Measuring the bias of incorrect application of feature selection when using cross-validation in radiomics.
Ontology highlight
ABSTRACT: Background
Many studies in radiomics are using feature selection methods to identify the most predictive features. At the same time, they employ cross-validation to estimate the performance of the developed models. However, if the feature selection is performed before the cross-validation, data leakage can occur, and the results can be biased. To measure the extent of this bias, we collected ten publicly available radiomics datasets and conducted two experiments. First, the models were developed by incorrectly applying the feature selection prior to cross-validation. Then, the same experiment was conducted by applying feature selection correctly within cross-validation to each fold. The resulting models were then evaluated against each other in terms of AUC-ROC, AUC-F1, and Accuracy.Results
Applying the feature selection incorrectly prior to the cross-validation showed a bias of up to 0.15 in AUC-ROC, 0.29 in AUC-F1, and 0.17 in Accuracy.Conclusions
Incorrect application of feature selection and cross-validation can lead to highly biased results for radiomic datasets.
SUBMITTER: Demircioglu A
PROVIDER: S-EPMC8613324 | biostudies-literature |
REPOSITORIES: biostudies-literature
Similar Datasets
Project description:Introduction: Imaging of tumors is a standard step in diagnosing cancer and making subsequent treatment decisions. The field of radiomics aims to develop imaging based biomarkers using methods rooted in artificial intelligence applied to medical imaging. However, a challenging aspect of developing predictive models for clinical use is that many quantitative features derived from image data exhibit instability or lack of reproducibility across different imaging systems or image-processing pipelines. Methods: To address this challenge, we propose a Bayesian sparse modeling approach for image classification based on radiomic features, where the inclusion of more reliable features is favored via a probit prior formulation. Results: We verify through simulation studies that this approach can improve feature selection and prediction given correct prior information. Finally, we illustrate the method with an application to the classification of head and neck cancer patients by human papillomavirus status, using as our prior information a reliability metric quantifying feature stability across different imaging systems.
| S-EPMC10030957 | biostudies-literature
Project description:Sparse and robust classification models have the potential for revealing common predictive patterns that not only allow for categorizing objects into classes but also for generating mechanistic hypotheses. Identifying a small and informative subset of features is their main ingredient. However, the exponential search space of feature subsets and the heuristic nature of selection algorithms limit the coverage of these analyses, even for low-dimensional datasets. We present methods for reducing the computational complexity of feature selection criteria allowing for higher efficiency and coverage of screenings. We achieve this by reducing the preparation costs of high-dimensional subsets [Formula: see text] to those of one-dimensional ones [Formula: see text]. Our methods are based on a tight interaction between a parallelizable cross-validation traversal strategy and distance-based classification algorithms and can be used with any product distance or kernel. We evaluate the traversal strategy exemplarily in exhaustive feature subset selection experiments (perfect coverage). Its runtime, fitness landscape, and predictive performance are analyzed on publicly available datasets. Even in low-dimensional settings, we achieve approximately a 15-fold increase in exhaustively generating distance matrices for feature combinations bringing a new level of evaluations into reach.
| S-EPMC9744898 | biostudies-literature
Project description:PurposeRadiomics are quantitative features extracted from medical images. Many radiomic features depend not only on tumor properties, but also on non-tumor related factors such as scanner signal-to-noise ratio (SNR), reconstruction kernel and other image acquisition settings. This causes undesirable value variations in the features and reduces the performance of prediction models. In this paper, we investigate whether we can use phantom measurements to characterize and correct for the scanner SNR dependence.MethodsWe used a phantom with 17 regions of interest (ROI) to investigate the influence of different SNR values. CT scans were acquired with 9 different exposure settings. We developed an additive correction model to reduce scanner SNR influence.ResultsSixty-two of 92 radiomic features showed high variance due to the scanner SNR. Of these 62 features, 47 showed at least a factor 2 significant standard deviation reduction by using the additive correction model. We assessed the clinical relevance of radiomics instability by using a 221 NSCLC patient cohort measured with the same scanner.ConclusionsPhantom measurements show that roughly two third of the radiomic features depend on the exposure setting of the scanner. The dependence can be modeled and corrected significantly reducing the variation in feature values with at least a factor of 2. More complex models will likely increase the correctability. Scanner SNR correction will result in more reliable radiomics predictions in NSCLC.
