The feature selection bias problem in relation to high-dimensional gene data

An improved binary particle swarm optimization algorithm for clinical cancer biomarker identification in microarray data

BACKGROUND AND OBJECTIVE

The limited number of samples and high-dimensional features in microarray data make selecting a small number of features for disease diagnosis a challenging problem. Traditional feature selection methods based on evolutionary algorithms are difficult to search for the optimal set of features in a limited time when dealing with the high-dimensional feature selection problem. New solutions are proposed to solve the above problems.

METHODS

In this paper, we propose a hybrid feature selection method (C-IFBPFE) for biomarker identification in microarray data, which combines clustering and improved binary particle swarm optimization while incorporating an embedded feature elimination strategy. Firstly, an adaptive redundant feature judgment method based on correlation clustering is proposed for feature screening to reduce the search space in the subsequent stage. Secondly, we propose an improved flipping probability-based binary particle swarm optimization (IFBPSO), better applicable to the binary particle swarm optimization problem. Finally, we also design a new feature elimination (FE) strategy embedded in the binary particle swarm optimization algorithm. This strategy gradually removes poorer features during iterations to reduce the number of features and improve accuracy.

RESULTS

We compared C-IFBPFE with other published hybrid feature selection methods on eight public datasets and analyzed the impact of each improvement. The proposed method outperforms other current state-of-the-art feature selection methods in terms of accuracy, number of features, sensitivity, and specificity. The ablation study of this method validates the efficacy of each component, especially the proposed feature elimination strategy significantly improves the performance of the algorithm.

CONCLUSIONS

The hybrid feature selection method proposed in this paper helps address the issue of high-dimensional microarray data with few samples. It can select a small subset of features and achieve high classification accuracy on microarray datasets. Additionally, independent validation of the selected features shows that those chosen by C-IFBPFE have strong correlations with disease phenotypes and can identify important biomarkers from data related to biomedical problems.

Novel ensemble feature selection techniques applied to high-grade gastroenteropancreatic neuroendocrine neoplasms for the prediction of survival

BACKGROUND AND OBJECTIVE

Determining the most informative features for predicting the overall survival of patients diagnosed with high-grade gastroenteropancreatic neuroendocrine neoplasms is crucial to improve individual treatment plans for patients, as well as the biological understanding of the disease. The main objective of this study is to evaluate the use of modern ensemble feature selection techniques for this purpose with respect to (a) quantitative performance measures such as predictive performance, (b) clinical interpretability, and (c) the effect of integrating prior expert knowledge.

METHODS

The Repeated Elastic Net Technique for Feature Selection (RENT) and the User-Guided Bayesian Framework for Feature Selection (UBayFS) are recently developed ensemble feature selectors investigated in this work. Both allow the user to identify informative features in datasets with low sample sizes and focus on model interpretability. While RENT is purely data-driven, UBayFS can integrate expert knowledge a priori in the feature selection process. In this work, we compare both feature selectors on a dataset comprising 63 patients and 110 features from multiple sources, including baseline patient characteristics, baseline blood values, tumor histology, imaging, and treatment information.

RESULTS

Our experiments involve data-driven and expert-driven setups, as well as combinations of both. In a five-fold cross-validated experiment without expert knowledge, our results demonstrate that both feature selectors allow accurate predictions: A reduction from 110 to approximately 20 features (around 82%) delivers near-optimal predictive performances with minor variations according to the choice of the feature selector, the predictive model, and the fold. Thereafter, we use findings from clinical literature as a source of expert knowledge. In addition, expert knowledge has a stabilizing effect on the feature set (an increase in stability of approximately 40%), while the impact on predictive performance is limited.

CONCLUSIONS

The features WHO Performance Status, Albumin, Platelets, Ki-67, Tumor Morphology, Total MTV, Total TLG, and SUVmax are the most stable and predictive features in our study. Overall, this study demonstrated the practical value of feature selection in medical applications not only to improve quantitative performance but also to deliver potentially new insights to experts.

"Shortcuts" Causing Bias in Radiology Artificial Intelligence: Causes, Evaluation, and Mitigation

