Identifying acute illness phenotypes via deep temporal interpolation and clustering network on physiologic signatures

Abstract

Using clustering analysis for early vital signs, unique patient phenotypes with distinct pathophysiological signatures and clinical outcomes may be revealed and support early clinical decision-making. Phenotyping using early vital signs has proven challenging, as vital signs are typically sampled sporadically. We proposed a novel, deep temporal interpolation and clustering network to simultaneously extract latent representations from irregularly sampled vital signs and derive phenotypes. Four distinct clusters were identified. Phenotype A (18%) had the greatest prevalence of comorbid disease with increased prevalence of prolonged respiratory insufficiency, acute kidney injury, sepsis, and long-term (3-year) mortality. Phenotypes B (33%) and C (31%) had a diffuse pattern of mild organ dysfunction. Phenotype B's favorable short-term clinical outcomes were tempered by the second highest rate of long-term mortality. Phenotype C had favorable clinical outcomes. Phenotype D (17%) exhibited early and persistent hypotension, high incidence of early surgery, and substantial biomarker incidence of inflammation. Despite early and severe illness, phenotype D had the second lowest long-term mortality. After comparing the sequential organ failure assessment scores, the clustering results did not simply provide a recapitulation of previous acuity assessments. This tool may impact triage decisions and have significant implications for clinical decision-support under time constraints and uncertainty.

Figure 1

Schematic representation of deep temporal interpolation and clustering network architecture. (A) The detailed architecture of deep interpolation network, specifically tailored for handling sparse and irregularly sampled time-series data. $t_i$ is the time point of raw time-series data, and $x_i$ is the corresponding observed value. $r_i$ is the reference time point of interpolated time-series data, and $({\chi }_i,{\tau }_i,{\lambda }_i)$ represents the three different channel outputs from interpolation network capturing the smooth trends, transient and local observation intensity respectively. $h_i$ is the hidden status from decoder network at reference time point $r_i$. (B) The full architecture of deep temporal interpolation and clustering network, designed to concurrently extract the feature representation and determine cluster assignments.

Figure 2

Vital sign representations across identified phenotypes. (A) Distribution of the vital signs recorded within the initial six hours following hospital admission. (B) Visualization of initial phenotypes, as assigned by the pre-trained deep temporal interpolation network, without the integration of the clustering network. The t-distributed stochastic neighbor embedding (t-SNE) technique was utilized to reduce the original 128-dimensional vital sign representations to two dimensions. Each dot signifies an individual patient, with separate colors indicating different phenotypes. (C) Visualization of final phenotypes, as assigned by the deep temporal interpolation and clustering network utilizing the t-SNE technique. The network simultaneously learns feature representation and cluster assignments, thus facilitating clustering-friendly representation learning.

Figure 3

Contributions of vital signs to assigned clusters. The pairwise phenotype comparisons of vital sign values, which have been standardized to a mean of 0 and standard deviation of 1. The comparison reveals that oxygen saturation and temperature contribute the least to the differences between phenotypes. Conversely, respiratory rate and heart rate exhibit considerable variation across all phenotypes, with the exception of phenotypes C and D, where these vital signs appear more consistent. Temp: temperature; SpO2: peripheral capillary oxygen saturation; DBP: diastolic blood pressure; SBP: systolic blood pressure; RR: respiratory rate; HR: heart rate.

Figure 4

Survival curves and adjusted Cox proportional hazards modeling. (A) The survival curves for each phenotype, considering adjustments for both comorbidities and demographic information. (B) The adjusted Cox proportional hazards models, incorporating comorbidities and demographic information into the analysis. CCI: Charlson Comorbidity Index.