Machine Learning Based Acute Kidney Injury Sub-Phenotyping with Time Series Serum Creatinine Data

Acute Kidney injury (AKI) is a life-threatening clinical syndrome prevalent in hospitalized patients and is associated with increased mortality and poor patient outcomes. The pathophysiology of AKI is highly divergent. AKI sub-phenotyping can lead to better understanding of the AKI pathophysiology and help develop etiology-based personalized treatment. Most existing sub-phenotyping methods are based on static data, ignoring time-series data like vital signs and lab measurements. To address these limitations, we propose to use the Dynamic Time Warp algorithm to capture information from the full time-series of Serum Creatinine (SCr) measurements for AKI sub-phenotyping. We aimed to discover sub-phenotypes of AKI patients in ICU
$(\mathrm{n}=5781)$
and AKI patients in general units
$(\mathrm{n}=5034)$
. For each patient cohort, we identified 4 sub-phenotypes and analyses showed that they are divided by two major characteristics: AKI stage-l vs AKI stage-2/3 and resolving vs Non-resolving AKI.