Kinetics of the serum metabolic fingerprint to predict mortality of critically ill patients

Introduction: Accurate prognosis in intensive care units (ICUs), especially for mortality, remains a persistent challenge, mainly due to the heterogeneity of patients' demographic and clinical status, which is associated with a very dynamic pathophysiological state. The current work explores how the kinetics of the whole-serum molecular fingerprint, as acquired by FTIR spectroscopy, can be applied to develop predictive models of mortality in the ICU. Methods: The FTIR spectra of the serum of 44 patients, of which 23 were deceased, were acquired from ICU-hospitalized patients. Linear discriminant analysis and support vector machine models were developed to predict patient's mortality based on serum collected 3 to 7 days before the clinical outcome, i.e., the patient's death or patient's discharge. The application of diverse spectra pre-processing methods, feature selection, and kinetics of the spectral features was evaluated. Results: It was possible to develop only reasonable predictive models based on whole spectra, with accuracy < 75%. These models’ prediction was increased by feature selection, which resulted in a maximum accuracy of mortality prediction of 84% only 3 days before death. To extend the prediction window prior to the patient's death, the kinetics of these spectral features were considered, resulting in an accuracy prediction of 81% when considering the kinetics of the metabolic fingerprint between the 7^th and the 5^th days before death. Conclusions: Based on the kinetics of the serum spectral features, it was possible to develop a good predictive model of mortality 5 days before the patient’s outcome. These findings highlight the potential of the applied platform to develop good predictive models, based on rapid, simple, and minimally invasive analysis, towards the support of clinical decisions and the management of a relevant clinical infrastructure. Acknowledgements: This work was supported by the DSAIPA/DS/0117/2020, PTDC from the Portuguese Foundation for Science and Technology, and by the IPL/IDI&CA2024/R-DICIP_ISEL financed by Instituto Politécnico de Lisboa.