Since diagnostic errors can lead to severe consequences for a patient, medical schools are searching for advanced methodologies to teach and asses diagnostic reasoning. Computer software or actors portraying patients are used to present the symptoms of a disease to the students, and the students' approach in diagnosing that disease are recorded. An expert analyzes that record to classify each approach as being correct or incorrect and then recommends appropriate training for each student. With the goal of automating this classification process, neural network-based supervised algorithms are tested in this investigation. The task here is to reliably classify the diagnosis skills of 200 third-year medical students in diagnosing a back pain case. Each student could acquire information from a set of 22 questions, physical examinations and diagnostic tests to diagnose the disease. The potential of using less than 22 features are also investigated. Classification results provided by neural networks are compared with that provided by two conventional statistical classifiers.