Fig. 3

Visualization of the patient phenoscape. The latent phenotype is visualized in two dimensions using PHATE. In this plot each point represents a patient and the coloring is based on the value of different clinical variables. (A) All patients in the Yale cohort are plotted and colored according to TTD label (top-left), TTD (log-transformed, top-right), heart rate (middle), SpO2 (bottom-left), and GCS (bottom-right). Each point represents a patient. These plots uncover the continuous structure of the patients’ latent phenotype and highlight the correlation between different clinical variables and the time-to-death. The longitudinal variables were transformed into scalar variables using different transformations. (range) computes the average of the five highest observations minus the average of the five lowest observations in the clinical history of the patient. (mean) computes the average of the observations in the clinical history of the patient. (min) computes the average of the five lowest observations. Different transformations extract different patterns from the time series, enabling a finer interpretation of the dynamical patterns for a given clinical variable. For instance, we observed that heart rate (range) correlates with the TTD label, unlike heart rate (mean), suggesting the variation in heart rate is more important than the average value. The visualization of the whole cohort suggests two distincts groups of patients, characterized by high or low TTD. (B) Focus visualization of the identified cluster of patients with TTD<120 min. This uncovers a finer grained structure in this specific cohort of patients. We colored the patients by TTD (log-transformed, top-left), range of heart rate (middle-left), minimum SpO2 (middle-right), average GCS (bottom-left), and BMI (bottom-right). (C) We clustered the patients in the zoomed-in group of patients according to the similarity of their latent phenotype. We obtained three clusters: A, B, and C. Guided by the visualization of panel B, we examined the specific phenotype of patients from cluster A, which appear to have higher TTD than the rest of the patients. We show boxplots and corresponding independent t-tests p-values for difference of means between clusters, for various clinical variables. This analysis characterizes cluster A as a subgroup of patients with high TTD, high range of heart rate, low minimum SpO2, high GCS, as well as low BMI.