Different Sets Of Features Research Articles

Natural language processing to identify lupus nephritis phenotype in electronic health records

BackgroundSystemic lupus erythematosus (SLE) is a rare autoimmune disorder characterized by an unpredictable course of flares and remission with diverse manifestations. Lupus nephritis, one of the major disease manifestations of SLE for organ damage and mortality, is a key component of lupus classification criteria. Accurately identifying lupus nephritis in electronic health records (EHRs) would therefore benefit large cohort observational studies and clinical trials where characterization of the patient population is critical for recruitment, study design, and analysis. Lupus nephritis can be recognized through procedure codes and structured data, such as laboratory tests. However, other critical information documenting lupus nephritis, such as histologic reports from kidney biopsies and prior medical history narratives, require sophisticated text processing to mine information from pathology reports and clinical notes. In this study, we developed algorithms to identify lupus nephritis with and without natural language processing (NLP) using EHR data from the Northwestern Medicine Enterprise Data Warehouse (NMEDW).MethodsWe developed five algorithms: a rule-based algorithm using only structured data (baseline algorithm) and four algorithms using different NLP models. The first NLP model applied simple regular expression for keywords search combined with structured data. The other three NLP models were based on regularized logistic regression and used different sets of features including positive mention of concept unique identifiers (CUIs), number of appearances of CUIs, and a mixture of three components (i.e. a curated list of CUIs, regular expression concepts, structured data) respectively. The baseline algorithm and the best performing NLP algorithm were externally validated on a dataset from Vanderbilt University Medical Center (VUMC).ResultsOur best performing NLP model incorporated features from both structured data, regular expression concepts, and mapped concept unique identifiers (CUIs) and showed improved F measure in both the NMEDW (0.41 vs 0.79) and VUMC (0.52 vs 0.93) datasets compared to the baseline lupus nephritis algorithm.ConclusionOur NLP MetaMap mixed model improved the F-measure greatly compared to the structured data only algorithm in both internal and external validation datasets. The NLP algorithms can serve as powerful tools to accurately identify lupus nephritis phenotype in EHR for clinical research and better targeted therapies.

Battery Remaining Useful Life Prediction Using Machine Learning Models: A Comparative Study

Predicting the remaining useful life (RUL) of lithium-ion (Li-ion) batteries is crucial to preventing system failures and enhancing operational performance. Knowing the RUL of a battery enables one to perform preventative maintenance or replace the battery before its useful life expires, which is vital in safety-critical applications. The prediction of the RUL of Li-ion batteries plays a critical role in their optimal utilization throughout their lifetime and supporting sustainable practices. This paper conducts a comparative analysis to assess the effectiveness of multiple machine learning (ML) models in predicting the capacity fade and RUL of Li-ion batteries. Three case studies are analyzed to assess the performances of the state-of-the-art ML models, considering two distinct datasets. These case studies are conducted under various operating conditions such as temperature, C-rate, state of charge (SOC), and depth of discharge (DOD) of the batteries in Cases 1 and 2, and a different set of features and charging policies for the second dataset in Case 3. Meanwhile, diverse extracted features from the initial cycles of the second dataset are considered in Case 3 to predict the RUL of Li-ion batteries in all cycles. In addition, a multi-feature multi-target (MFMT) feature mapping is introduced to investigate the performance of the developed ML models in predicting the battery capacity fade and RUL in the entire life cycle. Multiple ML models that are developed for the comparison analysis in the proposed methodology include Random Forest (RF), extreme gradient boosting (XGBoost), light gradient-boosting machine (LightGBM), multi-layer perceptron (MLP), long short-term memory (LSTM), and attention-LSTM. Furthermore, hyperparameter tuning is applied to improve the performance of the XGBoost and LightGBM models. The results demonstrate that the extreme gradient boosting with hyperparameter tuning (XGBoost-HT) model outperforms the other ML models in terms of the root-mean-squared error (RMSE) and mean absolute percentage error (MAPE) of the battery capacity fade and RUL for all cycles. The obtained RMSE and MAPE values for XGBoost-HT in terms of cycle life are 69 cycles and 6.5%, respectively, for the third case. In addition, the XGBoost-HT model handles the MFMT feature mapping within an acceptable range of RMSE and MAPE, compared to the rest of the developed ML models and similar benchmarks.

