Infection Diagnosis using Biomedical Signals in Small Data Scenarios

Being able to detect an infection at an early stage is a clinical problem of the utmost importance. An infection not diagnosed on time might not only severely affect the infected patient's health, but also to spread and start a focus of contagion to other people. In this paper, we propose a clinical decision support system to automatically diagnose infections using physiological signals from the patients. The focus of the system is put on being able to deal with very small amounts of data (one aggregated record per patient and day), which eases the potential of the system in environments with low resources. Data has been acquired between April 2018 and January 2019 in two nursing homes in Spain, where nurses had also tested patients for infections. Machine learning models have then been created by aggregating measurements from days prior to the infection (lead) and after the infection started (lag) in order to generate features. The best model attained reports an AUROC of 0.722, using data from up to two days after the infection started. Interestingly, an AUROC of up to 0.692 is achieved when infection prognosis is considered; i.e., using only measurements prior to the manual recording of the infection to compose the dataset.