Despite the expert-level performance of artificial intelligence (AI) models for various medical imaging tasks, real-world performance failures with disparate outputs for various subgroups limit the usefulness of AI in improving patients' lives. Many definitions of fairness have been proposed, with discussions of various tensions that arise in the choice of an appropriate metric to use to evaluate bias; for example, should one aim for individual or group fairness? One central observation is that AI models apply "shortcut learning" whereby spurious features (such as chest tubes and portable radiographic markers on intensive care unit chest radiography) on medical images are used for prediction instead of identifying true pathology. Moreover, AI has been shown to have a remarkable ability to detect protected attributes of age, sex, and race, while the same models demonstrate bias against historically underserved subgroups of age, sex, and race in disease diagnosis. Therefore, an AI model may take shortcut predictions from these correlations and subsequently generate an outcome that is biased toward certain subgroups even when protected attributes are not explicitly used as inputs into the model. As a result, these subgroups became nonprivileged subgroups. In this review, the authors discuss the various types of bias from shortcut learning that may occur at different phases of AI model development, including data bias, modeling bias, and inference bias. The authors thereafter summarize various tool kits that can be used to evaluate and mitigate bias and note that these have largely been applied to nonmedical domains and require more evaluation for medical AI. The authors then summarize current techniques for mitigating bias from preprocessing (data-centric solutions) and during model development (computational solutions) and postprocessing (recalibration of learning). Ongoing legal changes where the use of a biased model will be penalized highlight the necessity of understanding, detecting, and mitigating biases from shortcut learning and will require diverse research teams looking at the whole AI pipeline.

Using Machine Learning Methods to Study Colorectal Cancer Tumor Micro-Environment and Its Biomarkers

Colorectal cancer (CRC) is a leading cause of cancer deaths worldwide, and the identification of biomarkers can improve early detection and personalized treatment. In this study, RNA-seq data and gene chip data from TCGA and GEO were used to explore potential biomarkers for CRC. The SMOTE method was used to address class imbalance, and four feature selection algorithms (MCFS, Borota, mRMR, and LightGBM) were used to select genes from the gene expression matrix. Four machine learning algorithms (SVM, XGBoost, RF, and kNN) were then employed to obtain the optimal number of genes for model construction. Through interpretable machine learning (IML), co-predictive networks were generated to identify rules and uncover underlying relationships among the selected genes. Survival analysis revealed that INHBA, FNBP1, PDE9A, HIST1H2BG, and CADM3 were significantly correlated with prognosis in CRC patients. In addition, the CIBERSORT algorithm was used to investigate the proportion of immune cells in CRC tissues, and gene mutation rates for the five selected biomarkers were explored. The biomarkers identified in this study have significant implications for the development of personalized therapies and could ultimately lead to improved clinical outcomes for CRC patients.

A Predictive Visual Analytics System for Studying Neurodegenerative Disease Based on DTI Fiber Tracts

Diffusion tensor imaging (DTI) has been used to study the effects of neurodegenerative diseases on neural pathways, which may lead to more reliable and early diagnosis of these diseases as well as a better understanding of how they affect the brain. We introduce a predictive visual analytics system for studying patient groups based on their labeled DTI fiber tract data and corresponding statistics. The system's machine-learning-augmented interface guides the user through an organized and holistic analysis space, including the statistical feature space, the physical space, and the space of patients over different groups. We use a custom machine learning pipeline to help narrow down this large analysis space and then explore it pragmatically through a range of linked visualizations. We conduct several case studies using DTI and T1-weighted images from the research database of Parkinson's Progression Markers Initiative.

Classification of Alzheimer's disease using robust TabNet neural networks on genetic data

Alzheimer's disease (AD) is one of the most common neurodegenerative diseases and its onset is significantly associated with genetic factors. Being the capabilities of high specificity and accuracy, genetic testing has been considered as an important technique for AD diagnosis. In this paper, we presented an improved deep learning (DL) algorithm, namely differential genes screening TabNet (DGS-TabNet) for AD binary and multi-class classifications. For performance evaluation, our proposed approach was compared with three novel DLs of multi-layer perceptron (MLP), neural oblivious decision ensembles (NODE), TabNet as well as five classical machine learnings (MLs) including decision tree (DT), random forests (RF), gradient boosting decision tree (GBDT), light gradient boosting machine (LGBM) and support vector machine (SVM) on the public data set of gene expression omnibus (GEO). Moreover, the biological interpretability of global important genetic features implemented for AD classification was revealed by the Kyoto encyclopedia of genes and genomes (KEGG) and gene ontology (GO). The results demonstrated that our proposed DGS-TabNet achieved the best performance with an accuracy of 93.80% for binary classification, and with an accuracy of 88.27% for multi-class classification. Meanwhile, the gene pathway analyses demonstrated that there existed two most important global genetic features of AVIL and NDUFS4 and those obtained 22 feature genes were partially correlated with AD pathogenesis. It was concluded that the proposed DGS-TabNet could be used to detect AD-susceptible genes and the biological interpretability of susceptible genes also revealed the potential possibility of being AD biomarkers.

