Differences In Predictive Performance Research Articles

Particle size distribution models for soils of the humid tropics

Standardisation of particle size distribution (PSD) is a prerequisite to achieve compatibility of soil data among various countries with different texture classification systems. Therefore, several mathematical models have been proposed to accurately represent PSD. Previous studies evaluated the performance of such models to describe PSD of soils from temperate regions. This study aims at evaluating the performance of models for describing PSD of soils from the humid tropics based on a large dataset. A dataset of 1,412 soils from Central Africa representing 11 different FAO Soil Groups was used. Ten PSD models with two to four fitting parameters were selected: simple log-normal (LN_2p), van Genuchten-type1 (VG_2p), van Genuchten-type2 (vG_3p), Fredlund-type1 (F_3p), Fredlund-type2 (F_4p), Weibull (W_3p), Skaggs (Sk_3p), Gompertz-type1 (G_2p), Gompertz-type2 (G_4p) and Andersson (A_4p). The fitting performance of the PSD models was evaluated by three statistical indices: the adjusted coefficient of determination, the Akaike information criterion and the relative error. Clustered columns and box plots were also used to get more insights. The predictive ability of the best PSD models was tested using a leave-one-out method and 1:1 plots. A table of initial values for the fitting parameters of each PSD model was provided for future applications. Some models like VG_2p, VG_3p, Sk_3p and G_4p were not suitable to describe PSD of soils in the humid tropics. On the other hand, F_3p, F_4p, W_3p and A_4p models showed outstanding fitting performance. The fitting performance of PSD models was also dependent of the textural class, the broad textural group and the bimodal character of the soil. For the most frequent textural classes in the dataset, the F_3p and A_4p models were the best closely followed by the W_3p model. While the F_3p model performed better than the A_4p model for coarse-textured soils, the opposite was observed for fine-textured soils. The W_3p model showed acceptable fitting performance for fine, medium and coarse-textured soils. The performance of the PSD models was found to be better for bimodal soils, which are common in the humid tropics, than for unimodal soils. Great differences in fitting and prediction performance were found between the PSD models. Soil texture as well as the bimodal character of the soil significantly affect their respective performance. Some models like VG_2p, VG_3p, Sk_3p and G_4p are not suitable to describe PSD of soils of the humid tropics. On the other hand, F_3p, F_4p, W_3p and A_4p models showed outstanding fitting performance. Therefore, they are highly recommended in order to get a better description of the PSD of soils of the humid tropics.

The Multilingual Naming Test in Alzheimer's Disease: Clues to the Origin of Naming Impairments

The current study explored the picture naming performance of patients with Alzheimer’s disease (AD). First, we evaluated the utility of the multilingual naming test (MINT; Gollan et al., 2011), which was designed to assess naming skills in speakers of multiple languages, for detecting naming impairments in monolingual AD and amnestic mild cognitive impairment (MCI). If the MINT were sensitive to linguistic impairment in AD, using it in clinical practice might have advantages over using tests exclusively designed for English monolinguals. We found that the MINT can be used with both monolinguals and bilinguals: A 32-item subset of the MINT is best for distinguishing monolingual patients from controls, while the full MINT is best for assessing degree of bilingualism and language dominance in bilinguals. We then investigated the cognitive mechanisms underlying naming impairment in AD. To this end, we explored which MINT item characteristics best predicted performance differences between monolingual patients and controls. We found that contextual diversity and imageability, but not word frequency (nor words’ number of senses), contributed unique variance to explaining naming impairments in AD. These findings suggest a semantic component to the naming impairment in AD (modulated by names’ semantic richness and network size).

Frequency response areas in the inferior colliculus: nonlinearity and binaural interaction.

