Dataset Bias Research Articles

AI Bias in Healthcare: Using ImpactPro as a Case Study for Healthcare Practitioners’ Duties to Engage in Anti-Bias Measures

The introduction of ImpactPro to identify patients with complex health needs suggests that current bias and impacts of bias in healthcare AIs stem from historically biased practices leading to biased datasets, a lack of oversight, as well as bias in practitioners who are overseeing AIs. In order to improve these outcomes, healthcare practitioners need to engage in current best practices for anti-bias training.

Deep learning-based unlearning of dataset bias for MRI harmonisation and confound removal

Increasingly large MRI neuroimaging datasets are becoming available, including many highly multi-site multi-scanner datasets. Combining the data from the different scanners is vital for increased statistical power; however, this leads to an increase in variance due to nonbiological factors such as the differences in acquisition protocols and hardware, which can mask signals of interest.We propose a deep learning based training scheme, inspired by domain adaptation techniques, which uses an iterative update approach to aim to create scanner-invariant features while simultaneously maintaining performance on the main task of interest, thus reducing the influence of scanner on network predictions. We demonstrate the framework for regression, classification and segmentation tasks with two different network architectures.We show that not only can the framework harmonise many-site datasets but it can also adapt to many data scenarios, including biased datasets and limited training labels. Finally, we show that the framework can be extended for the removal of other known confounds in addition to scanner. The overall framework is therefore flexible and should be applicable to a wide range of neuroimaging studies.

Measurements of dielectron production in Au+Au collisions at [formula omitted] with the STAR experiment

Dielectrons are ideal probes of the hot and dense medium created in relativistic heavy-ion collisions due to their minimal interactions with the partonic and hadronic medium. They carry the information from the initial to the final stage of a collision. We present the dielectron spectra in Au+Au collisions at sNN=27,54.4 and 200GeV measured with the STAR experiment. The large minimum bias datasets at sNN=27(54.4)GeV taken in 2017 (2018) significantly enhance the precision of the in-medium ρ modification measurement compared to previous STAR BES-I results. The Heavy Flavor Tracker (HFT) installed at STAR between 2014 and 2016 enables a better understanding of the semi-leptonic charm decay contributions by providing high-precision tracking and vertex information. Strategies to extract the dielectron spectra and suppress the charm component to the spectrum with the HFT in sNN=200GeV Au+Au collisions are discussed.

Interdisciplinary Research in Artificial Intelligence: Challenges and Opportunities.

The use of artificial intelligence (AI) in a variety of research fields is speeding up multiple digital revolutions, from shifting paradigms in healthcare, precision medicine and wearable sensing, to public services and education offered to the masses around the world, to future cities made optimally efficient by autonomous driving. When a revolution happens, the consequences are not obvious straight away, and to date, there is no uniformly adapted framework to guide AI research to ensure a sustainable societal transition. To answer this need, here we analyze three key challenges to interdisciplinary AI research, and deliver three broad conclusions: 1) future development of AI should not only impact other scientific domains but should also take inspiration and benefit from other fields of science, 2) AI research must be accompanied by decision explainability, dataset bias transparency as well as development of evaluation methodologies and creation of regulatory agencies to ensure responsibility, and 3) AI education should receive more attention, efforts and innovation from the educational and scientific communities. Our analysis is of interest not only to AI practitioners but also to other researchers and the general public as it offers ways to guide the emerging collaborations and interactions toward the most fruitful outcomes.

Three Frontiers for the Future of Biodiversity Research Using Citizen Science Data

AbstractCitizen science is fundamentally shifting the future of biodiversity research. But although citizen science observations are contributing an increasingly large proportion of biodiversity data, they only feature in a relatively small percentage of research papers on biodiversity. We provide our perspective on three frontiers of citizen science research, areas that we feel to date have had minimal scientific exploration but that we believe deserve greater attention as they present substantial opportunities for the future of biodiversity research: sampling the undersampled, capitalizing on citizen science's unique ability to sample poorly sampled taxa and regions of the world, reducing taxonomic and spatial biases in global biodiversity data sets; estimating abundance and density in space and time, develop techniques to derive taxon-specific densities from presence or absence and presence-only data; and capitalizing on secondary data collection, moving beyond data on the occurrence of single species and gain further understanding of ecological interactions among species or habitats. The contribution of citizen science to understanding the important biodiversity questions of our time should be more fully realized.

