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

Area of Science:

• Biomedical Informatics
• Critical Care Medicine
• Machine Learning

Background:

• Phenotyping patients using early vital signs is challenging due to sporadic data sampling.
• Identifying unique patient phenotypes can support early clinical decision-making and reveal distinct pathophysiological signatures.

Purpose of the Study:

• To develop and validate a novel deep temporal interpolation and clustering network for phenotyping patients using irregularly sampled vital signs.
• To identify distinct patient phenotypes based on early vital sign data and analyze their clinical outcomes.

Main Methods:

• A novel deep temporal interpolation and clustering network was developed to extract latent representations from irregularly sampled vital signs.
• The network was used to derive patient phenotypes from vital sign data.
• Identified clusters were analyzed for prevalence, comorbid diseases, organ dysfunction, and mortality.

Main Results:

• Four distinct patient phenotypes (A, B, C, D) were identified, comprising 18%, 33%, 31%, and 17% of the cohort, respectively.
• Phenotype A showed high comorbidity, organ dysfunction, and long-term mortality. Phenotype D exhibited early hypotension and inflammation but lower long-term mortality.
• Phenotype C demonstrated favorable outcomes, while Phenotype B had mixed short-term and long-term results.

Conclusions:

• The deep learning model effectively identifies distinct patient phenotypes from vital signs, offering insights beyond traditional scoring systems.
• These phenotypes have unique pathophysiological characteristics and clinical outcomes, potentially impacting triage and clinical decision support in critical care.
• This approach holds promise for improving patient management under time constraints and uncertainty.