The tuning, binaural properties, and encoding characteristics of neurons in the central nucleus of the inferior colliculus (CNIC) were investigated to shed light on nonlinearities in the responses of these neurons. Results were analyzed for three types of neurons (I, O, and V) in the CNIC of decerebrate cats. Rate responses to binaural stimuli were characterized using a 1st- plus 2nd-order spectral integration model. Parameters of the model were derived using broadband stimuli with random spectral shapes (RSS). This method revealed four characteristics of CNIC neurons: (1) Tuning curves derived from broadband stimuli have fixed (i. e., level tolerant) bandwidths across a 50–60 dB range of sound levels; (2) 1st-order contralateral weights (particularly for type I and O neurons) were usually larger in magnitude than corresponding ipsilateral weights; (3) contralateral weights were more important than ipsilateral weights when using the model to predict responses to untrained noise stimuli; and (4) 2nd-order weight functions demonstrate frequency selectivity different from that of 1st-order weight functions. Furthermore, while the inclusion of 2nd-order terms in the model usually improved response predictions related to untrained RSS stimuli, they had limited impact on predictions related to other forms of filtered broadband noise [e. g., virtual-space stimuli (VS)]. The accuracy of the predictions varied considerably by response type. Predictions were most accurate for I neurons, and less accurate for O and V neurons, except at the lowest stimulus levels. These differences in prediction performance support the idea that type I, O, and V neurons encode different aspects of the stimulus: while type I neurons are most capable of producing linear representations of spectral shape, type O and V neurons may encode spectral features or temporal stimulus properties in a manner not easily explained with the low-order model. Supported by NIH grant DC00115.

Prediction of intracerebral hemorrhage following thrombolytic therapy for acute ischemic stroke using multiple artificial neural networks

Objectives: Artificial neural networks (ANNs) have been increasingly used in diagnosis and the prediction of outcome, mortality, and risk factors in ischemic stroke. Each model may have different accuracy, sensitivity, and specificity in processing the same clinical information. Thus, using only one model of ANNs may mislead the prediction. The present study aimed to predict symptomatic intracerebral hemorrhage (SICH) following thrombolysis in acute ischemic stroke based on clinical, laboratory, and imaging data using multiple ANN models.Methods: Models for radial basis function (RBF), multilayer perceptron (MLP), probabilistic neural network (PNN), and support vector machine (SVM) were generated to analyze 194 datasets with 29 predictive variables. The relative importance of each predictor variable was calculated using sensitivity analysis.Results: Comparison among the models based on the areas under the receiver operating characteristic curves (AUC) showed no significantly statistical difference in predictive performance among RBF, MLP, and PNN. PNN showed significantly better performance than SVM. With a minimum importance score of 50 together with an AUC value ⩾0·50, three models identified stroke subtype as an important predictive variable for SICH. Other potential predictors were stroke location, prothrombin time, low-density-lipoprotein cholesterol, diastolic blood pressure, International Normalized Ratio, and brain computed tomography findings.Discussion: Although ANN models showed similar performance, the classification results were not totally alike, suggesting an advantage of using multiple classification models over a single model. The predictive results are supported by previous statistical studies on different datasets, suggesting generalizability of the utility of ANN analyses.

Regional and global modeling estimates of policy relevant background ozone over the United States

Abstract Policy Relevant Background (PRB) ozone, as defined by the US Environmental Protection Agency (EPA), refers to ozone concentrations that would occur in the absence of all North American anthropogenic emissions. PRB enters into the calculation of health risk benefits, and as the US ozone standard approaches background levels, PRB is increasingly important in determining the feasibility and cost of compliance. As PRB is a hypothetical construct, modeling is a necessary tool. Since 2006 EPA has relied on global modeling to establish PRB for their regulatory analyses. Recent assessments with higher resolution global models exhibit improved agreement with remote observations and modest upward shifts in PRB estimates. This paper shifts the paradigm to a regional model (CAMx) run at 12 km resolution, for which North American boundary conditions were provided by a low-resolution version of the GEOS-Chem global model. We conducted a comprehensive model inter-comparison, from which we elucidate differences in predictive performance against ozone observations and differences in temporal and spatial background variability over the US. In general, CAMx performed better in replicating observations at remote monitoring sites, and performance remained better at higher concentrations. While spring and summer mean PRB predicted by GEOS-Chem ranged 20–45 ppb, CAMx predicted PRB ranged 25–50 ppb and reached well over 60 ppb in the west due to event-oriented phenomena such as stratospheric intrusion and wildfires. CAMx showed a higher correlation between modeled PRB and total observed ozone, which is significant for health risk assessments. A case study during April 2006 suggests that stratospheric exchange of ozone is underestimated in both models on an event basis. We conclude that wildfires, lightning NOx and stratospheric intrusions contribute a significant level of uncertainty in estimating PRB, and that PRB will require careful consideration in the ozone standard setting process.

