Iterative Variable Selection Research Articles

Prediction of Early Adverse Events After THA: A Comparison of Different Machine-Learning Strategies Based on 262,356 Observations From the Nordic Arthroplasty Register Association (NARA) Dataset.

Preoperative risk prediction models can support shared decision-making before total hip arthroplasties (THAs). Here, we compare different machine-learning (ML) approaches to predict the six-month risk of adverse events following primary THA to obtain accurate yet simple-to-use risk prediction models. We extracted data on primary THAs (N = 262,356) between 2010 and 2018 from the Nordic Arthroplasty Register Association dataset. We benchmarked a variety of ML algorithms in terms of the area under the receiver operating characteristic curve (AUROC) for predicting the risk of revision caused by periprosthetic joint infection (PJI), dislocation or periprosthetic fracture (PPF), and death. All models were internally validated against a randomly selected test cohort (one-third of the data) that was not used for training the models. The incidences of revisions because of PJI, dislocation, and PPF were 0.8%, 0.4%, and 0.3%, respectively, and the incidence of death was 1.2%. Overall, Lasso regression with stable iterative variable selection (SIVS) produced models using only four to five input variables but with AUROC comparable to more complex models using all 32 variables available. The SIVS-based Lasso models based on age, sex, preoperative diagnosis, bearing couple, fixation, and surgical approach predicted the risk of revisions caused by PJI, dislocations, and PPF, as well as death, with AUROCs of 0.61, 0.67, 0.76, and 0.86, respectively. Our study demonstrates that satisfactory predictive potential for adverse events following THA can be reached with parsimonious modeling strategies. The SIVS-based Lasso models may serve as simple-to-use tools for clinical risk assessment in the future.

Serum hydroxysteroid (17beta) dehydrogenase 1 concentration in pregnant women correlates with pregnancy-associated plasma protein A but does not serve as an independent marker for preeclampsia†.

Hydroxysteroid (17beta) dehydrogenase 1 (HSD17B1) is a steroid synthetic enzyme expressed in ovarian granulosa cells and placental syncytiotrophoblasts. Here, HSD17B1 serum concentration was measured with a validated immunoassay during pregnancy at three time points (12-14, 18-20 and 26-28weeks of gestation). The concentration increased 2.5-fold (P < 0.0001) and 1.7-fold (P = 0.0019) during the follow-up period for control women and women who later developed preeclampsia (PE), respectively, and a significant difference was observed at weeks 26-28 (P = 0.0266). HSD17B1 concentration at all the three time points positively correlated with serum PAPPA measured at the first time point (first time point r = 0.38, P = 1.1 × 10-10; second time point r = 0.27, P = 5.9 × 10-6 and third timepoint r = 0.26, P = 2.3 × 10-5). No correlation was observed between HSD17B1 and placental growth factor (PLGF). Serum HSD17B1 negatively correlated with the mother's weight and body mass index (BMI), mirroring the pattern observed for PAPPA. The univariable logistic regression identified a weak association between HSD17B1 at 26-28weeks and later development of PE (P = 0.04). The best multivariable model obtained using penalized logistic regression with stable iterative variable selection at 26-28weeks included HSD17B1, together with PLGF, PAPPA and mother's BMI. While the area under the receiver operating characteristic curve of the model was higher than that of the adjusted PLGF, the difference was not statistically significant. In summary, the serum concentration of HSD17B1 correlated with PAPPA, another protein expressed in syncytiotrophoblasts, and with mother's weight and BMI but could not be considered as an independent marker for PE.

