MACHINE LEARNING ALGORITHM IMPROVES SEVERITY PREDICTION OF HANTAVIRUS PULMONARY SYNDROME BY INTEGRATING SEROLOGICAL AND CYTOLOGICAL DATA

WIEBKE L; KLINGSTRÖM J; SCHIERLOH P; CASCO V; RAUSCH A

Tucuman

Congreso; LXVII Reunión Anual SAI; 2019

Sociedad Argentina de Inmunología

AIM: Hantavirus Pulmonary syndrome (HPS) is a life threatening emergent viral disease caused by New World Hantaviruses. Early clinical decision regarding patient´s handle strongly influence the fatality rates that in our country vary between 15-35%. Therefore, rapid prognostic methods based on early stage biomarkers (BM) are strongly needed. M&M: We employed a previously validated data set obtained from 105 HPS cases where 22 serum BM were determined shortly after (< 4 days) prodrome symptoms debut by a 16-plex customized Luminex immunoassay+6 ELISA assays (R&D). Several BM combinations and data transformation methods were employed to train, validate and critically evaluate compared several machine learning algorithms (ANN, Random forest, linear discriminant and decision tree) as predictors of outcome variables: Severity and fatality. All analysis was conducted with R software. Results: A feed forward Artificial neural network (FANN) that combine only 4 common BM (IL-6/IL-10/IL-15/C5a) with one classic blood parameter (platelet count) significatively increase all predictive values of severity (AUC ROC=0.93; Sensitivity=91%; Specificity=86%) compared with our previously identified best independent serum predictors (IL-6 and I-FABP, p