Application of horizontal wells in three-dimensional shale reservoir modeling: A case study of Longmaxi–Wufeng shale in Fuling gas field, Sichuan Basin: Reply

Three-dimensional reservoir modeling is an important aspect to determine the heterogeneity of organic-rich shale reservoirs, an area of study that continues to be explored and refined. A large proportion of data acquired from horizontal wells causes issues in the structural and property modeling for shale reservoirs. Since horizontal wells are designed to drill into a specific, narrow zone, their horizontal section tends to parallel or nearly parallel formation surfaces. As a result, formation surfaces have a much more complex spatial location relationship with horizontal wellbores than with vertical wellbores. The existing algorithms are not good at addressing this issue during structural modeling. The major problem of using horizontal well data in property modeling is the biased data set because their horizontal section tends to stay within a narrow zone. The property distribution feature estimated from this biased data set, as a significant, default input of geostatistical simulation algorithms, causes the constructed property models to deviate away from the real case in the subsurface. A method to infer more formation tops in pseudovertical wells according to a series of assumptions was developed to provide more constraint points for structural modeling within the areas of the horizontal well section. To use the biased database from horizontal wells, distribution function and trend model methods were developed for continuous property modeling, and percentage and probability trend models were developed for discrete property modeling. The Longmaxi–Wufeng shale in the Fuling gas field of Sichuan Basin was used as an example to express and verify these methods.

Deep Data Source Fusion with Bias-Undoing for Lung Adenocarcinoma Classification

Computer-aided diagnosis of early-stage lung adenocarcinoma based on deep learning is prospective for assisting the prevention and treatment of the deathly disease of lung cancer, however, relevant works face the problem of limited training data. The technique of data source fusion with the training of deep models on multiple relevant datasets is promising to resolve the lack of training data, while the bias of data distribution from different data sources exists as a universal issue to affect the learning performance. In this paper, we propose a deep learning framework based on bias-undoing data source fusion to classify early stages of lung adenocarcinoma in computed tomography (CT) images. The framework conducts learning on the integrated datasets for respectively natural image, lung nodule CT and lung adenocarcinoma CT, as designed with an organization of base parameters and bias parameters to adapt to the data distribution with bias. Experimental results demonstrate that the proposed bias-undoing framework is effective to improve the performance of deep learning for lung adenocarcinoma classification, and is with great superiority to those general fusion frameworks on alleviating the effect of dataset bias.

Improved cytokine\u2013receptor interaction prediction by exploiting the negative sample space

BackgroundCytokines act by binding to specific receptors in the plasma membrane of target cells. Knowledge of cytokine–receptor interaction (CRI) is very important for understanding the pathogenesis of various human diseases—notably autoimmune, inflammatory and infectious diseases—and identifying potential therapeutic targets. Recently, machine learning algorithms have been used to predict CRIs. “Gold Standard” negative datasets are still lacking and strong biases in negative datasets can significantly affect the training of learning algorithms and their evaluation. To mitigate the unrepresentativeness and bias inherent in the negative sample selection (non-interacting proteins), we propose a clustering-based approach for representative negative sample selection.ResultsWe used deep autoencoders to investigate the effect of different sampling approaches for non-interacting pairs on the training and the performance of machine learning classifiers. By using the anomaly detection capabilities of deep autoencoders we deduced the effects of different categories of negative samples on the training of learning algorithms. Random sampling for selecting non-interacting pairs results in either over- or under-representation of hard or easy to classify instances. When K-means based sampling of negative datasets is applied to mitigate the inadequacies of random sampling, random forest (RF) together with the combined feature set of atomic composition, physicochemical-2grams and two different representations of evolutionary information performs best. Average model performances based on leave-one-out cross validation (loocv) over ten different negative sample sets that each model was trained with, show that RF models significantly outperform the previous best CRI predictor in terms of accuracy (+ 5.1%), specificity (+ 13%), mcc (+ 0.1) and g-means value (+ 5.1). Evaluations using tenfold cv and training/testing splits confirm the competitive performance.ConclusionsA comparative analysis was performed to assess the effect of three different sampling methods (random, K-means and uniform sampling) on the training of learning algorithms using different evaluation methods. Models trained on K-means sampled datasets generally show a significantly improved performance compared to those trained on random selections—with RF seemingly benefiting most in our particular setting. Our findings on the sampling are highly relevant and apply to many applications of supervised learning approaches in bioinformatics.

Unifying Relational Sentence Generation and Retrieval for Medical Image Report Composition.

