Published Paper Details:

In some clinical applications, the performance of a machine learning algorithm may be negatively impacted by patient incorrect label uncertainty during training for a supervised learning job. For instance, because of uncertainty in the patient's condition or the unreliability of the diagnostic criteria, even clinical professionals may be less confident when making a medical diagnosis for some patients. Thus, certain examples utilised in algorithm training could have incorrect labels applied to them, which would negatively impact the algorithm's performance. In certain situations, professionals might be able to measure their diagnostic uncertainty, though. In order to account for such clinical diagnostic ambiguity while training an algorithm to predict individuals who develop acute respiratory distress syndrome (ARDS), we provide a robust technique that uses support vector machines (SVM). A condition of the severely sick known as ARDS is identified using clinical criteria that are known to be unreliable. Our method of representing ambiguity in the diagnosis of ARDS involves assigning a graded weight of confidence to every training label. In order to limit overfitting, we also employed a unique time-series sampling technique to address the issue of intercorrelation among the longitudinal clinical data from each patient utilised in model training. Based on preliminary results, we may compare our technique that takes into account the uncertainty of training labels with a traditional SVM algorithm and obtain a significant improvement in the system's ability to diagnose patients with ARDS on a hold-out sample.