Prediction of mortality in intensive care unit (ICU) patients is an important task in critical care medicine. Prior work in creating mortality risk models falls into two major categories: domain-expert-created scoring systems, and black box machine learning (ML) models. Both of these have disadvantages: black box models are unacceptable for use in hospitals, whereas manual creation of models (including hand-tuning of logistic regression parameters) relies on humans to perform high-dimensional constrained optimization, which leads to a loss in performance. In this work, we bridge the gap between accurate black box models and hand-tuned interpretable models. We build on modern interpretable ML techniques to design accurate and interpretable mortality risk scores. We leverage the largest existing public ICU monitoring datasets, namely the MIMIC III and eICU datasets. By evaluating risk across medical centers, we are able to study generalization across domains. In order to customize our risk score models, we develop a new algorithm, GroupFasterRisk, which has several important benefits: (1) it uses hard sparsity constraint, allowing users to directly control the number of features; (2) it incorporates group sparsity to allow more cohesive models; (3) it allows for monotonicity correction on models for including domain knowledge; (4) it produces many equally-good models at once, which allows domain experts to choose among them. GroupFasterRisk creates its risk scores within hours, even on the large datasets we study here. GroupFasterRisk's risk scores perform better than risk scores currently used in hospitals, and have similar prediction performance to black box ML models (despite being much sparser). Because GroupFasterRisk produces a variety of risk scores and handles constraints, it allows design flexibility, which is the key enabler of practical and trustworthy model creation.

We introduce Epsilon*, a new privacy metric for measuring the privacy risk of a single model instance prior to, during, or after deployment of privacy mitigation strategies. The metric requires only black-box access to model predictions, does not require training data re-sampling or model re-training, and can be used to measure the privacy risk of models not trained with differential privacy. Epsilon* is a function of true positive and false positive rates in a hypothesis test used by an adversary in a membership inference attack. We distinguish between quantifying the privacy loss of a trained model instance, which we refer to as empirical privacy, and quantifying the privacy loss of the training mechanism which produces this model instance. Existing approaches in the privacy auditing literature provide lower bounds for the latter, while our metric provides an empirical lower bound for the former by relying on an (${\epsilon}$, ${\delta}$)-type of quantification of the privacy of the trained model instance. We establish a relationship between these lower bounds and show how to implement Epsilon* to avoid numerical and noise amplification instability. We further show in experiments on benchmark public data sets that Epsilon* is sensitive to privacy risk mitigation by training with differential privacy (DP), where the value of Epsilon* is reduced by up to 800% compared to the Epsilon* values of non-DP trained baseline models. This metric allows privacy auditors to be independent of model owners, and enables visualizing the privacy-utility landscape to make informed decisions regarding the trade-offs between model privacy and utility.

The unequal representation of different groups in a sample population can lead to discrimination of minority groups when machine learning models make automated decisions. To address these issues, fairness-aware machine learning jointly optimizes two (or more) metrics aiming at predictive effectiveness and low unfairness. However, the inherent under-representation of minorities in the data makes the disparate treatment of subpopulations less noticeable and difficult to deal with during learning. In this paper, we propose a novel adversarial reweighting method to address such \emph{representation bias}. To balance the data distribution between the majority and the minority groups, our approach deemphasizes samples from the majority group. To minimize empirical risk, our method prefers samples from the majority group that are close to the minority group as evaluated by the Wasserstein distance. Our theoretical analysis shows the effectiveness of our adversarial reweighting approach. Experiments demonstrate that our approach mitigates bias without sacrificing classification accuracy, outperforming related state-of-the-art methods on image and tabular benchmark datasets.

Spatial and intensity normalization are nowadays a prerequisite for neuroimaging analysis. Influenced by voxel-wise and other univariate comparisons, where these corrections are key, they are commonly applied to any type of analysis and imaging modalities. Nuclear imaging modalities such as PET-FDG or FP-CIT SPECT, a common modality used in Parkinson's Disease diagnosis, are especially dependent on intensity normalization. However, these steps are computationally expensive and furthermore, they may introduce deformations in the images, altering the information contained in them. Convolutional Neural Networks (CNNs), for their part, introduce position invariance to pattern recognition, and have been proven to classify objects regardless of their orientation, size, angle, etc. Therefore, a question arises: how well can CNNs account for spatial and intensity differences when analysing nuclear brain imaging? Are spatial and intensity normalization still needed? To answer this question, we have trained four different CNN models based on well-established architectures, using or not different spatial and intensity normalization preprocessing. The results show that a sufficiently complex model such as our three-dimensional version of the ALEXNET can effectively account for spatial differences, achieving a diagnosis accuracy of 94.1% with an area under the ROC curve of 0.984. The visualization of the differences via saliency maps shows that these models are correctly finding patterns that match those found in the literature, without the need of applying any complex spatial normalization procedure. However, the intensity normalization -- and its type -- is revealed as very influential in the results and accuracy of the trained model, and therefore must be well accounted.

