility distribution can inform our response to current disparities in access to public and private physical activity venues. Methods: We used data from the National Establishment Time-Series (NETS), a longitudinal database of U.S. businesses, focusing on 4528 census tracts (23 counties) in the New York City metropolitan area and on decennial intervals for which population data were also available through the Census or American Community Survey (1990, 2000, 2010). Commercial physical activity facilities (e.g., gyms, tennis courts, martial arts studios) were defined based on Standard Industrial Classification (SIC) codes and name searches. Facility counts were aggregated to 2010 census tract boundaries and linked to local population characteristics. Comparisons across decennial intervals were used to define increasing count of physical activity facilities and shifting population demographics. Associations were evaluated using lasso logistic regression to estimate relationships with predictor variables and their interactions with model shrinkage and variable subset selection through 10-fold cross-validation for minimization of test set model deviance. Results: Census tracts with at least one physical activity facility increased over time (1990=1172, 2000=2295, 2010=2365). Greater tract-level median income, larger land area, and higher previous total physical activity facilities at start of decade were positively associated with greater odds for local increase in physical activity facilities (OR=1.27 per SD median income; OR=1.30 per SD land area; OR=1.14 per SD lagged facility count). Inclusion of two-way interaction terms increased R2 estimates from 0.30 to 0.33, suggesting explanation of an additional 3% of the variation in facility count increase. Subset selection through lasso to minimize cross-validation error resulted in retention of 11 of 21 possible two-way predictor interactions. The association between 10-year increase in median income with increased physical activity facility count was stronger in geographically larger census tracts (interaction OR=1.05); similarly, a stronger relationship was found for 10-year population count increase with physical facility count increase in larger census tracts (interaction OR=1.05). Conclusion: Local population, geographic, and business environment characteristics are associated with change in physical activity facilities. Inclusion of interaction terms improved prediction.

QSPR study on solubility of some fullerenes derivatives using the genetic algorithms — Multiple linear regression

A quantitative structure–property relation study was performed on the solubility of C60 and C70 fullerene derivatives. Topological and geometrical as well as quantum mechanical energy-related and charge distribution-related descriptors, generated from CODESSA, were calculated to define the molecule structures requirement for measuring their correlations with solubility. The best four variables among the other subsets were selected by the genetic algorithm variable subset selection procedure. Modeling of the relationship between selected molecular descriptors and solubility data was achieved by multiple linear regression method (R2train=0.801, R2test=0.792, Q2LOO=0.716, Q2BOOT=0.674). The robustness and the predictive performance of the proposed model were verified using both internal (cross-validation by leave one out, bootstrap, Y-scrambling) and external statistical validations (external test set by splitting the original data set into training and test sets by k-nearest neighbor (kNN) classification method). Further, the external predictive power of the developed model was examined by considering modified r2 and concordance correlation coefficient values. The reactivity, the polar interactions, the electron–electron repulsion energy, the electronuclear attraction energy, the nuclear–nuclear repulsion energy, and the rotational–vibrational energies were the main independent factors contributing to the solubility of the fullerenes.

Molecular based models for estimation of critical properties of pure refrigerants: Quantitative structure property relationship (QSPR) approach

In the present survey, new quantitative structure–property relationships (QSPR) have been proposed to estimate critical properties of pure refrigerants. For developing these models, the experimental data of TC for 198 pure compounds, the experimental data of VC for 194 pure compounds, and the experimental data of PC for 197 pure compounds were applied. For each compound, 1650 molecular descriptors were determined. Enhanced replacement method (ERM), as an effectual tool for subset variable selection, was utilized. The proposed models are simple multivariate linear equations with six variables for estimation of TC, five variables for estimation of VC, and six variables for estimation of PC. The variables of models are computed from the molecular structure of refrigerants. The average absolute relative deviation (AARD, %) of the proposed models over all experimental data are 2.65%, 3.76%, and 4.65% for estimation of TC, VC, and PC, respectively.

Estudio de la Relación Cuantitativa Estructura-Actividad de pesticidas mediante técnicas de clasificación

El objetivo de este trabajo fue la comparación entre los métodos de clasificación del vecino más cercano (κ-NN) y las redes neuronales artificiales de contrapropagación (CP-ANN) para modelar la toxicidad de un conjunto de 192 pesticidas organoclorados, organofosforados, carbamatos y piretroides, medidos como Concentración Efectiva (EC50) y que fueron divididos en tres clases, es decir, baja, intermedia y alta toxicidad. Se calcularon 4885 descriptores moleculares usando el programa DRAGON, los que fueron simultáneamente analizados mediante el método κ-NN acoplado con la técnica de selección de variables de los Algoritmos Genéticos (GA-VSS). Los modelos fueron apropiadamente validados mediante un subconjunto de predicción. Los resultados claramente sugieren que los descriptores 3D no ofrecen información relevante para modelar las clases. Por otro lado, κ-NN muestra mejores resultados que CP-ANN.