Beyond generating long and topic-coherent paragraphs in traditional captioning tasks, the medical image report composition task poses more task-oriented challenges by requiring both the highly accurate medical term diagnosis and multiple heterogeneous forms of information, including impression and findings. Current methods often generate the most common sentences due to dataset bias for the individual case, regardless of whether the sentences properly capture key entities and relationships. Such limitations severely hinder their applicability and generalization capability in medical report composition, where the most critical sentences lie in the descriptions of abnormal diseases that are relatively rare. Moreover, some medical terms appearing in one report are often entangled with each other and co-occurred, for example, symptoms associated with a specific disease. To enforce the semantic consistency of medical terms to be incorporated into the final reports and encourage the sentence generation for rare abnormal descriptions, we propose a novel framework that unifies template retrieval and sentence generation to handle both common and rare abnormality while ensuring the semantic coherency among the detected medical terms. Specifically, our approach exploits hybrid-knowledge co-reasoning: 1) explicit relationships among all abnormal medical terms to induce the visual attention learning and topic representation encoding for better topic-oriented symptoms descriptions and 2) adaptive generation mode that changes between the template retrieval and sentence generation according to a contextual topic encoder. The experimental results on two medical report benchmarks demonstrate the superiority of the proposed framework in terms of both human and metrics evaluation.

Known unknowns: media bias in the reporting of political violence

ABSTRACT How does sourcing affect which events are included in international relations datasets? The increasing number of machine-coded datasets offers the promise of coding a larger corpus of documents more quickly, but existing automated processes rely exclusively on databases of news reports for coverage. We exploit source variation in the UCDP GED dataset, which includes events from media reports and non-media sources, to explore the bias introduced by including only media reports in international relations datasets. Unlike previous studies, our approach allows us to compare subnational and cross-national determinants of bias. We find that media sources severely underreport events in African countries, and coverage is also associated with country-level factors like international trade and subnational factors like access to communication technology. Non-media sources cover a significant number of events not included in media sources; their inclusion can expand coverage and reduce bias in datasets.

Detect and correct bias in multi-site neuroimaging datasets.

The desire to train complex machine learning algorithms and to increase the statistical power in association studies drives neuroimaging research to use ever-larger datasets. The most obvious way to increase sample size is by pooling scans from independent studies. However, simple pooling is often ill-advised as selection, measurement, and confounding biases may creep in and yield spurious correlations. In this work, we combine 35,320 magnetic resonance images of the brain from 17 studies to examine bias in neuroimaging. In the first experiment, Name That Dataset, we provide empirical evidence for the presence of bias by showing that scans can be correctly assigned to their respective dataset with 71.5% accuracy. Given such evidence, we take a closer look at confounding bias, which is often viewed as the main shortcoming in observational studies. In practice, we neither know all potential confounders nor do we have data on them. Hence, we model confounders as unknown, latent variables. Kolmogorov complexity is then used to decide whether the confounded or the causal model provides the simplest factorization of the graphical model. Finally, we present methods for dataset harmonization and study their ability to remove bias in imaging features. In particular, we propose an extension of the recently introduced ComBat algorithm to control for global variation across image features, inspired by adjusting for unknown population stratification in genetics. Our results demonstrate that harmonization can reduce dataset-specific information in image features. Further, confounding bias can be reduced and even turned into a causal relationship. However, harmonization also requires caution as it can easily remove relevant subject-specific information. Code is available at https://github.com/ai-med/Dataset-Bias.

Quantifying effects of tracking data bias on species distribution models

Abstract Telemetry datasets are becoming increasingly large and covering a wider range of species using different technologies (GPS, Argos, light‐based geolocation). Together, such datasets hold tremendous potential to understand species' space use at broad spatial scale, through the development of species distribution or habitat suitability models (SDMs) to predict environmental dependencies of species across space and time. However, tracking datasets can be heavily biased and an assessment of how such biases affect SDM predictions, and therefore, our interpretation of animal distributions is lacking. We generated simulated tracks based on predetermined environmental values for a random predator and a central place forager, and then sampled positions from those tracks based on a combination of five common biases in tracking datasets: (a) tagging location; (b) tracking device; (c) data gaps within tracks; (d) premature tag detachment (or failure) and (e) different processing methods. We then used 240 combinations of the resulting biased simulated datasets to develop binomial generalised linear (GLM) and additive (GAM) models to estimate habitat suitability in different environmental sets (cool deep, cool coastal, warm deep and warm coastal environments). Our results show that tagging location and length of tracks have the largest effects in decreasing model performance, but that these biases can be overcome by adding a small percentage of additional, relatively less biased tracks to the dataset. In comparison, the effects from all other biases were almost negligible, including for low resolution tracking datasets for which sufficient tracks are available. We also highlight the need for a cautionary approach when using processing methods that can introduce other biases (e.g. interpolated locations). Similar trends were obtained for the random predator and the central place forager, but with relatively lower model performance for the latter. We provide evidence that even non‐GPS tracking datasets can be readily used to improve the knowledge of large‐scale space use by species without the need for detailed processing and tracking reconstruction. This is especially relevant in the current context of rapid increase in data acquisition and the urgent need to address the large spatial scale ecological consequences of global change.

