Link prediction is an unbalanced early retrieval problem, whose goal is to prioritize a small cohort of positive links on top of a list largely populated by unlabelled links. Differently from binary classification, here the evaluation focuses on how the predictor prioritizes the positive class because, in practice, a negative class does not exist. Previous studies explained that AUC-ROC is not apt for unbalanced class problems and is misleading for early retrieval problems, therefore standard AUC-ROC is not appropriate for evaluation of link prediction. However, some scholars argue that an AUC-ROC like evaluation accounting for the relative positioning of the few positive links among the vastness of unlabelled links remains a valid concept to pursue. Here we propose the area under the magnified ROC (AUC-mROC), a new measure that adjusts the standard AUC-ROC to work also for unbalanced early retrieval problems such as link prediction.

Cancer is a disease caused by the abnormal growth of cells in different parts of body is one of the top causes of death globally. Microarray gene expression data plays a critical role in the identification and classification of cancer tissues. Due to recent advancements in Machine Learning (ML) techniques, researchers are analyzing gene expression data using a variety of such techniques to model the progression rate & treatment of cancer patients with great effect. But high dimensionality alongside the presence of highly correlated columns in gene expression datasets leads to computational difficulties. This paper aims to propose the use of ML classification techniques- Linear Discriminant Analysis (LDA) & Random Forest (RF) for classifying five types of cancer (breast cancer, kidney cancer, colon cancer, lung cancer and prostate cancer) based on high dimensional microarray gene expression data. Principal component analysis (PCA) was used for dimensionality reduction, and principal component scores of the raw data for classification. Six distinct categorization performance measures were used to evaluate these approaches; RF method provided us with higher accuracy than LDA method. The method and results of this article should be helpful to researchers who are dealing with many genes in microarray data.

A recent study claimed that genome-wide polygenic scores (GPSs) for five common diseases could identify individuals with risk equivalent to monogenic mutations. Receiver operator curve analyses were reported to have areas under the curve (AUCs) ranging from 0.63 for inflammatory bowel disease up to 0.81 for coronary artery disease (CAD) but these models also included age and sex, themselves strong predictors of risk. The GPS for CAD identified 8% of the population at threefold increased risk, which it was claimed was comparable to the excess risk from monogenic mutations. In the present study attempts were made to model the distribution of the GPS for CAD to match the information provided. These models were based on the reported distribution of prevalence by centile of GPS and on the distribution of GPS in controls and cases and were fitted to the reported results using linear approximations to the distributions and using simulations of a liability-threshold model. It was impossible to produce a compatible model in which the GPS produced an AUC as high as 0.81 and the most plausible estimate was that the true AUC was only 0.65. The reported distributions of the GPS in cases and controls overlap so much that they are not compatible with an AUC of 0.7 or higher. The AUC of the GPS for these diseases is modest. Furthermore, the literature robustly demonstrates that true CAD risk associated with monogenic mutations is much higher than the threefold increase which is predicted by the GPS. Together, these findings cast doubt on the clinical utility of the GPS.

In the literature, the question of the amount of data necessary and sufficient to validate different models of the occurrence of risk of adverse events for patients or the classification of the presence or absence of pathological features has been repeatedly raised. In the presented study, we propose a new approach to determine the necessary and enough studies for validation of medical software based on artificial intelligence technology, whose main task is to classify medical X-rays according to the presence of normality and pathology. It is shown that for several studies in a dataset, when AUC ROC has maximum heterogeneity, it varies depending on the balance of "norm"/"pathology" classes. Thus, for a balance of "normal"/"abnormality", where 90% is "normal" and 10% is "abnormality", maximum heterogeneity is achieved for 190 studies, for a balance of 80% ("normal")/20% ("abnormality") for 80 studies, for a balance of 70% ("norm")/30% ("abnormality") - 120 trials, for a balance of 60% ("norm")/40% ("abnormality") the maximum heterogeneity is reached at 110 trials, and for a balance of 50% ("norm")/50% ("abnormality") - at 70 trials. The obtained data are in good agreement with the previous results. They allow us to determine a sufficient (necessary) number of studies in the dataset to perform an unbiased assessment of AUC ROC.