Development of prediction model for alanine transaminase elevations during the first 6 months of conventional synthetic DMARD treatment

Frequent laboratory monitoring is recommended for early identification of toxicity when initiating conventional synthetic disease-modifying antirheumatic drugs (csDMARDs). We aimed at developing a risk prediction model to individualize laboratory testing at csDMARD initiation. We identified inflammatory joint disease patients (N = 1196) initiating a csDMARD in Turku University Hospital 2013–2019. Baseline and follow-up safety monitoring results were drawn from electronic health records. For rheumatoid arthritis patients, diagnoses and csDMARD initiation/cessation dates were manually confirmed. Primary endpoint was alanine transaminase (ALT) elevation of more than twice the upper limit of normal (ULN) within 6 months after treatment initiation. Computational models for predicting incident ALT elevations were developed using Lasso Cox proportional hazards regression with stable iterative variable selection (SIVS) and were internally validated against a randomly selected test cohort (1/3 of the data) that was not used for training the models. Primary endpoint was reached in 82 patients (6.9%). Among baseline variables, Lasso model with SIVS predicted subsequent ALT elevations of > 2 × ULN using higher ALT, csDMARD other than methotrexate or sulfasalazine and psoriatic arthritis diagnosis as important predictors, with a concordance index of 0.71 in the test cohort. Respectively, at first follow-up, in addition to baseline ALT and psoriatic arthritis diagnosis, also ALT change from baseline was identified as an important predictor resulting in a test concordance index of 0.72. Our computational model predicts ALT elevations after the first follow-up test with good accuracy and can help in optimizing individual testing frequency.

273 (PB053) - Comparison of biomarker selection methods in high-dimensional genomic data

Background: High-dimensional genomic data are indispensable in drug research and development. Genomic data-based biomarker discovery provides comprehensive insights to drug mechanism of action and efficacy prediction. However, it faces the difficulty of selecting suitable signature genes from several thousand or more genes with highly fluctuating and often tightly correlated expression patterns. This study aims to compare the performance of three recent signature gene selection algorithms in several large in vitro cell assays. Methods: Drug response datasets of irinotecan, cetuximab, pelitinib, gefitinib, erlotinib and KRAS (G12C) inhibitor-12 were obtained from the Genomics of Drug Sensitivity in Cancer (GDSC). Cell line gene expression data were gathered from our internal dataset of 20,432 genes, with 1,068 cell lines, across 23 tissues. The three algorithms evaluated were Stable Iterative Variable Selection (SIVS), Precision Lasso (PL) and Whitening Lasso (WL). Efficacy variance between different cancer types (ANOVA p-value) and prediction accuracy of drug efficacy were calculated to evaluate the results. Results: The accuracy of PL predictions was more positively correlated with the number of genes than SIVS and WL. The number of signature genes was limited from 5 to 50 and the most precise results were obtained for all drugs with 30 to 50 genes composite biomarkers. The PL, WL and SIVS methods separately selected 30, 17 and 6 genes as composite biomarkers for irinotecan with an accuracy of 85.7%, 78.2% and 87.3% respectively, in the training set. In validation set, the prediction accuracy of the SIVS panel was the highest at 83.1%. For erlotinib, SIVS, WL and PL picked out 22, 23 and 32 genes as the best composite biomarker with an accuracy of 73%, 83% and 76.8%, respectively. Overall, composite biomarkers selected by SIVS was the superior choice in four of six drugs. In most cases, these three algorithms took less time than 100 iterations of Boruta under the same conditions. Conclusions: Compared with PL and WL, the SIVS method appears to obtain a higher prediction accuracy with a relatively smaller number of genes in this research. PL tended to select many more signature genes than SIVS and WL to yield comparable performance. Overlapping signature genes between the three algorithms was generally low, including for genes belonging to same molecular pathways. The cost and practicality of in vivo experiments would have made it difficult to select a large number of cancer types, drug response, and gene expression data. Therefore, selecting the best predictive composite biomarker in vitro screening for in vivo validation can significantly reduce costs and accelerate the early drug development process. No conflict of interest.

Improving Symbolic Regression for Predicting Materials Properties with Iterative Variable Selection.