Constrained neuro fuzzy inference methodology for explainable personalised modelling with applications on gene expression data

Interpretable machine learning models for gene expression datasets are important for understanding the decision-making process of a classifier and gaining insights on the underlying molecular processes of genetic conditions. Interpretable models can potentially support early diagnosis before full disease manifestation. This is particularly important yet, challenging for mental health. We hypothesise this is due to extreme heterogeneity issues which may be overcome and explained by personalised modelling techniques. Thus far, most machine learning methods applied to gene expression datasets, including deep neural networks, lack personalised interpretability. This paper proposes a new methodology named personalised constrained neuro fuzzy inference (PCNFI) for learning personalised rules from high dimensional datasets which are structurally and semantically interpretable. Case studies on two mental health related datasets (schizophrenia and bipolar disorders) have shown that the relatively short and simple personalised fuzzy rules provided enhanced interpretability as well as better classification performance compared to other commonly used machine learning methods. Performance test on a cancer dataset also showed that PCNFI matches previous benchmarks. Insights from our approach also indicated the importance of two genes (ATRX and TSPAN2) as possible biomarkers for early differentiation of ultra-high risk, bipolar and healthy individuals. These genes are linked to cognitive ability and impulsive behaviour. Our findings suggest a significant starting point for further research into the biological role of cognitive and impulsivity-related differences. With potential applications across bio-medical research, the proposed PCNFI method is promising for diagnosis, prognosis, and the design of personalised treatment plans for better outcomes in the future.

Application of artificial intelligence in medical technologies: A systematic review of main trends

Objective

Artificial intelligence (AI) has been increasingly applied in various fields of science and technology. In line with the current research, medicine involves an increasing number of artificial intelligence technologies. The introduction of rapid AI can lead to positive and negative effects. This is a multilateral analytical literature review aimed at identifying the main branches and trends in the use of using artificial intelligence in medical technologies.

Methods

The total number of literature sources reviewed is n = 89, and they are analyzed based on the literature reporting evidence-based guideline PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) for a systematic review.

Results

As a result, from the initially selected 198 references, 155 references were obtained from the databases and the remaining 43 sources were found on open internet as direct links to publications. Finally, 89 literature sources were evaluated after exclusion of unsuitable references based on the duplicated and generalized information without focusing on the users.

Conclusions

This article is identifying the current state of artificial intelligence in medicine and prospects for future use. The findings of this review will be useful for healthcare and AI professionals for improving the circulation and use of medical AI from design to implementation stage.

AltWOA: Altruistic Whale Optimization Algorithm for feature selection on microarray datasets

The data-driven modern era has enabled the collection of large amounts of biomedical and clinical data. DNA microarray gene expression datasets have mainly gained significant attention to the research community owing to their ability to identify diseases through the "bio-markers" or specific alterations in the gene sequence that represent that particular disease (for example, different types of cancer). However, gene expression datasets are very high-dimensional, while only a few of those are "bio-markers". Meta-heuristic-based feature selection effectively filters out only the relevant genes from a large set of attributes efficiently to reduce data storage and computation requirements. To this end, in this paper, we propose an Altruistic Whale Optimization Algorithm (AltWOA) for the feature selection problem in high-dimensional microarray data. AltWOA is an improvement on the basic Whale Optimization Algorithm. We embed the concept of altruism in the whale population to help efficient propagation of candidate solutions that can reach the global optima over the iterations. Evaluation of the proposed method on eight high dimensional microarray datasets reveals the superiority of AltWOA compared to popular and classical techniques in the literature on the same datasets both in terms of accuracy and the final number of features selected. The relevant codes for the proposed approach are available publicly at https://github.com/Rohit-Kundu/AltWOA.

Gene selection for microarray data classification via multi-objective graph theoretic-based method

In recent decades, the improvement of computer technology has increased the growth of high-dimensional microarray data. Thus, data mining methods for DNA microarray data classification usually involve samples consisting of thousands of genes. One of the efficient strategies to solve this problem is gene selection, which improves the accuracy of microarray data classification and also decreases computational complexity. In this paper, a novel social network analysis-based gene selection approach is proposed. The proposed method has two main objectives of the relevance maximization and redundancy minimization of the selected genes. In this method, on each iteration, a maximum community is selected repetitively. Then among the existing genes in this community, the appropriate genes are selected by using the node centrality-based criterion. The reported results indicate that the developed gene selection algorithm while increasing the classification accuracy of microarray data, will also decrease the time complexity.