Perceptual bias and technical metapictures: critical machine vision as a humanities challenge

In many critical investigations of machine vision, the focus lies almost exclusively on dataset bias and on fixing datasets by introducing more and more diverse sets of images. We propose that machine vision systems are inherently biased not only because they rely on biased datasets but also because their perceptual topology, their specific way of representing the visual world, gives rise to a new class of bias that we call perceptual bias. Concretely, we define perceptual topology as the set of those inductive biases in machine vision systems that determine its capability to represent the visual world. Perceptual bias, then, describes the difference between the assumed “ways of seeing” of a machine vision system, our reasonable expectations regarding its way of representing the visual world, and its actual perceptual topology. We show how perceptual bias affects the interpretability of machine vision systems in particular, by means of a close reading of a visualization technique called “feature visualization”. We conclude that dataset bias and perceptual bias both need to be considered in the critical analysis of machine vision systems and propose to understand critical machine vision as an important transdisciplinary challenge, situated at the interface of computer science and visual studies/Bildwissenschaft.

Sampbias, a method for quantifying geographic sampling biases in species distribution data

Geo‐referenced species occurrences from public databases have become essential to biodiversity research and conservation. However, geographical biases are widely recognized as a factor limiting the usefulness of such data for understanding species diversity and distribution. In particular, differences in sampling intensity across a landscape due to differences in human accessibility are ubiquitous but may differ in strength among taxonomic groups and data sets. Although several factors have been described to influence human access (such as presence of roads, rivers, airports and cities), quantifying their specific and combined effects on recorded occurrence data remains challenging. Here we present sampbias, an algorithm and software for quantifying the effect of accessibility biases in species occurrence data sets. sampbias uses a Bayesian approach to estimate how sampling rates vary as a function of proximity to one or multiple bias factors. The results are comparable among bias factors and data sets. We demonstrate the use of sampbias on a data set of mammal occurrences from the island of Borneo, showing a high biasing effect of cities and a moderate effect of roads and airports. sampbias is implemented as a well‐documented, open‐access and user‐friendly R package that we hope will become a standard tool for anyone working with species occurrences in ecology, evolution, conservation and related fields.

Comparing Fingerprints for Ligand-Based Virtual Screening: A Fast and Scalable Approach for Unbiased Evaluation.

Ligand-based virtual screening is a useful tool for drug and probe discovery due to its high accessibility and scalability. The recent identification of bias in many data sets that were used in performance evaluation, quantified by the asymmetric validation embedding (AVE) score, has prompted the reanalysis of models to determine which performs best. Based on the understanding that ligand data are made up of blocks of highly correlated instances, we introduce a technique that quickly generates splits with AVE distributed close to zero using a combination of clustering and removal of the most biased data. We used our technique to compare the performance of the Morgan and CATS fingerprints and show that, after debiasing, the implementation of the CATS fingerprint performs significantly better. The code to replicate these results and perform low-bias splits is available at https://github.com/ljmartin/fp_low_ave.

A smart tool wear prediction model in drilling of woven composites

Undetected tool wear during drilling of woven composites can cause laminate damage and fiber pull-out and fuzzing, causing subsurface damage. This diminishes the life of the produced part under fatigue loads. Thus, the producing of proper and reliable holes in woven composites requires accurate monitoring of the cutting tool wear level to safeguard the machined parts and increase process productivity and profitability. Available tool condition monitoring (TCM) systems mainly require long development lead time and extensive experimental efforts to predict the tool wear within predefined values of cutting conditions. The changes in these values require system relearning. Therefore, developing of a smart generalized TCM system that can accurately predict tool wear based on unlearned data during drilling of woven composite plates is crucial. In this work, an attempt was presented and discussed to predict the tool wear in drilling of woven composite plates at different and wide range of cutting conditions based on the drilling forces using biased learning data. A generalized heuristic model was proposed to accurately predict tool wear value. The performance of the proposed model was benchmarked with respect to four machine learning techniques namely regression tree, support vector machine (SVM), Gaussian process regression (GPR), and artificial neural network (ANN). Extensive experimental validation tests have showed that the GPR model has offered the lowest prediction error based on a reduced biased learning dataset, which represents 50% reduction in learning efforts compared with available literature. However, the developed heuristic model showed a comparable accuracy using significantly less learning efforts.

