A New Effective Machine Learning Framework for Sepsis Diagnosis

Abstract

There is a lack of early specific diagnosis and effective evaluation of sepsis, and the clinical treatment is not timely. As a result, the mortality is high, which seriously threatens the health of the people. Data were collected from the human blood samples of the hospital by gas chromatography mass spectrometry. Thirty-five healthy controls and 42 sepsis patients were enrolled. Machine-learning techniques were used to diagnose the sepsis. Using the metabolic data from the sepsis patients, the proposed method has got 81.6% recognition rate, 89.57% sensitivity, and 65.77% specificity. A new learning strategy was proposed to boost the performance of the kernel extreme learning machine, known as, chaotic fruit fly optimization, and two new mechanisms were introduced into the original a fruit fly optimization, including the chaotic population initialization and the chaotic local search strategy. To further enhance the diagnosis accuracy and identify the most important biomarkers, we performed the feature selection using the random forest before the construction of the classification model. The final established model, random forest-improved fruit fly optimization algorithm-kernel extreme learning machine, was used to effectively diagnose the sepsis. Experimental results demonstrate that the proposed method obtains better results than other methods across four performance metrics. We screened out five biomarkers and performed statistical analysis on these five substances. The level of acetic acid increased (p <; 0.05) in the sepsis group, while the level of linoleic acid and cholesterol decreased (p <; 0.05). The promising results suggest that the developed methodology can be a useful diagnostic tool for clinical decision support.