P4896Automated classification of health apps into medical specialties by using text analytics

Abstract

Introduction: Wide adoption of smartphones in the general population and proliferation of health-related apps open new possibilities for patient engagement, health literacy and prevention. However, among the >165,000 health apps available on the app stores, only a limited number have been tested for reliability and efficacy. Thus, finding the right app for a specific need and identifying the relevant features is a challenging task and a barrier to physicians' confidence in these tools. Purpose(s): To develop automated methods, based on text analytics, for extracting information from app webpages to classify mobile applications into medical specialties and to identify app features, as a basis for support tools aimed at patients and healthcare professionals. Method(s): Custom software was developed to use the HTML code in the apps webpage from the US iTunes store (as of Dec 20, 2016) to automatically identify specific tags and retrieve relevant attributes. It was applied to all apps in the Medical (M) and Health&Fitness (H&F) categories, resulting in a database including, among other attributes, the App ID, Name, Developer Name, App Description, Contacts, and URL. Pre-processing was performed to remove non-English apps and non-ASCII characters. Text analytics was applied using AlchemyAPI (IBM) and MetaMap (NLM, National Library of Medicine) to automatically classify the most appropriate medical specialty (among a list of 22) based on a relevance score (i.e., keywords and concepts relevant to the medical specialty) and to identify the app promoter. To test the classification accuracy, 900 apps' webpages out of >80,000 were manually reviewed, and results were compared with the automated ones. Result(s): A total of 33858 M and 51994 H&F apps' webpages were analysed to build the apps' database. After pre-processing, text analytics classified 21205 M and 34821 H&F apps. Overall, 57% of apps had no medical content (e.g., fitness, lifestyle, games), whereas the ten most relevant medical specialties were Nutrition (12%), Mental Health (8%), Diabetes Care (5%), Gynaecology & Obstetrics (5%), Cardiology (4%) Orthopaedics (4%), Neurology (4%), General Medicine (3%), Dermatology (2%), and Audiology (2%). Automated identification of the app promoter resulted in: 74% Independent Developers, 25% Manufacturers/Software Houses, 0.3% Government Services, 0.2% Scientific/Educational Organizations, 0.2% Healthcare Providers, 0.2% by Publishers, and 0.1% by Drug Companies. Comparison with manual classification resulted in accuracy of 80%. Conclusion(s): Text analytics could be used to automatically identify the medical specialty and promoter of apps in online stores, thus providing potential filtering capabilities when searching for specific health apps. This approach might constitute the basis for novel tools for app characterization to support patients and healthcare professionals in informed apps selection and adoption.