Early prediction of generalized infection in intensive care units from clinical data: a committee-based machine learning approach
Flávio Secco Fonseca, Arianne Sarmento Torcate, Ana Clara Gomes da Silva, Victor Hugo Wanderley Freire, Gilles Paiva M. de Farias, João Fausto Lorenzato de Oliveira, Flavio Monteiro De Oliveira, Jose Carlos Da Silva, Dayane Aparecida Gomes, Wellington Pinheiros Dos Santos · 2022
Sepsis is a serious health condition caused by the body’s exaggerated response to an infection that influences organ failure and death of individuals, accounting for about 20% of global mortality. Early detection of sepsis remains a medical challenge due to heterogeneity in the source of infection. The lack of a specific diagnosis and early predictions make adequate clinical treatment unfeasible and thus leads to a mortality rate. The prediction of sepsis through machine learning models has gained special attention due to the abundance of available data. It is in this context that the present research proposes a hybrid model, composed of a committee of classifiers for the detection of sepsis. For this, we use the PhysioNet database, where we apply the Gaussian distribution to deal with missing data and the SMOTE method for class balancing. We started our experiment in an exploratory way, seeking to identify among nine algorithms which were the most promising to perform the classification tasks. Then, we developed our hybrid model composed of a Light Gradient Boosting, Multilayer Perceptron, Logistic Regression and Random Forest, both acting on a classification committee that reached an accuracy of 95.08%, Kappa of 0.1512, Sensitivity of 0.2868, Specificity of 0.9629 and AUC of 0.7434. Finally, it is worth mentioning that the results obtained are promising and open a range of possibilities for future work.