Outlier detection using flexible categorization and interrogative agendas

Categorization is one of the basic tasks in machine learning and data analysis. Building on formal concept analysis (FCA), the starting point of the present work is that different ways to categorize a given set of objects exist, which depend on the choice of the sets of features used to classify them, and different such sets of features may yield better or worse categorizations, relative to the task at hand. In their turn, the (a priori) choice of a particular set of features over another might be subjective and express a certain epistemic stance (e.g. interests, relevance, preferences) of an agent or a group of agents, namely, their interrogative agenda. In the present paper, we represent interrogative agendas as sets of features, and explore and compare different ways to categorize objects w.r.t. different sets of features (agendas). We first develop a simple unsupervised FCA-based algorithm for outlier detection which uses categorizations arising from different agendas. We then present a supervised meta-learning algorithm to learn suitable (fuzzy) agendas for categorization as sets of features with different weights or masses. We combine this meta-learning algorithm with the unsupervised outlier detection algorithm to obtain a supervised outlier detection algorithm. We show that these algorithms perform at par with commonly used algorithms for outlier detection on commonly used datasets in outlier detection. These algorithms provide both local and global explanations of their results.

Graph Neural Network and BERT Model for Antimalarial Drug Predictions Using Plasmodium Potential Targets

Malaria continues to pose a significant global health burden despite concerted efforts to combat it. In 2020, nearly half of the world’s population faced the risk of malaria, underscoring the urgency of innovative strategies to tackle this pervasive threat. One of the major challenges lies in the emergence of the resistance of parasites to existing antimalarial drugs. This challenge necessitates the discovery of new, effective treatments capable of combating the Plasmodium parasite at various stages of its life cycle. Advanced computational approaches have been utilized to accelerate drug development, playing a crucial role in every stage of the drug discovery and development process. We have witnessed impressive and groundbreaking achievements, with GNNs applied to graph data and BERT from transformers across diverse NLP text analysis tasks. In this study, to facilitate a more efficient and effective approach, we proposed the integration of an NLP based model for SMILES (i.e., BERT) and a GNN model (i.e., RGCN) to predict the effect of antimalarial drugs against Plasmodium. The GNN model was trained using designed antimalarial drug and potential target (i.e., PfAcAS, F/GGPPS, and PfMAGL) graph-structured data with nodes representing antimalarial drugs and potential targets, and edges representing relationships between them. The performance of BERT-RGCN was further compared with that of Mordred-RGCN to evaluate its effectiveness. The BERT-RGCN and Mordred-RGCN models performed consistently well across different feature combinations, showcasing high accuracy, sensitivity, specificity, MCC, AUROC, and AUPRC values. These results suggest the effectiveness of the models in predicting antimalarial drugs against Plasmodium falciparum in various scenarios based on different sets of features of drugs and potential antimalarial targets.

Self-supervised prediction of the intention to interact with a service robot

A service robot can provide a smoother interaction experience if it has the ability to proactively detect whether a nearby user intends to interact, in order to adapt its behavior e.g. by explicitly showing that it is available to provide a service. In this work, we propose a learning-based approach to predict the probability that a human user will interact with a robot before the interaction actually begins; the approach is self-supervised because after each encounter with a human, the robot can automatically label it depending on whether it resulted in an interaction or not. We explore different classification approaches, using different sets of features considering the pose and the motion of the user. We validate and deploy the approach in three scenarios. The first collects 3442 natural sequences (both interacting and non-interacting) representing employees in an office break area: a real-world, challenging setting, where we consider a coffee machine in place of a service robot. The other two scenarios represent researchers interacting with service robots (200 and 72 sequences, respectively). Results show that, even in challenging real-world settings, our approach can learn without external supervision, and can achieve accurate classification (i.e. AUROC greater than 0.9) of the user’s intention to interact with an advance of more than 3 s before the interaction actually occurs.

I/O Access Patterns in HPC Applications: A 360-Degree Survey

The high-performance computing I/O stack has been complex due to multiple software layers, the inter-dependencies among these layers, and the different performance tuning options for each layer. In this complex stack, the definition of an “I/O access pattern” has been reappropriated to describe what an application is doing to write or read data from the perspective of different layers of the stack, often comprising a different set of features. It has become common to have to redefine what is meant when discussing a pattern in every new study, as no assumption can be made. This survey aims to propose a baseline taxonomy, harnessing the I/O community’s knowledge over the past 20 years. This definition can serve as a common ground for high-performance computing I/O researchers and developers to apply known I/O tuning strategies and design new strategies for improving I/O performance. We seek to summarize and bring a consensus to the multiple ways to describe a pattern based on common features already used by the community over the years.

