SCS Undergraduate Thesis Topics
|Jack Paparian||Chris Langmead||An Autoencoder Triaging Algorithm for Acute Pancreatitis|
As more data is collected from clinical studies, predictive medicine will advance to the state where triaging can be automated using raw patient data. For acute pancreatitis, several scoring systems, such as Ranson and APACHE II, are employed clinically to measure the severity of acute pancreatitis; however, these criteria have limited accuracy. Designing more accurate scoring systems is difficult due to the small number of patients typically enrolled in studies and the small percentage of patients with acute pancreatitis whose condition becomes severe. In this thesis, we present scoring systems derived from machine learning classifiers that are trained on both raw patient data and clinical scores. We focus on measuring severity by predicting whether a patient will develop organ failure, the most common cause of death in patients with acute pancreatitis. Using stacked autoencoders, we attempt to learn deep feature representations of organ failure in acute pancreatitis, and we compare the results to those of shallow architectures. Ultimately, we found single autoencoder networks to perform with the best sensitivity (lowest number of false negatives). In addition to our results, we discuss our techniques for compensating for imbalanced data classes and missing patient data.