Classimed logo

Outcome Prediction in Sepsis Patients by Machine Learning, a Pilot Study

G.Valet1), W.Kellermann2)

1) Max-Planck-Institut für Biochemie, Martinsried
2) Abt.Anästhesiologie, Krankenhaus-Schwabing der TUM, Munich, Germany

1. Introduction: Intensive care patients are in a life threatening condition when affected by sepsis or non infectious shock, so the prediction of the imminent danger for the development of these states is of high clinical importance.
Clinical sepsis research focuses frequently on the determination of biomarker levels in patient blood samples. Biomarkers like cytokines are in many instances liberated from immune cells (lympho-/monocytes). These mediators act in part upon effector cells like granulocytes. Depending on immune and effector cell status, given mediator levels may result in stronger or weaker cellular responses, that is mediator levels are not directly correlated with effector cell activities. Lymphocytes defend the organism by cellular and humoral (antibodies) immunity, which typically requires weeks to be established while sepsis often develops within hours. Once the lymphocyte dependent immune defense wall has been penetrated, granulocytes and monocytes act as important effectors, eliminating microorganisms or tissue breakdown products by phagocytosis and degradation through oxidative or enzymatic action. The overshooting release of granulocyte enzymes like elastase, of reactive oxygen species like H2O2, O2-, OH. or of pharmacologically active mediators like histamine may endanger the organism in case these potent functionalities escape inhibitory control mechanisms like in non infectious shock.
It seemed, therefore, promising to investigate granulocyte and monocyte effector functions in blood samples of intensive care unit (ICU) patients to determine early outcome predictors for individual ICU patients by data pattern classification.

2. Concept and Goals: Flow cytometric monitoring of bacteria phagocytosis by granulocytes was promising but too complicated to perform in automatated instruments (CL1). Cell function assays for the assessment of oxidative burst and proteolytic activities in mono- and granulocytes were therefore developed as an alternative using:
1. humoral stimulators like e.g. cytokines and
2. newly developed sensitive oxidative burst indicator dyes dihydrorhodamine123 (DHR) ( 8, 10, 11, 14) and specific rhodamine110 substrates for the determination of protease activity ( 12, 13, 17-22, 24) in vital cells. These developments have substantially simplified the determination of blood cell functions in infection, sepsis or non infectious shock.

3. Results: The early flow cytometric work ( 1, 2) using bacterial phagocytosis ( 6, 7), ADB intracellular pH and esterase ( 1, 2) measurements as well as acridine orange as indicator of cellular and bacterial RNA and DNA( 7) had shown for the first time that the prediction of imminent danger of sepsis and non infectious shock in intensive care (IC) patients was possible two to three days prior to the appearence of life threatening clinical symptoms (CL1). Simplified assays using intracellular oxidative burst and proteolytic capacities support these findings (details see below), providing a significantly increased therapeutic lead time for the clinician (CL2,CL3).

4. CLASSIF1 Data Pattern Analysis Flow cytometric data of such measurements are typically collected as list mode files. They are evaluated in a standardized and automated way by the CLASSIF1 (CL2,CL3) multiparameter data classification program. The analysis of the entire data set in this way reveiled that the incubation of the blood samples:
- as collected (ex-vivo status)
- with physiological stimulators such as: suboptimal concentrations of FMLP (formyl-methionyl-leucyl-phenylalanyl bacterial peptide), TNF-alpha (tumor necrosis factor-alpha), FMLP+TNF-alpha and
- with phorbol ester (PMA, phorbol-myristate-acetate) as maximum stimulus
provides a sufficient amount of predictive information (CL2) similarly as the determination of proteolytic enzyme activities like cysteine or serine proteinases
- The optimization of the classification process for the same group of septically admitted IC patients showed (CL3) that the most discriminatory predictive information was contained in the FMLP and TNF-alpha stimulated oxidative burst (DHR123) assays.
- As a practical consequence of the CLASSIF1 multiparameter data classification, only two out of the seven performed assays were really required for survival prediction in this group of ICU patients (CL3).

5. Conclusions:
A. Functional granulocyte and monocyte parameters provide individualized predictive information for the two to three days in advance recognition of life threatening sepsis occurence in ICU patients.
B. The proposed cell assays are suitable for automated preparation, cytometric measurement and standardized data classification.

Literature References:
CL1. G.Rothe, W.Kellermann, G.Valet: Flow cytometric parameters of neutrophil function as early indicators of sepsis or trauma-related pulmonary or cardiovascular organ failure. J.Lab.Clin.Invest.115:52-61(1990)
CL2. G.Rothe, W.Kellermann, J.Briegel, B.Schaerer, G.Valet: Activation of neutrophils by tumor necrosis factor-alpha during sepsis, in: Immune Consequences of Trauma, Shock and Sepsis Vol.II, Ed: E.Faist, J.Ninnemann, D.Green, Springer Verlag, Berlin 1993, p.727-733
CL3. G.Valet, G.Roth, W.Kellermann: Risk assessment for intensive care patients by automated classification of flow cytometric oxidative burst, serine and cysteine proteinase measurements using CLASSIF1 triple matrix analysis, in: Cytometric Cellular Analysis, Eds: J.P.Robinson, G.Babcock, Wiley-Liss, New York 1998, p.289-306

© 2017 G.Valet
Last update: Nov 08, 2017
First display: Feb 26,1996