Symbolic regression offers a promising avenue for describing the structure-property relationships of materials with explicit mathematical expressions, yet it meets challenges when the key variables are unclear because of the high complexity of the problems. In this work, we propose to solve the difficulty by automatically searching for important variables from a large pool of input features. A new algorithm that integrates symbolic regression with iterative variable selection (VS) was designed for optimization of the model with a large amount of input features. Using the recent method SISSO for symbolic regression and random search for variable selection, we show that the VS-assisted SISSO (VS-SISSO) can effectively manage even hundreds of input features that the SISSO alone was computationally hindered, and it fastly converges to (near) optimal solutions when the model complexity is not high. The efficiency of this approach for improving the accuracy of symbolic regression in materials science was demonstrated in the two showcase applications of learning approximate equations for the band gap of inorganic halide perovskites and the stability of single-atom alloy catalysts.

Predictors of laminitis development in a cohort of nonlaminitic ponies.

Quantifying risk factors for laminitis development requires improvement. To identify the most useful physical examination, metabolic and management factors to predict laminitis development in client-owned, nonlaminitic ponies. Prospective cohort study. Physical examination, metabolic and management data were collected from a pony cohort 6 monthly for up to 4years. Ponies were monitored for the development of laminitis. Metabolic data included basal plasma concentrations of ACTH ([ACTH]), adiponectin ([adiponectin]), triglycerides and glucose. Serum insulin concentrations ([insulin]) were measured in the unfasted basal state ([insulin]T0) and 60minutes ([insulin]T60) after administration of corn syrup (0.3ml/kg). Separate multivariable Cox proportional-hazards models were developed for physical, management/signalment and metabolic data and later combined into two final models. Low-, medium- and high-laminitis risk categories were defined based on basal or T60 [insulin]. Overall, 374 ponies (age 5-32years) and 891 pony-years were included in the main analysis. Laminitis incidence (95% confidence interval (CI)) was 4.8 (3.5-6.5) cases/100 pony-years. Laminitis development was associated with numerous univariable factors. Significant (P<.05) factors retained in the final multivariable models included [insulin]T0, [insulin]T60, [adiponectin] and divergent hoof growth. [ACTH] was not independently associated with laminitis. Based on [Insulin]T0, low- (<21.6 µIU/ml), medium- (21.6-45.2 µIU/ml) and high-risk (>45.2µIU/ml) categories encompassed 70, 20 and 10% of the population and had estimated 4-year laminitis incidences (95%CI) of 6 (2-9)%, 22 (10-33)% and 69 (48-82)% respectively. Based on [Insulin]T60 the low- (<53.4µIU/ml), medium- (53.4-153µIU/ml) and high-risk (≥153µIU/ml) categories comprised 60, 30 and 10% of the population and had estimated 4-year laminitis incidences (95%CI) of 3 (0-6)%, 20 (10-29)% and 73 (52-84)% respectively. Results may not apply to different insulin assays, geographical regions, breeds or management types. [Insulin]T0 or [insulin]T60 best quantify the risk of future laminitis development in nonlaminitic ponies.

Stable Iterative Variable Selection.

MotivationThe emergence of datasets with tens of thousands of features, such as high-throughput omics biomedical data, highlights the importance of reducing the feature space into a distilled subset that can truly capture the signal for research and industry by aiding in finding more effective biomarkers for the question in hand. A good feature set also facilitates building robust predictive models with improved interpretability and convergence of the applied method due to the smaller feature space.ResultsHere, we present a robust feature selection method named Stable Iterative Variable Selection (SIVS) and assess its performance over both omics and clinical data types. As a performance assessment metric, we compared the number and goodness of the selected feature using SIVS to those selected by Least Absolute Shrinkage and Selection Operator regression. The results suggested that the feature space selected by SIVS was, on average, 41% smaller, without having a negative effect on the model performance. A similar result was observed for comparison with Boruta and caret RFE.Availability and implementationThe method is implemented as an R package under GNU General Public License v3.0 and is accessible via Comprehensive R Archive Network (CRAN) via https://cran.r-project.org/package=sivs.Supplementary information Supplementary data are available at Bioinformatics online.