Middle versus anterior cerebral artery Doppler for the prediction of perinatal outcome and neonatal neurobehavior in term small‐for‐gestational‐age fetuses with normal umbilical artery Doppler

To evaluate whether anterior cerebral artery (ACA) Doppler ultrasonography is superior to middle cerebral artery (MCA) Doppler in the prediction of perinatal outcome and neonatal neurobehavior in term small-for-gestational-age (SGA) fetuses with normal umbilical artery (UA) Doppler. MCA and ACA Doppler ultrasonography was performed in a cohort of SGA term fetuses with normal UA Doppler. Perinatal outcome and neonatal neurobehavioral performance were compared with a group of term appropriate-for-gestational age (AGA) infants. Neurobehavior was evaluated at 40 ( +/- 1) weeks of corrected age with the Neonatal Behavioral Assessment Scale. Differences between the study groups were adjusted for potential confounding variables by multiple linear or logistic regression analyis. A total of 199 newborns (98 SGA and 101 AGA) were included. Among the SGA fetuses, 28.6 and 17% had MCA and ACA redistribution, respectively. Cases with either type of redistribution had an increased risk for adverse outcome, with no differences in predictive performance between the two parameters. SGA fetuses with MCA redistribution compared with controls had an increased risk for abnormal neurobehavioral performance in motor (36 vs. 20%; adjusted P = 0.02) and state organization (25 vs. 17.5%; adjusted P = 0.03) areas. SGA fetuses with ACA redistribution had only an increased risk for abnormal neurobehavioral performance area in state organization compared with controls (30 vs. 17.5%; adjusted P = 0.021). In term SGA newborns with no signs of brain-sparing, ACA Doppler investigation does not provide any benefit over MCA in terms of the prediction of adverse perinatal outcome.

Prediction of protein protein interactions from primary sequences

The prediction of protein-protein interactions is a difficult problem in biology. In this study, an efficient method is presented to predict protein-protein interactions with sequence composition information. Four kinds of basic building blocks of protein sequences are investigated. The experimental results show that there is minor difference in prediction performance among the four kinds of different building blocks. The method based on combination of all building blocks out performs any of the building blocks. We also demonstrate that the use of Latent Semantic Analysis (LSA) can efficiently remove noise and improve the prediction efficiency without significantly degrading the performance.

Local stream habitat variables predicted from catchment scale characteristics are useful for predicting fish distribution

South-east Queensland (Australia) streams were described by 21 local habitat variables that were chosen because of their potential association with fish distribution. An Assessment by a Nearest Neighbour Analysis (ANNA) model used large-scale variables that are robust to human influence to predict what the values of each of the 21 local habitat variables at each site would be without modification from human activity. The ANNA model used elevation, stream order, distance from source and longitude to predict the local habitat variables; other candidate predictor variables (mean rainfall, latitude and catchment area) were not found to be useful. The ANNA model was able to predict five of the 21 local habitat variables (average width, sand (%), cobble (%), rocks (%) and large woody debris) with an R2 of at least 0.2. The observed values of these five local habitat variables were used to model the distributions of individual fish species. The species distribution models were developed using logistic regression based on a subset of the data (some of the data were withheld for model validation) and a forward stepwise model selection procedure. There was no difference in predictive performance of fish distribution models for model predictions based on observed values and model predictions based on ANNA predicted values of local habitat variables in the withheld data (p-value = 0.85). Therefore, it is possible to predict the suitability of sites as habitat for given fish species using estimated (estimates based on large-scale variables) natural values of local habitat variables.