Systematic literature reviews (SLRs) play an essential role in summarising, synthesising and validating scientific evidence. In recent years, there has been a growing interest in using machine learning techniques to automate the identification of relevant studies for SLRs. However, the lack of standardised evaluation datasets makes comparing the performance of such automated literature screening systems difficult. In this paper, we analyse the citation screening evaluation datasets, revealing that many of the available datasets are either too small, suffer from data leakage or have limited applicability to systems treating automated literature screening as a classification task, as opposed to, for example, a retrieval or question-answering task. To address these challenges, we introduce CSMeD, a meta-dataset consolidating nine publicly released collections, providing unified access to 325 SLRs from the fields of medicine and computer science. CSMeD serves as a comprehensive resource for training and evaluating the performance of automated citation screening models. Additionally, we introduce CSMeD-FT, a new dataset designed explicitly for evaluating the full text publication screening task. To demonstrate the utility of CSMeD, we conduct experiments and establish baselines on new datasets.

Neglecting the effect that decisions have on individuals (and thus, on the underlying data distribution) when designing algorithmic decision-making policies may increase inequalities and unfairness in the long term--even if fairness considerations were taken in the policy design process. In this paper, we propose a novel framework for achieving long-term group fairness in dynamical systems, in which current decisions may affect an individual's features in the next step, and thus, future decisions. Specifically, our framework allows us to identify a time-independent policy that converges, if deployed, to the targeted fair stationary state of the system in the long-term, independently of the initial data distribution. We model the system dynamics with a time-homogeneous Markov chain and optimize the policy leveraging the Markov chain convergence theorem to ensure unique convergence. We provide examples of different targeted fair states of the system, encompassing a range of long-term goals for society and policy makers. Furthermore, we show how our approach facilitates the evaluation of different long-term targets by examining their impact on the group-conditional population distribution in the long term and how it evolves until convergence.

Model extraction attacks are designed to steal trained models with only query access, as is often provided through APIs that ML-as-a-Service providers offer. ML models are expensive to train, in part because data is hard to obtain, and a primary incentive for model extraction is to acquire a model while incurring less cost than training from scratch. Literature on model extraction commonly claims or presumes that the attacker is able to save on both data acquisition and labeling costs. We show that the attacker often does not. This is because current attacks implicitly rely on the adversary being able to sample from the victim model's data distribution. We thoroughly evaluate factors influencing the success of model extraction. We discover that prior knowledge of the attacker, i.e. access to in-distribution data, dominates other factors like the attack policy the adversary follows to choose which queries to make to the victim model API. Thus, an adversary looking to develop an equally capable model with a fixed budget has little practical incentive to perform model extraction, since for the attack to work they need to collect in-distribution data, saving only on the cost of labeling. With low labeling costs in the current market, the usefulness of such attacks is questionable. Ultimately, we demonstrate that the effect of prior knowledge needs to be explicitly decoupled from the attack policy. To this end, we propose a benchmark to evaluate attack policy directly.

Out-of-distribution (OOD) detection has attracted a large amount of attention from the machine learning research community in recent years due to its importance in deployed systems. Most of the previous studies focused on the detection of OOD samples in the multi-class classification task. However, OOD detection in the multi-label classification task, a more common real-world use case, remains an underexplored domain. In this research, we propose YolOOD - a method that utilizes concepts from the object detection domain to perform OOD detection in the multi-label classification task. Object detection models have an inherent ability to distinguish between objects of interest (in-distribution) and irrelevant objects (e.g., OOD objects) in images that contain multiple objects belonging to different class categories. These abilities allow us to convert a regular object detection model into an image classifier with inherent OOD detection capabilities with just minor changes. We compare our approach to state-of-the-art OOD detection methods and demonstrate YolOOD's ability to outperform these methods on a comprehensive suite of in-distribution and OOD benchmark datasets.

In recent years, the industrial sector has evolved towards its fourth revolution. The quality control domain is particularly interested in advanced machine learning for computer vision anomaly detection. Nevertheless, several challenges have to be faced, including imbalanced datasets, the image complexity, and the zero-false-negative (ZFN) constraint to guarantee the high-quality requirement. This paper illustrates a use case for an industrial partner, where Printed Circuit Board Assembly (PCBA) images are first reconstructed with a Vector Quantized Generative Adversarial Network (VQGAN) trained on normal products. Then, several multi-level metrics are extracted on a few normal and abnormal images, highlighting anomalies through reconstruction differences. Finally, a classifer is trained to build a composite anomaly score thanks to the metrics extracted. This three-step approach is performed on the public MVTec-AD datasets and on the partner PCBA dataset, where it achieves a regular accuracy of 95.69% and 87.93% under the ZFN constraint.

In many applications, it is of interest to identify a parsimonious set of features, or panel, from multiple candidates that achieves a desired level of performance in predicting a response. This task is often complicated in practice by missing data arising from the sampling design or other random mechanisms. Most recent work on variable selection in missing data contexts relies in some part on a finite-dimensional statistical model, e.g., a generalized or penalized linear model. In cases where this model is misspecified, the selected variables may not all be truly scientifically relevant and can result in panels with suboptimal classification performance. To address this limitation, we propose a nonparametric variable selection algorithm combined with multiple imputation to develop flexible panels in the presence of missing-at-random data. We outline strategies based on the proposed algorithm that achieve control of commonly used error rates. Through simulations, we show that our proposal has good operating characteristics and results in panels with higher classification and variable selection performance compared to several existing penalized regression approaches in cases where a generalized linear model is misspecified. Finally, we use the proposed method to develop biomarker panels for separating pancreatic cysts with differing malignancy potential in a setting where complicated missingness in the biomarkers arose due to limited specimen volumes.