Cross-national analysis of food security drivers: comparing results based on the Food Insecurity Experience Scale and Global Food Security Index

The second UN Sustainable Development Goal establishes food security as a priority for governments, multilateral organizations, and NGOs. These institutions track national-level food security performance with an array of metrics and weigh intervention options considering the leverage of many possible drivers. We studied the relationships between several candidate drivers and two response variables based on prominent measures of national food security: the 2019 Global Food Security Index (GFSI) and the Food Insecurity Experience Scale’s (FIES) estimate of the percentage of a nation’s population experiencing food security or mild food insecurity (FI<mod). We compared the contributions of explanatory variables in regressions predicting both response variables, and we further tested the stability of our results to changes in explanatory variable selection and in the countries included in regression model training and testing. At the cross-national level, the quantity and quality of a nation’s agricultural land were not predictive of either food security metric. We found mixed evidence that per-capita cereal production, per-hectare cereal yield, an aggregate governance metric, logistics performance, and extent of paid employment work were predictive of national food security. Household spending as measured by per-capita final consumption expenditure (HFCE) was consistently the strongest driver among those studied, alone explaining a median of 92% and 70% of variation (based on out-of-sample R2) in GFSI and FI<mod, respectively. The relative strength of HFCE as a predictor was observed for both response variables and was independent of the countries used for model training, the transformations applied to the explanatory variables prior to model training, and the variable selection technique used to specify multivariate regressions. The results of this cross-national analysis reinforce previous research supportive of a causal mechanism where, in the absence of exceptional local factors, an increase in income drives increase in food security. However, the strength of this effect varies depending on the countries included in regression model fitting. We demonstrate that using multiple response metrics, repeated random sampling of input data, and iterative variable selection facilitates a convergence of evidence approach to analyzing food security drivers.

Iterative Variable Selection for High-Dimensional Data: Prediction of Pathological Response in Triple-Negative Breast Cancer

Over the last decade, regularized regression methods have offered alternatives for performing multi-marker analysis and feature selection in a whole genome context. The process of defining a list of genes that will characterize an expression profile remains unclear. It currently relies upon advanced statistics and can use an agnostic point of view or include some a priori knowledge, but overfitting remains a problem. This paper introduces a methodology to deal with the variable selection and model estimation problems in the high-dimensional set-up, which can be particularly useful in the whole genome context. Results are validated using simulated data and a real dataset from a triple-negative breast cancer study.

Development and Evaluation of a Machine Learning-Based In-Hospital COVID-19 Disease Outcome Predictor (CODOP): A Multicontinental Retrospective Study

Background: More contagious SARS-CoV-2 virus variants, breakthrough infections, waning immunity, and differential access to COVID-19 vaccines account for the worst yet numbers of hospitalization and deaths during the COVID-19 pandemic, particularly in resource-limited countries. There is an urgent need for clinically valuable, generalizable, and parsimonious triage tools assisting appropriate allocation of hospital resources during the pandemic. We aimed to develop and extensively validate a machine learning-based tool for accurately predicting the clinical outcome of hospitalized COVID-19 patients.Methods: CODOP was built using modified stable iterative variable selection and linear regression with lasso regularisation. To avoid generalization problems, CODOP was trained and tested with three time-sliced and geographically distinct cohorts encompassing 40 511 blood-based analyses of COVID-19 patients from more than 110 hospitals in Spain and the USA during 2020-21. We assessed the discriminative ability of the model using the Area Under the Receiving Operative Curve (AUROC) as well as horizon and Kaplan-Meier risk stratification analyses. To reckon the fluctuating pressure levels in hospitals through the pandemic, we offer two online CODOP calculators suited for undertriage or overtriage scenarios. We challenged their generalizability and clinical utility throughout an evaluation with datasets gathered in five hospitals from three Latin American countries. Findings: CODOP uses 12 clinical parameters commonly measured at hospital admission and associated with the pathophysiology of COVID-19. CODOP reaches high discriminative ability up to nine days before clinical resolution (AUROC: 0·90-0·96, 95% CI 0·879-0·970), it is well calibrated, and it enables an effective dynamic risk stratification during hospitalization. The two CODOP online calculators predicted the clinical outcome of the majority of patients (73-100% sensitivity and 84-100% specificity) from the distinctive Latin American evaluation cohort.Interpretation: The high predictive performance of CODOP in geographically disperse patient cohorts and the easiness-of-use, strongly suggest its clinical utility as a global triage tool, particularly in resource-limited countries.Funding: The Max Planck Society.Declaration of Interest: The authors declare no conflict of interest.Ethical Approval: This study was approved by the Provincial Research Ethics Committee of Málaga (Spain) and the Institutional Research Ethics Committees of each participating hospital.