Development of Trio Optimal Feature Extraction Model for Attention-Based Adaptive Weighted RNN-Based Lung and Colon Cancer Detection Framework Using Histopathological Images

Due to the combination of genetic diseases as well as a variety of biomedical abnormalities, the fatal disease named cancer is caused. Colon and lung cancer are regarded as the two leading diseases for disability and death. The most significant component for demonstrating the best course of action is the histopathological identification of such malignancies. So, in order to minimize the mortality rate caused by cancer, there is a need for early detection of the aliment on both fronts accordingly. In this case, both the deep and machine learning techniques have been utilized to speed up the detection process of cancer which may also help the researchers to study a huge amount of patients over a short period and less loss. Hence, it is highly essential to design a new lung and colon detection model based on deep learning approaches. Initially, a different set of histopathological images is collected from benchmark resources to perform effective analysis. Then, to attain the first set of features, the collected image is offered to the dilated net for attaining deep image features with the help of the Visual Geometry Group (VGG16) and Residual Neural Network (ResNet). Further, the second set of features is attained by the below process. Here, the collected image is given to pre-processing phase and the image is pre–pre-processed with the help of Contrast-limited Adaptive Histogram Equalization (CLAHE) and filter technique. Then, the pre-processed image is offered to the segmentation phase with the help of adaptive binary thresholding and offered to a dilated network that holds VGG16 and ResNet and attained the second set of features. The parameters of adaptive binary thresholding are tuned with the help of a developed hybrid approach called Sand Cat swarm JAya Optimization (SC-JAO) via Sand Cat swarm Optimization (SCO) and JAYA (SC-JAO). Finally, the third set of features is attained by offering the image to pre-processing phase. Then, the pre-processed image is offered to the segmentation phase and the image is a segmented phase and features are tuned by developed SC-JAO. Further, the segmented features are offered to attain the textural features like Gray-Level Co-Occurrence Matrix (GLCM) and Local Weber Pattern (LWP) and attained the third set of features. Then, the attained three different sets of features are given to the optimal weighted feature phase, where the parameters are optimized by the SC-JAO algorithm and then given to the disease prediction phase. Here, disease prediction is made with the help of Attention-based Adaptive Weighted Recurrent Neural Networks (AAW-RNN), and their parameters are tuned by developed SC-JAO. Thus, the developed model achieved an effective lung and colon detection rate over conventional approaches over multiple experimental analyses.

Assessing consciousness in patients with disorders of consciousness using soft-clustering

Consciousness is something we experience in our everyday life, more especially between the time we wake up in the morning and go to sleep at night, but also during the rapid eye movement (REM) sleep stage. Disorders of consciousness (DoC) are states in which a person’s consciousness is damaged, possibly after a traumatic brain injury. Completely locked-in syndrome (CLIS) patients, on the other hand, display covert states of consciousness. Although they appear unconscious, their cognitive functions are mostly intact. Only, they cannot externally display it due to their quadriplegia and inability to speak. Determining these patients’ states constitutes a challenging task. The ultimate goal of the approach presented in this paper is to assess these CLIS patients consciousness states. EEG data from DoC patients are used here first, under the assumption that if the proposed approach is able to accurately assess their consciousness states, it will assuredly do so on CLIS patients too. This method combines different sets of features consisting of spectral, complexity and connectivity measures in order to increase the probability of correctly estimating their consciousness levels. The obtained results showed that the proposed approach was able to correctly estimate several DoC patients’ consciousness levels. This estimation is intended as a step prior attempting to communicate with them, in order to maximise the efficiency of brain–computer interfaces (BCI)-based communication systems.

AWC‐NIDS: Attack‐wise customized network intrusion detection system using machine learning, concurrency, and distributed systems

SummaryWith digitization and modern network applications, information security has gained a tremendous importance. Therefore, accurate and efficient detection systems are crucial for maintaining proactive security in computer networks. Machine learning (ML) has shown great potential as a promising solution since it can teach a machine to distinguish malicious and normal network activities. However, recently proposed methods are suffering from at least one of the following: detection accuracy, false alarm rate, and computational complexity issues. The main reason behind this problem is the complexity of the model in terms of attack types. From the ML perspective, intrusion detection is a classification problem where each attack type is identified by a set of different features, and features are used for classifying network activities. Thus, training an ML algorithm to detect more than one attack type leads to a more complex model; the increasing number of used features contributes positively to the model complexity, and may result in relatively lower detection accuracy or a higher false positive rate. To tackle this problem, this study proposes an attack‐wise customized network intrusion detection system (AWC‐NIDS) based on ML, concurrency, and distributed systems to achieve accurate and efficient network‐wide intrusion detection. Since CICIDS2017 contains many modern attacks, it was used for model development and performance evaluation. The experimental results showed that the proposed methodology achieved high classification performance for all datasets with a small number of features. However, it was observed that the lowest accuracy was achieved for the comprehensive dataset (which contains all attack types); for the single attack‐type datasets, the obtained accuracy was above 99%. This finding proves the concept of attack‐wise customization for intrusion detection and shows the significance of the proposed methodology. In conclusion, this framework is promising for implementing robust and accurate cybersecurity systems for traditional and modern networking.

