SEPTRON : Predição de sepse usando transformer, ontologia e algoritmo bioinspirado

Abstract

The use of Machine Learning (ML) techniques for sepsis prediction has evolved globally, but the selection of clinically relevant variables, essential for building robust predictive models, remains one of the main challenges. This aspect is particularly important for patients with suspected sepsis, who require intensive monitoring during the first hours after admission to Intensive Care Units (ICUs), in order to enable early diagnosis and treatment. In this context, this work proposes the model SEpsis Prediction with TRansformer, Ontology and BioiNspired Algorithm (SEPTRON), built on a Transformer architecture adapted for structured data. The model uses variables initially selected from a conceptual ontological model of sepsis and further refined by a Genetic Algorithm (GA). The contributions include: (i) a structured and replicable pipeline, covering the entire process from data extraction to model evaluation; (ii) creation of the curated dataset MIMIC-IV-EXT-ONTO-SEPSE, derived from the MIMIC-IV database; (iii) Transformer architecture adapted for structured data; and (iv) evidence of the relevance of the chosen clinical variables. The comparative evaluation focused on the performance of the Transformer against four classic models: Logistic Regression, Random Forest, XGBoost, and Long Short-Term Memory (LSTM). Using 4-hour time windows over the first 72 hours of ICU stay, the Transformer with all 17 clinical variables achieved superior performance (AUC-ROC 0.85; recall 0.83). In turn, the SEPTRON, which integrates the Genetic Algorithm to reduce the set of clinical variables from 17 to 8, maintained competitive performance (AUC-ROC 0.83; recall 0.76), combining robustness, interpretability, and computational feasibility. This effectiveness was reinforced by the temporal analysis of the evolution of sepsis probability over the 72 hours. It is concluded, therefore, that SEPTRON represents a relevant proposal to the field of digital health, providing a predictive model capable of anticipating the risk of sepsis in ICU with a good balance between performance, interpretability, and clinical applicability.