| S-EPMC6690665 | biostudies-literature
Project description:Educators seek to harness knowledge from educational corpora to improve student performance outcomes. Although prior studies have compared the efficacy of data mining methods (DMMs) in pipelines for forecasting student success, less work has focused on identifying a set of relevant features prior to model development and quantifying the stability of feature selection techniques. Pinpointing a subset of pertinent features can (1) reduce the number of variables that need to be managed by stakeholders, (2) make "black-box" algorithms more interpretable, and (3) provide greater guidance for faculty to implement targeted interventions. To that end, we introduce a methodology integrating feature selection with cross-validation and rank each feature on subsets of the training corpus. This modified pipeline was applied to forecast the performance of 3225 students in a baccalaureate science course using a set of 57 features, four DMMs, and four filter feature selection techniques. Correlation Attribute Evaluation (CAE) and Fisher's Scoring Algorithm (FSA) achieved significantly higher Area Under the Curve (AUC) values for logistic regression (LR) and elastic net regression (GLMNET), compared to when this pipeline step was omitted. Relief Attribute Evaluation (RAE) was highly unstable and produced models with the poorest prediction performance. Borda's method identified grade point average, number of credits taken, and performance on concept inventory assessments as the primary factors impacting predictions of student performance. We discuss the benefits of this approach when developing data pipelines for predictive modeling in undergraduate settings that are more interpretable and actionable for faculty and stakeholders.Supplementary informationThe online version contains supplementary material available at 10.1186/s41239-021-00279-6.
| S-EPMC8591701 | biostudies-literature
Project description:Using validation sets for outcomes can greatly improve the estimation of vaccine efficacy (VE) in the field (Halloran and Longini, 2001; Halloran and others, 2003). Most statistical methods for using validation sets rely on the assumption that outcomes on those with no cultures are missing at random (MAR). However, often the validation sets will not be chosen at random. For example, confirmational cultures are often done on people with influenza-like illness as part of routine influenza surveillance. VE estimates based on such non-MAR validation sets could be biased. Here we propose frequentist and Bayesian approaches for estimating VE in the presence of validation bias. Our work builds on the ideas of Rotnitzky and others (1998, 2001), Scharfstein and others (1999, 2003), and Robins and others (2000). Our methods require expert opinion about the nature of the validation selection bias. In a re-analysis of an influenza vaccine study, we found, using the beliefs of a flu expert, that within any plausible range of selection bias the VE estimate based on the validation sets is much higher than the point estimate using just the non-specific case definition. Our approach is generally applicable to studies with missing binary outcomes with categorical covariates.
| S-EPMC2766283 | biostudies-literature
Project description:Predictive modelling of cancer outcomes using radiomics faces dimensionality problems and data limitations, as radiomics features often number in the hundreds, and multi-institutional data sharing is ()often unfeasible. Federated learning (FL) and feature selection (FS) techniques combined can help overcome these issues, as one provides the means of training models without exchanging sensitive data, while the other identifies the most informative features, reduces overfitting, and improves model interpretability. Our proposed FS pipeline based on FL principles targets data-driven radiomics FS in a multivariate survival study of non-small cell lung cancer patients. The pipeline was run across datasets from three institutions without patient-level data exchange. It includes two FS techniques, Correlation-based Feature Selection and LASSO regularization, and Cox Proportional-Hazard regression with Overall Survival as endpoint. Trained and validated on 828 patients overall, our pipeline yielded a radiomic signature comprising "intensity-based energy" and "mean discretised intensity". Validation resulted in a mean Harrell C-index of 0.59, showcasing fair efficacy in risk stratification. In conclusion, we suggest a distributed radiomics approach that incorporates preliminary feature selection to systematically decrease the feature set based on data-driven considerations. This aims to address dimensionality challenges beyond those associated with data constraints and interpretability concerns.
