Model-interpreted outcomes of artificial neural networks classifying immune biomarkers associated with severe infections in ICU.

Introduction: Millions of deaths worldwide are a result of sepsis (viral and bacterial) and septic shock syndromes which originate from microbial infections and cause a dysregulated host immune response. These diseases share both clinical and immunological patterns that involve a plethora of biomarkers that can be quantified and used to explain the severity level of the disease. Therefore, we hypothesize that the severity of sepsis and septic shock in patients is a function of the concentration of biomarkers of patients. Methods: In our work, we quantified data from 30 biomarkers with direct immune function. We used distinct Feature Selection algorithms to isolate biomarkers to be fed into machine learning algorithms, whose mapping of the decision process would allow us to propose an early diagnostic tool. Results: We isolated two biomarkers, i.e., Programmed Death Ligand-1 and Myeloperoxidase, that were flagged by the interpretation of an Artificial Neural Network. The upregulation of both biomarkers was indicated as contributing to increase the severity level in sepsis (viral and bacterial induced) and septic shock patients. Discussion: In conclusion, we built a function considering biomarker concentrations to explain severity among sepsis, sepsis COVID, and septic shock patients. The rules of this function include biomarkers with known medical, biological, and immunological activity, favoring the development of an early diagnosis system based in knowledge extracted from artificial intelligence.

Modelos de riesgo para la predicción de mortalidad hospitalaria en ancianos con neumonía por COVID-19

Objective. To compare the prognostic value of 3 severity scales: the Pneumonia Severity Index (PSI), the CURB-65 pneumonia severity score, and the Severity Community-Acquired Pneumonia (SCAP) score. To build a new predictive model for in-hospital mortality in patients over the age of 75 years admitted with pneumonia due to the coronavirus disease 2019 (COVID-19). Methods. Retrospective study of patients older than 75 years admitted from the emergency department for COVID-19 pneumonia between March 12 and April 27, 2020. We recorded demographic (age, sex, living in a care facility or not), clinical (symptoms, comorbidities, Charlson Comorbidity Index [CCI]), and analytical (serum biochemistry, blood gases, blood count, and coagulation factors) variables. A risk model was constructed, and the ability of the 3 scales to predict all-cause in-hospital mortality was compared. Results. We included 186 patients with a median age of 85 years (interquartile range, 80-89 years);44.1% were men. Mortality was 47.3%. The areas under the receiver operating characteristic curves (AUCs) were as follows for each tool: PSI, 0.74 (95% CI, 0.64-0.82);CURB-65 score, 0.71 (95% CI, 0.62-0.79);and SCAP score, 0.72 (95% CI, 0.63-0.81). Risk factors included in the model were the presence or absence of symptoms (cough, dyspnea), the CCI, and analytical findings (aspartate aminotransferase, potassium, urea, and lactate dehydrogenase. The AUC for the model was 0.81 (95% CI, 0.73-0.88). Conclusions. This study shows that the predictive power of the PSI for mortality is moderate and perceptibly higher than the CURB-65 and SCAP scores. We propose a new predictive model for mortality that offers significantly better performance than any of the 3 scales compared. However, our model must undergo external validation. Objetivo. Los objetivos son comparar la utilidad pronóstica de tres escalas de gravedad (Pneumonia Severity Index: PSI;CURB-65 scale;Severity Community Acquired Pneumonia Score: SCAP) y diseñar un nuevo modelo predictivo de mortalidad hospitalaria en pacientes mayores de 75 años ingresados por neumonía por COVID-19. Método. Estudio retrospectivo de pacientes mayores de 75 años ingresados por neumonía por COVID-19 desde el servicio de urgencias entre el 12 de marzo y el 27 de abril de 2020. Se recogieron variables demográficas (edad, sexo, institucionalización), clínicas (síntomas, comorbilidades, índice de Charlson) y analíticas (bioquímica en suero, gasometría, hematimetría, hemostasia). Se derivó un modelo de riesgo y se compararon las escalas de gravedad PSI, CURB-65 y SCAP para predecir la mortalidad intrahospitalaria por cualquier causa. Resultados. Se incluyeron 186 pacientes, con una mediana de edad de 85 años (RIC 80-89), un 44,1% varones. La mortalidad fue del 47,3%. Las escalas PSI, CURB-65 y SCAP tuvieron un área bajo la curva (ABC) de 0,74 (IC 95% 0,64-0,82), 0,71 (IC 95% 0,62-0,79) y 0,72 (IC 95% 0,63-0,81), respectivamente. El modelo predictivo compuesto por la ausencia o presencia de síntomas (tos y disnea), comorbilidad (índice de Charlson) y datos analíticos (aspartato- aminotransferasa, potasio, urea y lactato-deshidrogenasa) tuvo un ABC de 0,81 (IC 95% 0,73-0,88). Conclusión. Este estudio muestra que la escala PSI tiene una capacidad predictiva de mortalidad moderada, notablemente mejor que las escalas CURB-65 y SCAP. Se propone un nuevo modelo predictivo de mortalidad que mejora significativamente el rendimiento de estas escalas, siendo necesario verificar su validez externa.