Analysis of the impact of errors made during health data collection using mobile phones: exploring error modeling and automatic diagnosis

Abstract

Mobile phones are near ubiquitous, and can be easily used to gather and store health data in remote or low resource settings. There exist many systems for supporting such data gathering, including Commcare, Frontline SMS, and OpenData Kit. Survey and health data is often collected by community health workers and frequently includes errors, due to mistakes, challenges with the input interface, systematic error or neglect [1,5]. Automatic detection of errors is important because of its potential impact on aggregate health statistics, and on individual patient treatment. In some important cases, such as tuberculosis diagnosis and monitoring, the space of possible medical diagnoses will generally be significantly smaller than the possible set of symptoms recorded. This suggests that it may be possible to build diagnostic systems whose recommendations are fairly robust to errors in the recorded patient symptoms.