Learning to diagnose from scratch by exploiting dependencies among labels

The field of medical diagnostics contains a wealth of challenges which resemble classical machine learning problems; practical constraints, , complicate the translation of these endpoints naively into classical. Many tasks in radiology, for example, are largely problems of-label classification wherein medical images are interpreted to indicate present or suspected pathologies. Clinical settings drive the for high accuracy simultaneously across a multitude of pathological and greatly limit the utility of tools which consider only a subset. issue is exacerbated by a general scarcity of training data and maximizes need to extract clinically relevant features from available samples -- without the use of pre-trained models which may carry forward biases from tangentially related tasks. We present and evaluate a solution to these constraints in using LSTMs to leverage among target labels in predicting 14 pathologic patterns from x-rays and establish state of the art results on the largest publicly chest x-ray dataset from the NIH without pre-training. Furthermore, propose and discuss alternative evaluation metrics and their relevance in practice.