Deep Learning Recurrent Neural Network for Concussion Classification in Adolescents Using Raw Electroencephalography Signals: Toward a Minimal Number of Sensors

Artificial neural networks (ANNs) are showing increasing promise as decision support tools in medicine and particularly in neuroscience and neuroimaging. Recently, there has been increasing work on using neural networks to classify individuals with concussion using electroencephalography (EEG) data. However, to date the need for research grade equipment has limited the applications to clinical environments. We recently developed a deep learning long short-term memory (LSTM) based recurrent neural network to classify concussion using raw, resting state data using 64 EEG channels and achieved high accuracy in classifying concussion. Here, we report on our efforts to develop a clinically practical system using a minimal subset of EEG sensors. EEG data from 23 athletes who had suffered a sport-related concussion and 35 non-concussed, control athletes were used for this study. We tested and ranked each of the original 64 channels based on its contribution toward the concussion classification performed by the original LSTM network. The top scoring channels were used to train and test a network with the same architecture as the previously trained network. We found that with only six of the top scoring channels the classifier identified concussions with an accuracy of 94%. These results show that it is possible to classify concussion using raw, resting state data from a small number of EEG sensors, constituting a first step toward developing portable, easy to use EEG systems that can be used in a clinical setting.

fMRI-Informed EEG for brain mapping of imagined lower limb movement: Feasibility of a brain computer interface

BACKGROUND

EEG and fMRI have contributed greatly to our understanding of brain activity and its link to behaviors by helping to identify both when and where the activity occurs. This is particularly important in the development of brain-computer interfaces (BCIs), where feed forward systems gather data from imagined brain activity and then send that information to an effector. The purpose of this study was to develop and evaluate a computational approach that enables an accurate mapping of spatial brain activity (fMRI) in relation to the temporal receptors (EEG electrodes) associated with imagined lower limb movement.

NEW METHOD

EEG and fMRI data from 16 healthy, male participants while imagining lower limb movement were used for this purpose. A combined analysis of fMRI data and EEG electrode locations was developed to identify EEG electrodes with a high likelihood of capturing imagined lower limb movement originating from various clusters of brain activity. This novel feature selection tool was used to develop an artificial neural network model to classify right and left lower limb movement.

RESULTS

Results showed that left versus right lower limb imagined movement could be classified with 66.5% accuracy using this approach. Comparison with existing methods: Adopting a purely data-driven approach for feature selection to use in the right/left classification task resulted in the same accuracy (66.6%) but with reduced interpretability.

CONCLUSIONS

The developed fMRI-informed EEG approach could pave the way towards improved brain computer interfaces for lower limb movement while also being applicable to other systems where fMRI could be helpful to inform EEG acquisition and processing.

DeepCOMBI: explainable artificial intelligence for the analysis and discovery in genome-wide association studies

Deep learning has revolutionized data science in many fields by greatly improving prediction performances in comparison to conventional approaches. Recently, explainable artificial intelligence has emerged as an area of research that goes beyond pure prediction improvement by extracting knowledge from deep learning methodologies through the interpretation of their results. We investigate such explanations to explore the genetic architectures of phenotypes in genome-wide association studies. Instead of testing each position in the genome individually, the novel three-step algorithm, called DeepCOMBI, first trains a neural network for the classification of subjects into their respective phenotypes. Second, it explains the classifiers’ decisions by applying layer-wise relevance propagation as one example from the pool of explanation techniques. The resulting importance scores are eventually used to determine a subset of the most relevant locations for multiple hypothesis testing in the third step. The performance of DeepCOMBI in terms of power and precision is investigated on generated datasets and a 2007 study. Verification of the latter is achieved by validating all findings with independent studies published up until 2020. DeepCOMBI is shown to outperform ordinary raw P-value thresholding and other baseline methods. Two novel disease associations (rs10889923 for hypertension, rs4769283 for type 1 diabetes) were identified.

Combing machine learning and elemental profiling for geographical authentication of Chinese Geographical Indication (GI) rice

Identification of geographical origin is of great importance for protecting the authenticity of valuable agri-food products with designated origins. In this study, a robust and accurate analytical method that could authenticate the geographical origin of Geographical Indication (GI) products was developed. The method was based on elemental profiling using inductively coupled plasma mass spectrometry (ICP-MS) in combination with machine learning techniques for model building and feature selection. The method successfully predicted and classified six varieties of Chinese GI rice. The elemental profiles of 131 rice samples were determined, and two machine learning algorithms were implemented, support vector machines (SVM) and random forest (RF), together with the feature selection algorithm Relief. Prediction accuracy of 100% was achieved by both Relief-SVM and Relief-RF models, using only four elements (Al, B, Rb, and Na). The methodology and knowledge from this study could be used to develop reliable methods for tracing geographical origins and controlling fraudulent labeling of diverse high-value agri-food products.

Prediction of venous thromboembolism with machine learning techniques in young-middle-aged inpatients

Accumulating studies appear to suggest that the risk factors for venous thromboembolism (VTE) among young-middle-aged inpatients are different from those among elderly people. Therefore, the current prediction models for VTE are not applicable to young-middle-aged inpatients. The aim of this study was to develop and externally validate a new prediction model for young-middle-aged people using machine learning methods. The clinical data sets linked with 167 inpatients with deep venous thrombosis (DVT) and/or pulmonary embolism (PE) and 406 patients without DVT or PE were compared and analysed with machine learning techniques. Five algorithms, including logistic regression, decision tree, feed-forward neural network, support vector machine, and random forest, were used for training and preparing the models. The support vector machine model had the best performance, with AUC values of 0.806-0.944 for 95% CI, 59% sensitivity and 99% specificity, and an accuracy of 87%. Although different top predictors of adverse outcomes appeared in the different models, life-threatening illness, fibrinogen, RBCs, and PT appeared to be more consistently featured by the different models as top predictors of adverse outcomes. Clinical data sets of young and middle-aged inpatients can be used to accurately predict the risk of VTE with a support vector machine model.

