Temporal phenotyping for characterisation of hospital care pathways of COVID19 patients

During the COVID19 crisis, Intensive Care Units admitted many patients with breathing disorders up to respiratory insufficiency. The care strategy of such patients was difficult to find and preventing patients to drift away toward a critical situation was one of the first challenge of physicians. In this study, we would like to characterize care pathways of patients that required a mechanical ventilation. The mechanical ventilation is an invasive treatment for the most critical respiratory insufficiencies. Through the analysis of the sequence of cares, we aim at supporting physicians to better understand patients evolution and let them propose new medical procedures to prevent some patients to be ventilated. This article proposes a method which combines a tensor factorization and sequence clustering. The tensor factorization enables to represent the care sequences as a sequence of daily phenotypes. Then, the sequences of phenotypes is clustered to extract typical care trajectories. This method is experimented on real data from Greater Paris university Hospital and is compared to a direct clustering of the sequences. The results show that the outputs are more easily interpretable with the proposed method.