ABSTRACT
OBJECTIVE: Clinical data on which artificial intelligence (AI) algorithms are trained and tested provide the basis to improve diagnosis or treatment of infectious diseases (ID). We aimed to identify important data for ID research to prioritise efforts being undertaken in AI programmes. METHODS: We searched for 1,000 articlesfrom high-impact ID journals on PubMed, selecting 288 of the latest articles from 10 top journals. We classified them into structured or unstructured data. Variables were homogenised and grouped into the following categories: epidemiology, admission, demographics, comorbidities, clinical manifestations, laboratory, microbiology, other diagnoses, treatment, outcomes and other non-categorizable variables. RESULTS: 4,488 individual variables were collected, from the 288 articles. 3,670 (81.8%) variables were classified as structured data whilst 818 (18.2%) as unstructured data. From the structured data, 2,319 (63.2%) variables were classified as direct-retrievable from electronic health records-whilst 1,351 (36.8%) were indirect. The most frequent unstructured data were related to clinical manifestations and were repeated across articles. Data on demographics, comorbidities and microbiology constituted the most frequent group of variables. CONCLUSIONS: This article identified that structured variables have comprised the most important data in research to generate knowledge in the field of ID. Extracting these data should be a priority when a medical centre intends to start an AI programme for ID. We also documented that the most important unstructured data in this field are those related to clinical manifestations. Such data could easily undergo some structuring with the use of semi-structured medical records focusing on a few symptoms.