Prediction of Early Dropouts in Patient Remote Monitoring Programs

The analysis of medical data is a significant opportunity worldwide for national health systems to reduce costs and at the same time improve healthcare. The utilization of these technologies is done in the context of monitoring health issues, counting health goals, as well as for recording medical data. In such a context, early detection of users at risk of lower compliance rates and patterns of use of a health monitoring application suggesting a risk of abandonment is an invaluable opportunity to implement tailored intervention strategies aimed at recovering and avoiding abandonment thoughts. This study aims to identify patterns of early dropout in users of an application for mobile intervention, having access to a database of users who have experienced the impact of a digital monitoring application to improve their quality of life for at least 6 months. At the experimental stage, many different approaches for early dropout prediction were implemented with a different set of features. Specifically, the current study proposes a methodology using the Neighborhood Cleaning Rule and a specific classification algorithm based on the Stacked Generalization learning method to predict the early abandonment of users of the health monitoring application. The results showed that the proposed algorithm was able to predict the early dropout of users from the application with an accuracy of 97.6%, making it reliable enough to be used as an early warning system.

What is a consistent glycan composition dataset?

Introduction: One of the main challenges in bioinformatics has been and still is, the comparison of entities through the development of algorithms for similarity scoring and data clustering according to biologically relevant aspects. Glycoinformatics also faces this challenge, in particular regarding the automated comparison of protein and/or tissue glycomes, that remains a relatively uncharted territory.Methods: Low and high throughput experimental glycomic and glycoproteomic results were collected, revealing a bias toward N-linked glycomes. Then, N-glycomes were considered and represented as networks of related glycan compositions as opposed to lists of glycans. They were processed and compared through a java application generating graphs and another producing a similarity matrix based on graph content. Several scoring schemes (e.g., Jaccard index or cosine) were tested and evaluated using the Matthews Correlation Coefficient, in order to capture a meaningful protein and tissue N-glycome similarity.Results: Assuming that a glycome corresponds to a well-connected graph of glycan compositions, graph comparison has revealed gaps that can be interpreted as inconsistencies. The outcome of systematic graph comparison is both formal and practical. In principle, it is shown that the idiosyncrasy of current glycome data limits the definition of appropriate estimates for systematically comparing N-glycomes. Yet, several potentially interesting criteria could be identified in a series of use cases detailed in the study.Discussion: Differentially expressed glycomes are usually compared manually, but the resulting work tends to remain in publications due to the lack of dedicated tools. Even manually, cross-comparison is challenging mostly because different sets of features are used from one study to the other. The work presented here enables laying down guidelines for developing a software tool comparing glycomes based on appropriate definitions of similarity and suitable methods for its evaluation and implementation.

Artificial Intelligence-Empowered Logistic Traffic Management System Using Empirical Intelligent XGBoost Technique in Vehicular Edge Networks

Recent advancements in computation and communication technologies and the increasing adoption of the Internet of Things (IoT) and Artificial Intelligence (AI) technologies have paved the way to tremendous developments in modern transportation systems. Driven by the massive number of connected vehicles and the stringent requirements of the public traffic management system, the transportation of data to and from the centralized cloud servers poses a great challenge. As a result, to meet the computational requirements and handle the massive amount of sensory data efficiently, the potential solution is to process/analyze the data at the edge of the network. Motivated by the challenges mentioned above, in this paper, we design a new empirically intelligent XGboost (EIXGB)-enabled logistic transportation system at the edge network for analyzing the data efficiently. Besides that, the proposed EIXGB technique intends to obtain real-time results based on the monitoring parameters of the public traffic management system with higher accuracy and minimum error. Extensive simulation results demonstrate the efficiency of the proposed EIXGB technique over the standard machine learning techniques using a set of parameters. The proposed technique achieves 87-97% accuracy over the different sets of features of a real-time dataset as per the simulation results.