Recurrent neural network-based acute concussion classifier using raw resting state EEG data

Concussion is a global health concern. Despite its high prevalence, a sound understanding of the mechanisms underlying this type of diffuse brain injury remains elusive. It is, however, well established that concussions cause significant functional deficits; that children and youths are disproportionately affected and have longer recovery time than adults; and that individuals suffering from a concussion are more prone to experience additional concussions, with each successive injury increasing the risk of long term neurological and mental health complications. Currently, the most significant challenge in concussion management is the lack of objective, clinically- accepted, brain-based approaches for determining whether an athlete has suffered a concussion. Here, we report on our efforts to address this challenge. Specifically, we introduce a deep learning long short-term memory (LSTM)-based recurrent neural network that is able to distinguish between non-concussed and acute post-concussed adolescent athletes using only short (i.e. 90 s long) samples of resting state EEG data as input. The athletes were neither required to perform a specific task nor expected to respond to a stimulus during data collection. The acquired EEG data were neither filtered, cleaned of artefacts, nor subjected to explicit feature extraction. The LSTM network was trained and validated using data from 27 male, adolescent athletes with sports related concussion, benchmarked against 35 non-concussed adolescent athletes. During rigorous testing, the classifier consistently identified concussions with an accuracy of > 90% and achieved an ensemble median Area Under the Receiver Operating Characteristic Curve (ROC/AUC) equal to 0.971. This is the first instance of a high-performing classifier that relies only on easy-to-acquire resting state, raw EEG data. Our concussion classifier represents a promising first step towards the development of an easy-to-use, objective, brain-based, automatic classification of concussion at an individual level.

Interpretable Machine Learning Reveals Dissimilarities Between Subtypes of Autism Spectrum Disorder

Autism spectrum disorder (ASD) is a heterogeneous neuropsychiatric disorder with a complex genetic background. Analysis of altered molecular processes in ASD patients requires linear and nonlinear methods that provide interpretable solutions. Interpretable machine learning provides legible models that allow explaining biological mechanisms and support analysis of clinical subgroups. In this work, we investigated several case-control studies of gene expression measurements of ASD individuals. We constructed a rule-based learning model from three independent datasets that we further visualized as a nonlinear gene-gene co-predictive network. To find dissimilarities between ASD subtypes, we scrutinized a topological structure of the network and estimated a centrality distance. Our analysis revealed that autism is the most severe subtype of ASD, while pervasive developmental disorder-not otherwise specified and Asperger syndrome are closely related and milder ASD subtypes. Furthermore, we analyzed the most important ASD-related features that were described in terms of gene co-predictors. Among others, we found a strong co-predictive mechanism between EMC4 and TMEM30A, which may suggest a co-regulation between these genes. The present study demonstrates the potential of applying interpretable machine learning in bioinformatics analyses. Although the proposed methodology was designed for transcriptomics data, it can be applied to other omics disciplines.

Item response theory as a feature selection and interpretation tool in the context of machine learning

Optimizing the number and utility of features to use in a classification analysis has been the subject of many research studies. Most current models use end-classifications as part of the feature reduction process, leading to circularity in the methodology. The approach demonstrated in the present research uses item response theory (IRT) to select features independent of the end-classification results without the biased accuracies that this circularity engenders. Dichotomous and polytomous IRT models were used to analyze 30 histological breast cancer features from 569 patients using the Wisconsin Diagnostic Breast Cancer data set. Based on their characteristics, three features were selected for use in a machine learning classifier. For comparison purposes, two machine learning-based feature selection protocols were run-recursive feature elimination (RFE) and ridge regression-and the three features selected from these analyses were also used in the subsequent learning classifier. Classification results demonstrated that all three selection processes performed comparably. The non-biased nature of the IRT protocol and information provided about the specific characteristics of the features as to why they are of use in classification help to shed light on understanding which attributes of features make them suitable for use in a machine learning context.