Multivariate optimization of headspace solid-phase microextraction coupled to gas chromatography-mass spectrometry for the analysis of terpenoids in sparkling wines

A headspace solid-phase microextraction coupled to gas-chromatography and mass spectrometry (HS-SPME-GC-MS) was developed and validated for the determination of 26 terpenes in sparkling wines. The use of a Box-Behnken experimental design, together with the desirability function D, allowed the extraction conditions of the method to be optimized. The optimal extraction conditions were found at a dilution ratio of 2:3, the addition of 3.5 g of NaCl, an extraction temperature of 46 °C and an extraction time of 52 min, using the DVB/CAR/PDMS fibre. Afterwards, the analytical method was successfully validated in terms of linearity, matrix effect, limit of detection and quantification, precision and accuracy. To test the developed method, 35 commercial sparkling wines from different grape varieties, geographical regions and ageing times were analysed and their terpenoid profile was monitored. The use of multivariate statistical tools made it possible to highlight differences in the samples related to the terpene profile. Finally, the most important compounds involved in the discrimination of the samples were identified by means of iterative variable selection procedures.

A novel variable selection algorithm for multi-layer perceptron with elastic net

The paper proposed a new iterative variable selection algorithm for multi-layer perceptron (MLP) to predict significant variables in complex industrial processes. The proposed algorithm starts from training an ordinary MLP with the original dataset. Following that, the elastic net (EN) algorithm is applied to perform the input weights shrinkage for the well-trained MLP, in which the optimal shrinkage parameters are calculated by grid search and cross-validation. On this basis, a local search strategy is combined to further optimize the input variable selection. The irrelative input variables will be deleted and a new dataset is constructed with remaining variables. This process is repeated until the prescribed termination rules are met. Benchmark datasets with different size and correlation and a practical air separation process are applied and comparisons with other algorithms are made to illustrate its performance. Statistical results show that the proposed algorithm can build more accurate model with fewer variables than other algorithms.

Metabolomic Profile Predicts Development of Microalbuminuria in Individuals with Type 1 Diabetes

Elevated urinary albumin excretion (microalbuminuria) is an early marker of diabetic nephropathy, but there is an unmet need for better biomarkers that capture the individuals at risk with higher accuracy and earlier than the current markers do. We performed an untargeted metabolomic study to assess baseline differences between individuals with type 1 diabetes who either developed microalbuminuria or remained normoalbuminuric. A total of 102 individuals progressed to microalbuminuria during a median follow-up of 3.2 years, whereas 98 sex-, age- and body mass index (BMI) matched non-progressors remained normoalbuminuric during a median follow-up of 7.1 years. Metabolomic screening identified 1,242 metabolites, out of which 111 differed significantly between progressors and non-progressors after adjustment for age of diabetes onset, baseline glycosylated hemoglobin A1c (HbA1c), and albumin excretion rate (AER). The metabolites that predicted development of microalbumiuria included several uremic toxins and carnitine metabolism related molecules. Iterative variable selection indicated erythritol, 3-phenylpropionate, and N-trimethyl-5-aminovalerate as the best set of variables to predict development of microalbuminuria. A metabolomic index based on these metabolites improved the prediction of incident microalbuminuria on top of the clinical variables age of diabetes onset, baseline HbA1c and AER (ROCAUC = 0.842 vs 0.797), highlighting their ability to predict early-phase diabetic nephropathy.