| S-EPMC10991291 | biostudies-literature
Project description:Due to advances in the acquisition and analysis of medical imaging, it is currently possible to quantify the tumor phenotype. The emerging field of Radiomics addresses this issue by converting medical images into minable data by extracting a large number of quantitative imaging features. One of the main challenges of Radiomics is tumor segmentation. Where manual delineation is time consuming and prone to inter-observer variability, it has been shown that semi-automated approaches are fast and reduce inter-observer variability. In this study, a semiautomatic region growing volumetric segmentation algorithm, implemented in the free and publicly available 3D-Slicer platform, was investigated in terms of its robustness for quantitative imaging feature extraction. Fifty-six 3D-radiomic features, quantifying phenotypic differences based on tumor intensity, shape and texture, were extracted from the computed tomography images of twenty lung cancer patients. These radiomic features were derived from the 3D-tumor volumes defined by three independent observers twice using 3D-Slicer, and compared to manual slice-by-slice delineations of five independent physicians in terms of intra-class correlation coefficient (ICC) and feature range. Radiomic features extracted from 3D-Slicer segmentations had significantly higher reproducibility (ICC?=?0.85±0.15, p?=?0.0009) compared to the features extracted from the manual segmentations (ICC?=?0.77±0.17). Furthermore, we found that features extracted from 3D-Slicer segmentations were more robust, as the range was significantly smaller across observers (p?=?3.819e-07), and overlapping with the feature ranges extracted from manual contouring (boundary lower: p?=?0.007, higher: p?=?5.863e-06). Our results show that 3D-Slicer segmented tumor volumes provide a better alternative to the manual delineation for feature quantification, as they yield more reproducible imaging descriptors. Therefore, 3D-Slicer can be employed for quantitative image feature extraction and image data mining research in large patient cohorts.
| S-EPMC4098900 | biostudies-literature
Project description:This paper describes a natural language processing system for the task of pneumonia identification. Based on the information extracted from the narrative reports associated with a patient, the task is to identify whether or not the patient is positive for pneumonia.A binary classifier was employed to identify pneumonia from a dataset of multiple types of clinical notes created for 426 patients during their stay in the intensive care unit. For this purpose, three types of features were considered: (1) word n-grams, (2) Unified Medical Language System (UMLS) concepts, and (3) assertion values associated with pneumonia expressions. System performance was greatly increased by a feature selection approach which uses statistical significance testing to rank features based on their association with the two categories of pneumonia identification.Besides testing our system on the entire cohort of 426 patients (unrestricted dataset), we also used a smaller subset of 236 patients (restricted dataset). The performance of the system was compared with the results of a baseline previously proposed for these two datasets. The best results achieved by the system (85.71 and 81.67 F1-measure) are significantly better than the baseline results (50.70 and 49.10 F1-measure) on the restricted and unrestricted datasets, respectively.Using a statistical feature selection approach that allows the feature extractor to consider only the most informative features from the feature space significantly improves the performance over a baseline that uses all the features from the same feature space. Extracting the assertion value for pneumonia expressions further improves the system performance.
| S-EPMC3422830 | biostudies-other
Project description:Feature selection (FS) can eliminate many redundant, irrelevant, and noisy features in high-dimensional data to improve machine learning or data mining models' prediction, classification, and computational performance. We proposed an improved whale optimization algorithm (IWOA) and improved k-nearest neighbors (IKNN) classifier approaches for feature selection (IWOAIKFS). Firstly, WOA is improved by using chaotic elite reverse individual, probability selection of skew distribution, nonlinear adjustment of control parameters and position correction strategy to enhance the search performance of the algorithm for feature subsets. Secondly, the sample similarity measurement criterion and weighted voting criterion based on the simulated annealing algorithm to solve the weight matrix M are proposed to improve the KNN classifier and improve the evaluation performance of the algorithm on feature subsets. The experimental results show: IWOA not only has better optimization performance when solving benchmark functions of different dimensions, but also when used with IKNN for feature selection, IWOAIKFS has better classification and robustness.
| S-EPMC9119564 | biostudies-literature
Project description:Radiomics involves high-throughput extraction of large numbers of quantitative features from medical images and analysis of these features to predict patients' outcome and support clinical decision-making. However, radiomics features are sensitive to several factors, including scanning protocols. The purpose of this study was to investigate the robustness of magnetic resonance imaging (MRI) radiomics features with various MRI scanning protocol parameters and scanners using an MRI radiomics phantom. The variability of the radiomics features with different scanning parameters and repeatability measured using a test-retest scheme were evaluated using the coefficient of variation and intraclass correlation coefficient (ICC) for both T1- and T2-weighted images. For variability measures, the features were categorized into three groups: large, intermediate, and small variation. For repeatability measures, the average T1- and T2-weighted image ICCs for the phantom (0.963 and 0.959, respectively) were higher than those for a healthy volunteer (0.856 and 0.849, respectively). Our results demonstrated that various radiomics features are dependent on different scanning parameters and scanners. The radiomics features with a low coefficient of variation and high ICC for both the phantom and volunteer can be considered good candidates for MRI radiomics studies. The results of this study will assist current and future MRI radiomics studies.
| S-EPMC7889870 | biostudies-literature