Improvement of q 2 Resolution in Semileptonic Decays Based on Machine Learning

The neutrino closure method is often used to obtain kinematics of semileptonic decays with one unreconstructed particle in hadron collider experiments. The kinematics of decays can be deducted by a twofold ambiguity with a quadratic equation. To resolve the twofold ambiguity, a novel method based on machine learning (ML) is proposed. We study the effect of different sets of features and regressors on the improvement of reconstructed invariant mass squared of ℓ ν system ( q 2 ). The result shows that the best performance is obtained by using the flight vector as the features and the multilayer perceptron (MLP) model as the regressor. Compared with the random choice, the MLP model improves the resolution of reconstructed q 2 by ~40%. Furthermore, the possibility of using this method on various semileptonic decays is shown.

Empire, Hegemony, Hyperpower?

Empire and Hegemony are among the crucial notions of science of international relations. Many scholars were investigating those concepts, and attempted to define them and estimate whether they retain explanatory value in contemporary situations of accelerated evolution of international system. Within the system governed by anarchy, those two prepositions were considered partial resolution to destructive tendencies generated within the international environment. Albeit those resolutions, it seems similar, their application required different conditions in order to become effective. Every mentioned above system is a complex and adequate answer to challenge posed by different set of features and variables generated by particular iteration of international system. The question remains: preconditions to occurrence of which system will be generated by late-Westphalian and subsequent generation of international environment? Contemporary iteration projects many individual characteristics, with regard to complexity, globalization and historical acceleration. One of the possible solutions to this issue is the recently introduced concept of Hyperpower, which could be positioned as a system in between those mentioned above. However its creation embraces new quality of international relations not yet encountered during course of history. This paper will be dedicated to the investigation of this concept and its usefulness for scientific analysis of contemporary international relations. Hyperpower embraces geopolitics and transnational in equal measures. It is product of a so-called “virtualization of state” and encompasses tools for effective interaction in both spheres of international environment. It is a generally much more passive system than mentioned above. Its activation consumes astounding amount of resources. Therefore, its primary application is in most cases passive, strictly limited to the shaping of the perception of remaining elements of international system. Keywords: Hyperpower, Hegemony, Empire, late-Westphalian international environment

Exploring the dominant features and data-driven detection of polycystic ovary syndrome through modified stacking ensemble machine learning technique

Polycystic ovary syndrome (PCOS) is the most frequent endocrinological anomaly in reproductive women that causes persistent hormonal secretion disruption, leading to the formation of numerous cysts within the ovaries and serious health complications. But the real-world clinical detection technique for PCOS is very critical since the accuracy of interpretations being substantially dependent on the physician's expertise. Thus, an artificially intelligent PCOS prediction model might be a feasible additional technique to the error prone and time-consuming diagnostic technique. In this study, a modified ensemble machine learning (ML) classification approach is proposed utilizing state-of-the-art stacking technique for PCOS identification with patients' symptom data; employing five traditional ML models as base learners and then one bagging or boosting ensemble ML model as the meta-learner of the stacked model. Furthermore, three distinct types of feature selection strategies are applied to pick different sets of features with varied numbers and combinations of attributes. To evaluate and explore the dominant features necessary for predicting PCOS, the proposed technique with five variety of models and other ten types of classifiers is trained, tested and assessed utilizing different feature sets. As outcomes, the proposed stacking ensemble technique significantly enhances the accuracy in comparison to the other existing ML based techniques in case of all varieties of feature sets. However, among various models investigated to categorize PCOS and non-PCOS patients, the stacking ensemble model with ‘Gradient Boosting’ classifier as meta learner outperforms others with 95.7% accuracy while utilizing the top 25 features selected using Principal Component Analysis (PCA) feature selection technique.

Modified Firefly Optimization Algorithm-Based IDS for Nature-Inspired Cybersecurity

The new paradigm of nature-inspired cybersecurity can establish a robust defense by utilizing well-established nature-inspired computing algorithms to analyze networks and act quickly. The existing research focuses primarily on the efficient selection of features for quick and optimized detection rates using firefly and other nature-inspired optimization techniques. However, selecting the most appropriate features may be specific to the network, and a different set of features may work better than the selected one. Therefore, there is a need for a generalized pre-processing step based on the standard network monitoring parameters for the early detection of suspicious nodes before applying feature-based or any other type of monitoring. This paper proposes a modified version of the firefly optimization algorithm to effectively monitor the network by introducing a novel health function for the early detection of suspicious nodes. We implement event management schemes based on the proposed algorithm and optimize the observation priority list based on a genetic evolution algorithm for real-time events in the network. The obtained simulation results demonstrate the effectiveness of the proposed algorithm under various attack scenarios. In addition, the results indicate that the proposed method reduces approximately 60–80% of the number of suspicious nodes while increasing the turnaround time by only approximately 1–2%. The proposed method also focuses specifically on accurate network health monitoring to protect the network proactively.

