In general, the human-readable rule refers to data shown in a format easily read by most humans - normally this is in the form of IF...THEN rules. This is the most convenient way for physicians to express their knowledge in medical diagnosis. In particular, if learned diagnostic rules can be presented in such a form, physicians are much more likely to trust and believe the consequent diagnoses. This paper investigates the performances of existing state-of-the-art classification algorithms, mainly rule induction and tree algorithms, on benchmark problems in medical data mining. The findings indicate that certain algorithms are better for generating rules that are both accurate and short; these algorithms are recommended for further research towards the goal of improved accuracy and readability in medical data mining. © 2009 Springer Science+Business Media, LLC.