Multiobjective Particle Swarm Optimization for Feature Selection With Fuzzy Cost

Feature selection (FS) is an important data processing technique in the field of machine learning. There have been various FS methods, but all assume that the cost associated with a feature is precise, which restricts their real applications. Focusing on the FS problem with fuzzy cost, a fuzzy multiobjective FS method with particle swarm optimization, called PSOMOFS, is studied in this article. The proposed method develops a fuzzy dominance relationship to compare the goodness of candidate particles and defines a fuzzy crowding distance measure to prune the elitist archive and determine the global leader of particles. Also, a tolerance coefficient is introduced into the proposed method to ensure that the Pareto-optimal solutions obtained satisfy decision makers' preferences. The developed method is used to tackle a series of the UCI datasets and is compared with three fuzzy multiobjective evolutionary methods and three typical multiobjective FS methods. Experimental results show that the proposed method can achieve feature sets with superior performances in approximation, diversity, and feature cost.

Novel Feature Selection for Artificial Intelligence Using Item Response Theory for Mortality Prediction

Feature selection is a critical component in supervised machine learning classification analyses. Extraneous features introduce noise and inefficiencies into the system leading to a need for feature reduction techniques. Many feature reduction models use the end-classification results in the feature reduction process, committing a circular error. Item Response Theory (IRT) examines the characteristics of features independent of the end-classification results, and provides high levels of information regarding feature utility. A two-parameter dichotomous IRT model was used to analyze 18 features from an intensive care unit data set with 2520 cases. The classification results showed that the features selected via IRT were comparable to that using more traditional machine learning approaches. Strengths and limitations of the IRT selection protocol are discussed.

Machine learning for radiomics-based multimodality and multiparametric modeling

Due to the recent developments of both hardware and software technologies, multimodality medical imaging techniques have been increasingly applied in clinical practice and research studies. Previously, the application of multimodality imaging in oncology has been mainly related to combining anatomical and functional imaging to improve diagnostic specificity and/or target definition, such as positron emission tomography/computed tomography (PET/CT) and single-photon emission CT (SPECT)/CT. More recently, the fusion of various images, such as multiparametric magnetic resonance imaging (MRI) sequences, different PET tracer images, PET/MRI, has become more prevalent, which has enabled more comprehensive characterization of the tumor phenotype. In order to take advantage of these valuable multimodal data for clinical decision making using radiomics, we present two ways to implement the multimodal image analysis, namely radiomic (handcrafted feature) based and deep learning (machine learned feature) based methods. Applying advanced machine (deep) learning algorithms across multimodality images have shown better results compared with single modality modeling for prognostic and/or prediction of clinical outcomes. This holds great potentials for providing more personalized treatment for patients and achieve better outcomes.

DeepCC: a novel deep learning-based framework for cancer molecular subtype classification

Molecular subtyping of cancer is a critical step towards more individualized therapy and provides important biological insights into cancer heterogeneity. Although gene expression signature-based classification has been widely demonstrated to be an effective approach in the last decade, the widespread implementation has long been limited by platform differences, batch effects, and the difficulty to classify individual patient samples. Here, we describe a novel supervised cancer classification framework, deep cancer subtype classification (DeepCC), based on deep learning of functional spectra quantifying activities of biological pathways. In two case studies about colorectal and breast cancer classification, DeepCC classifiers and DeepCC single sample predictors both achieved overall higher sensitivity, specificity, and accuracy compared with other widely used classification methods such as random forests (RF), support vector machine (SVM), gradient boosting machine (GBM), and multinomial logistic regression algorithms. Simulation analysis based on random subsampling of genes demonstrated the robustness of DeepCC to missing data. Moreover, deep features learned by DeepCC captured biological characteristics associated with distinct molecular subtypes, enabling more compact within-subtype distribution and between-subtype separation of patient samples, and therefore greatly reduce the number of unclassifiable samples previously. In summary, DeepCC provides a novel cancer classification framework that is platform independent, robust to missing data, and can be used for single sample prediction facilitating clinical implementation of cancer molecular subtyping.

A critical assessment of the feature selection methods used for biomarker discovery in current metaproteomics studies

Microbial community (MC) has great impact on mediating complex disease indications, biogeochemical cycling and agricultural productivities, which makes metaproteomics powerful technique for quantifying diverse and dynamic composition of proteins or peptides. The key role of biostatistical strategies in MC study is reported to be underestimated, especially the appropriate application of feature selection method (FSM) is largely ignored. Although extensive efforts have been devoted to assessing the performance of FSMs, previous studies focused only on their classification accuracy without considering their ability to correctly and comprehensively identify the spiked proteins. In this study, the performances of 14 FSMs were comprehensively assessed based on two key criteria (both sample classification and spiked protein discovery) using a variety of metaproteomics benchmarks. First, the classification accuracies of those 14 FSMs were evaluated. Then, their abilities in identifying the proteins of different spiked concentrations were assessed. Finally, seven FSMs (FC, LMEB, OPLS-DA, PLS-DA, SAM, SVM-RFE and T-Test) were identified as performing consistently superior or good under both criteria with the PLS-DA performing consistently superior. In summary, this study served as comprehensive analysis on the performances of current FSMs and could provide a valuable guideline for researchers in metaproteomics.

