Detección precoz de sepsis (se) y shock séptico (ss) utilizando técnicas de big data, inteligencia artificial y machine learning

Laburpena

Objectives: The aim of this study was to develop and validate predictive models for the detection of severe sepsis (SG) and septic shock (SS) in patients over the age of 14 at a University Hospital, utilizing advanced Big Data (BD), Artificial Intelligence (AI), and Machine Learning (ML) methodologies, and to compare their effectiveness with traditional diagnostic methods. We designed a study to refine the predictive ability of various models to effectively distinguish between patients with and without SG/SS, thus minimizing false positives and negatives. Methodological Design: A retrospective analysis was conducted on patients previously identified and prospectively validated as SG/SS cases by specialists from the Multidisciplinary Sepsis Unit (UMS). This analysis used a variety of data sources from the Electronic Health Record (EHR), both structured and unstructured (including free text and Natural Language Processing, NLP), to construct predictive models. Variables included demographic data, vital and clinical signs, laboratory results, pharmaceutical prescriptions, microbiological reports, triage information, and Urgent Care discharge summaries. Scope of Study: The research encompassed all hospitalization areas, Emergency and ICU departments of the University Hospital Son Llàtzer in Palma de Mallorca, Spain. Patients: All patients over the age of 14 included in the study period. Study Period: The analysis spanned from January 1, 2014, to December 31, 2018. Results: A total of 815,170 EHR records were examined, corresponding to 461,392 episodes from 203,755 patients, divided into two groups: those with SG/SS (4.56%) and those without sepsis (95.44%). Out of 2,829 identified variables, 229 (8.09%) demonstrated a significant correlation with SG/SS detection, validated by the UMS team. Notable variability was observed in the association of variables with SG/SS depending on the hospital service. The ML-based predictive models exhibited outstanding capability in detecting SG/SS, with the best model being a combination (ensemble) of ML plus the SEPSIS.2 criteria, achieving an AUC-ROC of 0.95, with sensitivity and specificity of 0.93 and 0.84, respectively. Conclusion: The application of advanced predictive models based on AI-ML has yielded more suitable, dynamic, and personalized tools for detecting SG/SS in patients across all hospital areas compared to conventional scores.