Towards Reducing Diagnostic Errors with Interpretable Risk Prediction