Deep learning-based clustering robustly identified two classes of sepsis with both prognostic and predictive values

Area of Science:

• Genomics
• Bioinformatics
• Critical Care Medicine

Background:

• Sepsis is a complex condition requiring individualized treatment.
• Previous gene expression profiling efforts faced classification instability and lacked robust prediction models.
• Developing a stable, validated model for sepsis subclassification is crucial for clinical utility.

Purpose of the Study:

• To develop a parsimonious 5-gene model for predicting sepsis class membership.
• To validate the model's prognostic and predictive capabilities in external datasets.
• To identify distinct sepsis subclasses with differential mortality and treatment responses.

Main Methods:

• Systematic search of Gene Expression Omnibus and ArrayExpress databases for adult sepsis gene expression data.
• Utilized autoencoder for feature extraction and k-means clustering to identify sepsis classes.
• Employed genetic algorithms (GA) to derive a 5-gene prediction model and validated it on external cohorts.

Main Results:

• Identified two sepsis classes in 1613 patients; Class 1 exhibited higher mortality due to immunosuppression.
• Developed a 5-gene model (C14orf159, AKNA, PILRA, STOM, USP4) with good predictive performance (AUC: 0.707).
• The 5-gene model outperformed Sepsis Response Signature (SRS) endotypes and matched APACHE II score in mortality prediction; differential hydrocortisone response observed between classes.

Conclusions:

• Two distinct sepsis classes with differing mortality rates and hydrocortisone treatment responses were identified.
• A robust 5-gene model was developed for predicting sepsis class membership, supporting personalized medicine approaches.
• Class 1 sepsis is linked to immunosuppression and higher mortality, while Class 2 shows differential response to hydrocortisone.