Illuminating the Path to Enhanced Resilience of Machine Learning Models Against the Shadows of Missing Labels

The sensitivity of state-of-the-art supervised classification models is compromised by contamination-prone biomedical datasets, which are vulnerable to the presence of missing or erroneous labels (i.e., inliers). Starting from codon frequencies, electrocardiogram signals, biomarkers, morphological features, and patient questionnaires, we attempted to cover a wide range of typical biomedical databases exposed to the risk of missing data labeled as negative values (inlier contamination). In some very niche fields, such as image recognition, missing labels have received a lot of attention, but in biomedical and clinical research, where outliers are almost systematically filtered, inliers have remained orphans. Our study introduced a pragmatic and innovative automated methodology that consists of upcycling one-class semi-supervised anomaly detection (OCSSAD) models for filtering potential inliers in training datasets. Five OCSSAD and two ensemble methods were benchmarked on 6 databases with 10 different contamination levels and 10 random samples, achieving an average Matthews correlation coefficient (MCC) of 78 $\pm$ 17% in validation, whereas 22 supervised classifiers achieved an average MCC score of 81 $\pm$ 9% trained with the complete and uncontaminated trainset.Therefore, by filtering the training set with an isolation forest, the average resilience to inliers of 22 tested Machine Learning models increased from 69 $\pm$ 11% to 95 $\pm$ 1%, including neural networks and gradient-boosting methods. Taken together, our study showcased the efficacy of our versatile approach in enhancing the resilience of Machine Learning models and highlighted the importance of accurately addressing the inliers challenge in the domains of medical and Life Sciences.