Oblique Survival Trees in Discrete Event Time Analysis

One of the main objectives of survival analysis is to predict the failure time that is usually considered as a continuous variable. In longitudinal studies, the data are often collected at every certain period of time, for example, monthly, quarterly, or yearly. Such data require appropriate techniques to handle the discrete time values that often have incomplete information about the failure occurrence-so-called "censored cases." Tree-based models are common, assumption-free methods of survival prediction. In this paper, the author proposes three recursive partitioning techniques able to cope with discrete-time censored survival data, which, in contrast to already-existing models limited to univariate trees, allow splits to have a form of any hyperplane. The performance of proposed methods, expressed as a mean absolute error, was examined on the basis of both synthetic and real data sets available in the literature and compared with existing tree-based models. To demonstrate the applicability of the methods in identifying subgroups of patients with a similar survival experience and to assess the influence of covariates on the risk of failure, a Veteran's Administration lung cancer data set was used. The results confirm proposed models to be good prediction tools for discrete-time survival data.

Variable selection and validation in multivariate modelling

MOTIVATION

Validation of variable selection and predictive performance is crucial in construction of robust multivariate models that generalize well, minimize overfitting and facilitate interpretation of results. Inappropriate variable selection leads instead to selection bias, thereby increasing the risk of model overfitting and false positive discoveries. Although several algorithms exist to identify a minimal set of most informative variables (i.e. the minimal-optimal problem), few can select all variables related to the research question (i.e. the all-relevant problem). Robust algorithms combining identification of both minimal-optimal and all-relevant variables with proper cross-validation are urgently needed.

RESULTS

We developed the MUVR algorithm to improve predictive performance and minimize overfitting and false positives in multivariate analysis. In the MUVR algorithm, minimal variable selection is achieved by performing recursive variable elimination in a repeated double cross-validation (rdCV) procedure. The algorithm supports partial least squares and random forest modelling, and simultaneously identifies minimal-optimal and all-relevant variable sets for regression, classification and multilevel analyses. Using three authentic omics datasets, MUVR yielded parsimonious models with minimal overfitting and improved model performance compared with state-of-the-art rdCV. Moreover, MUVR showed advantages over other variable selection algorithms, i.e. Boruta and VSURF, including simultaneous variable selection and validation scheme and wider applicability.

AVAILABILITY AND IMPLEMENTATION

Algorithms, data, scripts and tutorial are open source and available as an R package ('MUVR') at https://gitlab.com/CarlBrunius/MUVR.git.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Early isolated V-lesion may not truly represent rejection of the kidney allograft

Intimal arteritis is known to be a negative prognostic factor for kidney allograft survival. Isolated v-lesion (IV) is defined as intimal arteritis with minimal tubulointerstitial inflammation (TI). Although the Banff classification assesses IV as T cell-mediated rejection (TCMR), clinical, and prognostic significance of early IV (early IV, eIV) with negative C4d and donor-specific antibodies (DSA) remains unclear. To help resolve if such eIV truly represents acute rejection, a molecular study was performed. The transcriptome of eIV (n=6), T cell-mediated vascular rejection with rich TI (T cell-mediated vascular rejection, TCMRV, n=4) and non-rejection histologic findings (n=8) was compared using microarrays. A total of 310 genes were identified to be deregulated in TCMRV compared with eIV. Gene enrichment analysis categorized deregulated genes to be associated primarily with T-cells associated biological processes involved in an innate and adaptive immune and inflammatory response. Comparison of deregulated gene lists between the study groups and controls showed only a 1.7% gene overlap. Unsupervised hierarchical cluster analysis revealed clear distinction of eIV from TCMRV and showed similarity with a control group. Up-regulation of immune response genes in TCMRV was validated using RT-qPCR in a different set of eIV (n=12) and TCMRV (n=8) samples. The transcriptome of early IV (< 1 month) with negative C4d and DSA is associated with a weak immune signature compared with TCMRV and shows similarity with normal findings. Such eIV may feature non-rejection origin and reflect an injury distinct from an alloimmune response. The present study supports use of molecular methods when interpreting kidney allograft biopsy findings.

Differential privacy-based evaporative cooling feature selection and classification with relief-F and random forests

Motivation

Classification of individuals into disease or clinical categories from high-dimensional biological data with low prediction error is an important challenge of statistical learning in bioinformatics. Feature selection can improve classification accuracy but must be incorporated carefully into cross-validation to avoid overfitting. Recently, feature selection methods based on differential privacy, such as differentially private random forests and reusable holdout sets, have been proposed. However, for domains such as bioinformatics, where the number of features is much larger than the number of observations p≫n , these differential privacy methods are susceptible to overfitting.

