PROJECT TITLE :

Accounting for Label Uncertainty in Machine Learning for Detection of Acute Respiratory Distress Syndrome

ABSTRACT:

When training a machine learning algorithm for a supervised-learning job in some clinical applications, the system's performance may be harmed by ambiguity in the accurate labeling of some patients. Because of uncertainty in the patient's condition or insufficient reliability of the diagnostic criteria, even clinical professionals may have less confidence in assigning a medical diagnosis to some patients. As a result, some cases utilized in algorithm training may be mislabeled, causing the algorithm's performance to suffer. In certain circumstances, though, specialists may be able to quantify their diagnostic uncertainty. When training an algorithm to detect individuals who develop acute respiratory distress syndrome, we provide a robust technique based on support vector machines (SVM) to account for such clinical diagnostic ambiguity (ARDS). ARDS is a severely unwell syndrome that is diagnosed using clinical criteria that are acknowledged to be flawed. Uncertainty in the diagnosis of ARDS is represented by a graded weight of confidence assigned to each training label. In order to avoid overfitting, we employed a unique time-series sampling strategy to handle the problem of intercorrelation among the longitudinal clinical data from each patient used in model training. When we compare our method that accounts for the uncertainty of training labels with a traditional SVM algorithm, preliminary findings show that we can obtain considerable improvement in the performance of the system to predict patients with ARDS on a hold-out sample.