Unsupervised phenotyping of sepsis using nonnegative matrix factorization of temporal trends from a multivariate panel of physiological measurements

Abstract

Background: Sepsis is a highly lethal and heterogeneous disease. Utilization of an unsupervised method may identify novel clinical phenotypes that lead to targeted therapies and improved care.

Methods: Our objective was to derive clinically relevant sepsis phenotypes from a multivariate panel of physiological data using subgraph-augmented nonnegative matrix factorization. We utilized data from the Medical Information Mart for Intensive Care III database of patients who were admitted to the intensive care unit with sepsis. The extracted data contained patient demographics, physiological records, sequential organ failure assessment scores, and comorbidities. We applied frequent subgraph mining to extract subgraphs from physiological time series and performed nonnegative matrix factorization over the subgraphs to derive patient clusters as phenotypes. Finally, we profiled these phenotypes based on demographics, physiological patterns, disease trajectories, comorbidities and outcomes, and performed functional validation of their clinical implications.

Results: We analyzed a cohort of 5782 patients, derived three novel phenotypes of distinct clinical characteristics and demonstrated their prognostic implications on patient outcome. Subgroup 1 included relatively less severe/deadly patients (30-day mortality, 17%) and was the smallest-in-size group (n = 1218, 21%). It was characterized by old age (mean age, 73 years), a male majority (male-to-female ratio, 59-to-41), and complex chronic conditions. Subgroup 2 included the most severe/deadliest patients (30-day mortality, 28%) and was the second-in-size group (n = 2036, 35%). It was characterized by a male majority (male-to-female ratio, 60-to-40), severe organ dysfunction or failure compounded by a wide range of comorbidities, and uniquely high incidences of coagulopathy and liver disease. Subgroup 3 included the least severe/deadly patients (30-day mortality, 10%) and was the largest group (n = 2528, 44%). It was characterized by low age (mean age, 60 years), a balanced gender ratio (male-to-female ratio, 50-to-50), the least complicated conditions, and a uniquely high incidence of neurologic disease. These phenotypes were validated to be prognostic factors of mortality for sepsis patients.

Conclusions: Our results suggest that these phenotypes can be used to develop targeted therapies based on phenotypic heterogeneity and algorithms designed for monitoring, validating and intervening clinical decisions for sepsis patients.

Keywords: Clustering; Frequent subgraph mining; Gradient boosting machine; Intensive care unit; Nonnegative matrix factorization; Phenotyping; Physiological measurements; Sepsis; Unsupervised learning.

Fig. 1

Subgroup 1 trend group selected from representative frequent subgraphs of standardized physiological variable values over measurement period of six time windows

Fig. 2

Subgroup 2 trend group selected from representative frequent subgraphs of standardized physiological variable values over measurement period of six time windows

Fig. 3

Subgroup 3 trend group selected from representative frequent subgraphs of standardized physiological variable values over measurement period of six time windows

Fig. 4

7-Day sofa score charts

Fig. 5

Subgraph extraction from time series graph

Fig. 6

Frequent subgraph mining from subgraph corpus

Fig. 7

Flow diagram for NMF model construction, Sepsis-3 subgroup identification, and associated functional validations

Fig. 8

Gradient boosting machine ROC curve for patient group membership classification on frequent subgraphs—subgroup 1

Fig. 9

Gradient boosting machine ROC curve for patient group membership classification on frequent subgraphs—subgroup 2

Fig. 10

Gradient boosting machine ROC curve for patient group membership classification on frequent subgraphs—subgroup 3

Fig. 11

Gradient boosting machine ROC curve for patient mortality model on frequent subgraphs