Using classification trees to assess low birth weight outcomes

Low birth weight (LBW) is a major public health problem. Compared to normal weight infants, LBW is positively associated with infant mortality and negatively associated with normative childhood cognitive and physical development. In the past two decades, research has identified important risk factors of LBW. In this study, we used classification trees to study the interactive nature of these factors. In particular we: (1) identify subgroups of women who are at a high risk of a LBW outcome in seven geographical regions of Florida, and (2) study the predictive performance of classification trees by comparing the tree-based results to those obtained using logistic regression. The data, 181,690 singleton births, were derived from Florida birth certificates recorded in 1998. Classification trees and logistic regression models were built based on seven geographical regions. The outcome variable consisted of two classes, namely LBW (< 2500 g) and normal birth weight (> or = 2500 g) cases, while a large number of known risk factors was examined. Tree and logistic regression models were compared using Receiving Operating Curves, and sensitivity and specificity analyses. The use of classification trees has revealed a number of high-risk subgroups. For instance, White, Hispanic or Other non-white mothers who were healthy and smoked with a weight gain less than 20 lbs had a higher risk of a LBW birth compared to those with the same characteristics but with a weight gain of more than 20 lbs. Factors such as parity and marital status were important predictors for pregnancy outcomes among nonsmoker White, Hispanic or Other non-white mothers. Furthermore, we found that Black mothers were directly classified as a high-risk subgroup in the regions of Panhandle, Northeast, North Central, while in the Southern regions a series of other characteristics further defined the high-risk subgroup of Black mothers. Overall, the differences in predictive performance between tree models and logistic regression were minimal. The present study demonstrated that classification trees can be used to identify high-risk subgroups of mothers who are at risk of LBW outcomes. Although these exploratory tree analyses revealed a number of distinctive variable interactions for each geographical area, the variable selection was similar across all seven regions. This study also demonstrated that classification trees did not outperform logistic regression models or vice versa; both approaches provided useful analyses of the data.

Reliable prediction of T-cell epitopes using neural networks with novel sequence representations.

In this paper we describe an improved neural network method to predict T-cell class I epitopes. A novel input representation has been developed consisting of a combination of sparse encoding, Blosum encoding, and input derived from hidden Markov models. We demonstrate that the combination of several neural networks derived using different sequence-encoding schemes has a performance superior to neural networks derived using a single sequence-encoding scheme. The new method is shown to have a performance that is substantially higher than that of other methods. By use of mutual information calculations we show that peptides that bind to the HLA A*0204 complex display signal of higher order sequence correlations. Neural networks are ideally suited to integrate such higher order correlations when predicting the binding affinity. It is this feature combined with the use of several neural networks derived from different and novel sequence-encoding schemes and the ability of the neural network to be trained on data consisting of continuous binding affinities that gives the new method an improved performance. The difference in predictive performance between the neural network methods and that of the matrix-driven methods is found to be most significant for peptides that bind strongly to the HLA molecule, confirming that the signal of higher order sequence correlation is most strongly present in high-binding peptides. Finally, we use the method to predict T-cell epitopes for the genome of hepatitis C virus and discuss possible applications of the prediction method to guide the process of rational vaccine design.

The nebulous art of using wind tunnel aerofoil data for predicting rotor performance

AbstractThe objective of this study was threefold: to evaluate different two‐dimensional S809 aerofoil data sets in the prediction of rotor performance; to compare blade element momentum rotor predicted results with lifting surface, prescribed wake results; and to compare the NASA Ames combined experiment rotor measured data with the two different performance prediction methods. The S809 aerofoil data sets evaluated included those from Delft University of Technology, Ohio State University and Colorado State University. Substantial differences in prediction performance resulted from the different two‐dimensional aerofoil data sets. Predicted performance comparison with NASA Ames data documents shortcomings of these methods and recommends the use of the lifting surface, prescribed wake method over blade element momentum theory for future analytical improvements. Copyright © 2002 John Wiley &amp; Sons, Ltd.

