TEST AMPLIFICATION FOR MEDICAL APPS IMPLEMENTING LINEARLY-APPROXIMATABLE FUNCTIONS

Mobile (smartphone) apps are increasingly used in medical, health, and fitness settings, but their reliability can be poor, which has negative healthcare consequences. Researchers have started to scrutinize the correctness and reliability of such apps, but the techniques used are unsatisfactory. For example, prior work has used ad-hoc techniques to discover some errors and incorrect calculations, but we still lack automated procedures that systematically look for errors. We introduce a novel approach that provides test “amplification”: given a few test cases for an app, we use a linear approximation to model app behavior, compute an error function, and then look for its maxima. The error function maxima are further test cases that exhibit higher errors. Our approach can generally be applied to any calculations based on linearly-approximatable functions. We have applied our approach to 54 Android apps. We have primarily focused on apps computing the Basal Metabolic Rate (BMR). While manual testing discovered errors of at most 0.3%, amplification found errors as large as 5.4%. We also show how our approach can perform effective test amplification, from 2.4% to 8%, for recently-found issues in Android apps that calculate the Body-Surface Area.