Predicting Return to Work after Cardiac Rehabilitation.

To explore machine learning models for predicting return to work after cardiac rehabilitation. Patients who were admitted to the University of Malaya Medical Centre due to cardiac events. Eight different machine learning models were evaluated. The models included 3 different sets of features: full features; significant features from multiple logistic regression; and features selected from recursive feature extraction technique. The performance of the prediction models with each set of features was compared. The AdaBoost model with the top 20 features obtained the highest performance score of 92.4% (area under the curve; AUC) compared with other prediction models.

Leaf-based Classification of Important Indigenous Tree Species by Different Feature Extraction Techniques and Selected Classification Algorithms

The machine learning algorithms, namely, k-Nearest Neighbor (KNN), Support Vector Machine (SVM), Back-Propagation (BP) networks, and Convolutional Neural Networks (CNN) are four of the mostly used classifiers. Different sets of features are required as input in different application domains. In this paper, a set of significant leaf features and classification model was determined with a high accuracy in classifying important indigenous tree species. Leaf images were acquired using a scanner to control the image quality. The image dataset was then duplicated into two sets. The first set was labeled with their correct classes, preprocessed, and segmented in preparation for feature extraction. The leaf features extracted were leaf shape, leaf color, and leaf texture. Then, training and classification was done by KNN, SVM, and BP networks. On the other hand, the second set was unlabeled for training and classification by CNN. A CNN model was built and chosen with the best training and validation accuracy and the least training and validation loss rate. The study concluded that using all three leaf features for classification by BP networks resulted in a 93.48% accuracy with training done by supervised learning. However, the CNN achieved a high accuracy rate of 98.5% making it the best approach for classification of tree species using digital leaf images in the context of this study.

Application of Soft-Clustering to Assess Consciousness in a CLIS Patient.

Completely locked-in (CLIS) patients are characterized by sufficiently intact cognitive functions, but a complete paralysis that prevents them to interact with their surroundings. On one hand, studies have shown that the ability to communicate plays an important part in these patients' quality of life and prognosis. On the other hand, brain-computer interfaces (BCIs) provide a means for them to communicate using their brain signals. However, one major problem for such patients is the difficulty to determine if they are conscious or not at a specific time. This work aims to combine different sets of features consisting of spectral, complexity and connectivity measures, to increase the probability of correctly estimating CLIS patients' consciousness levels. The proposed approach was tested on data from one CLIS patient, which is particular in the sense that the experimenter was able to point out one time frame Δt during which he was undoubtedly conscious. Results showed that the method presented in this paper was able to detect increases and decreases of the patient's consciousness levels. More specifically, increases were observed during this Δt, corroborating the assertion of the experimenter reporting that the patient was definitely conscious then. Assessing the patients' consciousness is intended as a step prior attempting to communicate with them, in order to maximize the efficiency of BCI-based communication systems.

A pairwise strategy for imputing predictive features when combining multiple datasets.

In the training of predictive models using high-dimensional genomic data, multiple studies' worth of data are often combined to increase sample size and improve generalizability. A drawback of this approach is that there may be different sets of features measured in each study due to variations in expression measurement platform or technology. It is often common practice to work only with the intersection of features measured in common across all studies, which results in the blind discarding of potentially useful feature information that is measured in individual or subsets of studies. We characterize the loss in predictive performance incurred by using only the intersection of feature information available across all studies when training predictors using gene expression data from microarray and sequencing datasets. We study the properties of linear and polynomial regression for imputing discarded features and demonstrate improvements in the external performance of prediction functions through simulation and in gene expression data collected on breast cancer patients. To improve this process, we propose a pairwise strategy that applies any imputation algorithm to two studies at a time and averages imputed features across pairs. We demonstrate that the pairwise strategy is preferable to first merging all datasets together and imputing any resulting missing features. Finally, we provide insights on which subsets of intersected and study-specific features should be used so that missing-feature imputation best promotes cross-study replicability. The code is available at https://github.com/YujieWuu/Pairwise_imputation. Supplementary information is available at Bioinformatics online.