GWASinlps: non-local prior based iterative SNP selection tool for genome-wide association studies.

Multiple marker analysis of the genome-wide association study (GWAS) data has gained ample attention in recent years. However, because of the ultra high-dimensionality of GWAS data, such analysis is challenging. Frequently used penalized regression methods often lead to large number of false positives, whereas Bayesian methods are computationally very expensive. Motivated to ameliorate these issues simultaneously, we consider the novel approach of using non-local priors in an iterative variable selection framework. We develop a variable selection method, named, iterative non-local prior based selection for GWAS, or GWASinlps, that combines, in an iterative variable selection framework, the computational efficiency of the screen-and-select approach based on some association learning and the parsimonious uncertainty quantification provided by the use of non-local priors. The hallmark of our method is the introduction of 'structured screen-and-select' strategy, that considers hierarchical screening, which is not only based on response-predictor associations, but also based on response-response associations and concatenates variable selection within that hierarchy. Extensive simulation studies with single nucleotide polymorphisms having realistic linkage disequilibrium structures demonstrate the advantages of our computationally efficient method compared to several frequentist and Bayesian variable selection methods, in terms of true positive rate, false discovery rate, mean squared error and effect size estimation error. Further, we provide empirical power analysis useful for study design. Finally, a real GWAS data application was considered with human height as phenotype. An R-package for implementing the GWASinlps method is available at https://cran.r-project.org/web/packages/GWASinlps/index.html. Supplementary data are available at Bioinformatics online.

SELEÇÃO DE VARIÁVEIS PARA CATEGORIZAÇÃO DE AMOSTRAS QUÍMICAS

This paper presents a method to select the best variables to categorize chemical samples into two classes, say conforming or non-conforming. For that matter, PLS regression is combined with a data mining tool, the k-Nearest Neighbor classification technique, through an iterative variable selection process. The recommended subset of variables is chosen based on several criteria: sensitivity, specificity and percent of retained variables. When applied to two datasets related to wine analysis and one associated to QSAR, the proposed method significantly reduced the number of variables required for classification, while yielding superior categorization performance when compared to using all original variables.

A new framework for prediction and variable selection for uncommon events in a large prospective cohort study.

When prediction is a goal, validation utilizing data outside of the prediction effort is desirable. Typically, data is split into two parts: one for a development and one for validation. But this approach becomes less attractive when predicting uncommon events, as it substantially reduces power. When predicting uncommon events within a large prospective cohort study, we propose the use of a nested case-control design, which is an alternative to the full cohort analysis. By including all cases but only a subset of the non-cases, this design is expected to produce a result similar to the full cohort analysis. In our framework, variable selection is conducted and a prediction model is fit on those selected variables in the case-control cohort. Then, the fraction of true negative predictions (specificity) of the fitted prediction model in the case-control cohort is compared to that in the rest of the cohort (non-cases) for validation. In addition, we propose an iterative variable selection using random forest for missing data imputation, as well as a strategy for a valid classification. Our framework is illustrated with an application featuring high-dimensional variable selection in a large prospective cohort study.

Look before you leap: a new approach to mapping QTL

In this paper, we present an innovative and powerful approach for mapping quantitative trait loci (QTL) in experimental populations. This deviates from the traditional approach of (composite) interval mapping which uses a QTL profile to simultaneously determine the number and location of QTL. Instead, we look before we leap by employing separate detection and localization stages. In the detection stage, we use an iterative variable selection process coupled with permutation to identify the number and synteny of QTL. In the localization stage, we position the detected QTL through a series of one-dimensional interval mapping scans. Results from a detailed simulation study and real analysis of wheat data are presented. We achieve impressive increases in the power of QTL detection compared to composite interval mapping. We also accurately estimate the size and position of QTL. An R library, DLMap, implements the methods described here and is freely available from CRAN ( http://cran.r-project.org/ ).