The article introduces the problem of temporal phenotyping, which aims to discover hidden temporal patterns in electronic health records (EHR) data. It proposes a new tensor decomposition model called SWoTTeD that can extract temporal phenotypes, which are arrangements of medical events over time, unlike existing methods that only capture daily phenotypes.
The key highlights are:
SWoTTeD extends the classic tensor decomposition framework to handle the temporal dimension of EHR data. It decomposes an irregular third-order tensor into a set of temporal phenotypes and patient pathways that can accurately reconstruct the input data.
The model incorporates sparsity and non-succession regularization terms to enhance the interpretability and meaningfulness of the extracted phenotypes.
Experiments on both synthetic and real-world datasets show that SWoTTeD outperforms state-of-the-art tensor decomposition methods in terms of reconstruction accuracy and noise robustness. The qualitative analysis also demonstrates that the discovered phenotypes are clinically meaningful.
The authors provide an open-source, well-documented, and efficient implementation of SWoTTeD, making it accessible for further research and practical applications.
A case study on a COVID-19 dataset from the Greater Paris University Hospitals illustrates the utility of temporal phenotypes in understanding care pathways.
Egy másik nyelvre
a forrásanyagból
arxiv.org
Mélyebb kérdések