Infection prediction using physiological and social data in social environments

The ability of detecting infections at an early stage in clinical environments is an important clinical problem. When an infection is not diagnosed on time, it may not only affect the health of the infected patient, but also spread and infect other people. In this paper, we propose the development of a clinical decision support system (CDSS) for diagnosing infections using clinical signals from patients. This system is designed to be able to cope with small amounts of data (a single record per day and patient), making it convenient for environments under strict constraints (such as low resources or bad connectivity). Additionally, we have incorporated data from external sources, in order to enrich the quality of the models. In particular, we have considered social data arising from web searches, retrieved from Google Trends, as well as weather data. Clinical data was recorded between April 2018 and July 2019 in two nursing homes in Spain and one in Dominican Republic, where nurses had also tested patients for infections. Feature extraction was carried out by aggregating measurements from days before to the infection (lead) and after the infection was detected (lag), and these features were used to train supervised learning models. The best model attained using only clinical data attains an AUROC of 0.734. When data is enriched with external sources, this performance increases up to an AUROC of 0.798. In the case of prognosis (i.e., only measurements before the manual annotation of the infection are used) an AUROC of 0.719 is obtained using only clinical data, and up to 0.757 when combining additional sources of data. In conclusion, the CDSS provides a good recognition performance given the small amounts of data available. This performance can be increased by including social data, which are readily available, and can therefore be useful in scenarios where clinical data acquisition is expensive or unfeasible.