Comparing case-based reasoning classifiers for predicting high risk software components

Case-based reasoning (CBR) has been proposed for predicting the risk class of software components. Risky components can be defined as those that are fault-prone, or those that require a large amount of effort to maintain. Thus far evaluative studies of CBR classifiers have been promising, showing that their predictive performance is as good as or better than other types of classifiers. However, a CBR classifier can be instantiated in different ways by varying its parameters, and it is not clear which combination of parameters provides the best performance. In this paper we evaluate the performance of a CBR classifier with different parameters, namely: (a) different distance measures, (b) different standardization techniques, (c) use or non-use of weights, and (d) the number of nearest neighbors to use for the prediction. In total, we compared 30 different CBR classifiers. The study was conducted with a data set from a large real-time system, and the objective was to predict the fault-proneness of its components. Our results indicate that there is no difference in prediction performance when using any combination of parameters. Based on these results, we recommend using a simple CBR classifier with Euclidean distance, z-score standardization, no weighting scheme, and selecting the single nearest neighbor for prediction. The advantage of such a classifier is its intuitive appeal to nonspecialists, and the fact that it performs as well as more complex classifiers.

Using Machine Learning, Neural Networks, and Statistics to Predict Corporate Bankruptcy

Recent literature strongly suggests that machine learning approaches to classification outperform “classical” statistical methods. We make a comparison between the performance of linear discriminant analysis, classification trees, and neural networks in predicting corporate bankruptcy. Linear discriminant analysis represents the “classical” statistical approach to classification, whereas classification trees and neural networks represent artificial intelligence approaches. A proper statistical design is used to be able to test whether observed differences in predictive performance are statistically significant. The data set consists of a collection of 576 annual reports from Belgian construction companies. We use stratified 10–fold cross–validation on the training set to choose “good” parameter values for the different learning methods. The test set is used to obtain an unbiased estimate of the true prediction error. Using rigorous statistical testing, we cannot conclude that in the case of the data set studied, one learning method clearly outperforms the other methods.

Predicting Pure Tone Thresholds in Normal and Hearing-Impaired Ears with Distortion Product Emission and Age

Distortion product emission (DPE) growth functions, demographic data, and pure tone thresholds were recorded in 229 normal-hearing and hearing-impaired ears. Half of the data set (115 ears) was used by a discriminant analysis routine to classify DPE and demographic features into either a normal pure tone threshold (PTT) group or an impaired PTT group (PTT greater than 30 dB SPL) at six frequencies in the audiometric range. The six discriminant functions developed from this classification process were then used to predict pure tone threshold group membership in the remaining 114-ear data set. Frequency-specific prediction accuracy varied from 71% (correct classification of hearing impairment at 1025 Hz) to 92% (correct classification of normal hearing at 2050 Hz). Of the 45 DPE and demographic variables evaluated, the DPE amplitude associated with an f2 of moderate level (50 dB SPL) and a frequency corresponding to pure tone threshold was generally most predictive. DPE level was found to be weakly correlated with subject age and perhaps for this reason age was frequently included in discriminant functions. For our data, discriminant functions with one variable, two variables, or 5-10 variables showed no differences in predictive performance. This research suggests that DPE measures can reliably categorize pure tone thresholds as normal or impaired in large populations with varied cochlear hearing status.

Three‐way methods for the calibration of chromatographic systems: Comparing PARAFAC and three‐way PLS

AbstractFor the calibration of chromatographic systems, different methods can be used. One class of methods utilizes three‐way approaches. The calibration problem is stated in such a way that the decomposition of a three‐way array can serve for the prediction of retention on new stationary phases.Two three‐way approaches are presented: the Unfold‐PCA and PARAFAC models. The theory of both methods is presented and the differences are highlighted, the main difference being that PARAFAC is a trilinear decomposition whereas Unfold‐PCA is not. Both three‐way methods are evaluated on a small data set consisting of retention measurements of eight solutes at six mobile phase compositions on six stationary phases. The differences in performance of the two models are minor.For calibration purposes, two variants of the methods are discussed: three‐way PLS and an extension of PARAFAC. Again the theory and differences between the two methods are explained. The predictive performance of the two methods is compared using the same data set as earlier. The differences in predictive performance, however, are minor. Both methods are capable of predicting 98% of the variation in the test sets. Yet, there are other considerations when comparing methods than predictive performance, e.g. the quality of the predictions.