Methods

We introduce private Evaporative Cooling, a stochastic privacy-preserving machine learning algorithm that uses Relief-F for feature selection and random forest for privacy preserving classification that also prevents overfitting. We relate the privacy-preserving threshold mechanism to a thermodynamic Maxwell-Boltzmann distribution, where the temperature represents the privacy threshold. We use the thermal statistical physics concept of Evaporative Cooling of atomic gases to perform backward stepwise privacy-preserving feature selection.

Results

On simulated data with main effects and statistical interactions, we compare accuracies on holdout and validation sets for three privacy-preserving methods: the reusable holdout, reusable holdout with random forest, and private Evaporative Cooling, which uses Relief-F feature selection and random forest classification. In simulations where interactions exist between attributes, private Evaporative Cooling provides higher classification accuracy without overfitting based on an independent validation set. In simulations without interactions, thresholdout with random forest and private Evaporative Cooling give comparable accuracies. We also apply these privacy methods to human brain resting-state fMRI data from a study of major depressive disorder.

Availability and implementation

Code available at http://insilico.utulsa.edu/software/privateEC .

Contact

brett-mckinney@utulsa.edu.

Supplementary information

Supplementary data are available at Bioinformatics online.

Tree-based models for survival data with competing risks

OBJECTIVE

Tree-based models belong to common, assumption-free methods of data analysis. Their application in survival data is narrowed to univariate models, which partition the feature space with axis-parallel hyperplanes, meaning that each internal node involves a single feature. In this paper, I extend the idea of oblique survival tree induction for competing risks by modifying a piecewise-linear criterion function. Additionally, the use of tree-based ensembles to analyze the competing events is proposed.

METHOD AND MATERIALS

Two types of competing risks trees are proposed: a single event tree designed for analysis of the event of interest and a composite event tree, in which all the competing events are taken into account. The induction process is similar, except that the calculation of the criterion function is minimized for the individual tree nodes generation. These two tree types were also used for building the ensembles with aggregated cumulative incidence functions as an outcome. Nine real data sets, as well as a simulated data set, were taken to assess performance of the models, while detailed analysis was conducted on the basis of follicular cell lymphoma data.

RESULTS

The evaluation was focused on two measures: the prediction error expressed by an integrated Brier score (IBS), and the ranked measure of predictive ability calculated as a time-truncated concordance index (C-index). The proposed techniques were compared with the existing approaches of the Fine-Gray subdistribution hazard model, Fine-Gray regression model with backward elimination, and random survival forest for competing risks. The results for both the IBS and the C-index indicated statistically significant differences between these methods (p < .0001).

CONCLUSIONS

The prediction error of the individual trees was similar to the other methods, but the results of the C-index differ in comparison to the Fine-Gray subdistribution hazard model and the Fine-Gray regression with backward elimination. The ensembles prediction ability was comparable to existing algorithms, but their IBS values were better than either random survival forest or Fine-Gray regression with backward elimination.

Absolute cosine-based SVM-RFE feature selection method for prostate histopathological grading

OBJECTIVE

Feature selection (FS) methods are widely used in grading and diagnosing prostate histopathological images. In this context, FS is based on the texture features obtained from the lumen, nuclei, cytoplasm and stroma, all of which are important tissue components. However, it is difficult to represent the high-dimensional textures of these tissue components. To solve this problem, we propose a new FS method that enables the selection of features with minimal redundancy in the tissue components.

METHODOLOGY

We categorise tissue images based on the texture of individual tissue components via the construction of a single classifier and also construct an ensemble learning model by merging the values obtained by each classifier. Another issue that arises is overfitting due to the high-dimensional texture of individual tissue components. We propose a new FS method, SVM-RFE(AC), that integrates a Support Vector Machine-Recursive Feature Elimination (SVM-RFE) embedded procedure with an absolute cosine (AC) filter method to prevent redundancy in the selected features of the SV-RFE and an unoptimised classifier in the AC.

RESULTS

We conducted experiments on H&E histopathological prostate and colon cancer images with respect to three prostate classifications, namely benign vs. grade 3, benign vs. grade 4 and grade 3 vs. grade 4. The colon benchmark dataset requires a distinction between grades 1 and 2, which are the most difficult cases to distinguish in the colon domain. The results obtained by both the single and ensemble classification models (which uses the product rule as its merging method) confirm that the proposed SVM-RFE(AC) is superior to the other SVM and SVM-RFE-based methods.

CONCLUSION

We developed an FS method based on SVM-RFE and AC and successfully showed that its use enabled the identification of the most crucial texture feature of each tissue component. Thus, it makes possible the distinction between multiple Gleason grades (e.g. grade 3 vs. grade 4) and its performance is far superior to other reported FS methods.