Learning adaptive geometry for unsupervised domain adaptation

Unsupervised domain adaptation is an effective approach to solve the problem of dataset bias. However, most existing unsupervised domain adaptation methods assume that the geometry structures of data distributions are similar in the source and target domains. This assumption is invalid in many practical applications, because the training and test datasets usually differ in the variability modes and/or variation degrees. This paper handles the problem of inconsistent geometries by aligning both data representations and geometries. To overcome the lack of target labels in aligning geometries, this paper proposes learning the adaptive geometry that is derived from the domain-shared label space. Source and target geometries are aligned by constraining them with the unified criteria of the adaptive geometry. Combining the adaptive geometry learning and adversarial learning techniques, we develop a geometry-aware dual-stream network to learn the geometry-aligned representations. Experimental results show that our method achieves good performance on cross-dataset recognition tasks.

Controlling biases and diversity in diverse image-to-image translation

The task of unpaired image-to-image translation is highly challenging due to the lack of explicit cross-domain pairs of instances. We consider here diverse image translation (DIT), an even more challenging setting in which an image can have multiple plausible translations. This is normally achieved by explicitly disentangling content and style in the latent representation and sampling different styles codes while maintaining the image content. Despite the success of current DIT models, they are prone to suffer from bias. In this paper, we study the problem of bias in image-to-image translation. Biased datasets may add undesired changes (e.g. change gender or race in face images) to the output translations as a consequence of the particular underlying visual distribution in the target domain. In order to alleviate the effects of this problem we propose the use of semantic constraints that enforce the preservation of desired image properties. Our proposed model is a step towards unbiased diverse image-to-image translation (UDIT), and results in less unwanted changes in the translated images while still performing the wanted transformation. Experiments on several heavily biased datasets show the effectiveness of the proposed techniques in different domains such as faces, objects, and scenes.

The Anesthesia Patient Safety Foundation Stoelting Conference 2019: Perioperative Deterioration-Early Recognition, Rapid Intervention, and the End of Failure-to-Rescue.

The Anesthesia Patient Safety Foundation Stoelting Conference 2019: Perioperative Deterioration-Early Recognition, Rapid Intervention, and the End of Failure-to-Rescue.

Hate speech detection and racial bias mitigation in social media based on BERT model.

Disparate biases associated with datasets and trained classifiers in hateful and abusive content identification tasks have raised many concerns recently. Although the problem of biased datasets on abusive language detection has been addressed more frequently, biases arising from trained classifiers have not yet been a matter of concern. In this paper, we first introduce a transfer learning approach for hate speech detection based on an existing pre-trained language model called BERT (Bidirectional Encoder Representations from Transformers) and evaluate the proposed model on two publicly available datasets that have been annotated for racism, sexism, hate or offensive content on Twitter. Next, we introduce a bias alleviation mechanism to mitigate the effect of bias in training set during the fine-tuning of our pre-trained BERT-based model for hate speech detection. Toward that end, we use an existing regularization method to reweight input samples, thereby decreasing the effects of high correlated training set’ s n-grams with class labels, and then fine-tune our pre-trained BERT-based model with the new re-weighted samples. To evaluate our bias alleviation mechanism, we employed a cross-domain approach in which we use the trained classifiers on the aforementioned datasets to predict the labels of two new datasets from Twitter, AAE-aligned and White-aligned groups, which indicate tweets written in African-American English (AAE) and Standard American English (SAE), respectively. The results show the existence of systematic racial bias in trained classifiers, as they tend to assign tweets written in AAE from AAE-aligned group to negative classes such as racism, sexism, hate, and offensive more often than tweets written in SAE from White-aligned group. However, the racial bias in our classifiers reduces significantly after our bias alleviation mechanism is incorporated. This work could institute the first step towards debiasing hate speech and abusive language detection systems.