Diagnosis and prognosis disease biomarkers on critically ill patients with COVID towards a precision medicine – a machine learning approach

Funder

Organizational Unit

Publications

Predicting Critically Ill Patients Outcome in the ICU usinh UHPLC-HRMS data

Publication . Calado, Cecília; Fonseca, T.A.H.; Rekowski, C.P. Von; Araújo, R.; Oliveira, M. Conceição; Bento, L.; Justino, G.C.

The available scores to predict patients’ outcomes in specific settings generally present low sensitivities and specificities when applied to intensive care units’ (ICUs) populations. Advancements in analytical techniques, notably Ultra-High Performance Liquid Chromatography- Mass Spectrometry (UHPLC HRMS) transformed biomarker identification, enabling a comprehensive profiling of biofluids, including serum. In the current work, untargeted metabolomics, utilizing UHPLC-HRMS serum analysis, was performed on 16 ICU patients, categorized as either discharged (n=8), or deceased (n=8) in average seven days post sample collection. Linear discriminant analysis (LDA) or principal component analysis (PCA)-LDA models involving different metabolite sets were developed, enabling to predict patients’ outcomes in the ICU with 92% accuracy and 83% sensitivity on validation datasets. These results highlight the advantages of UHPLC-HRMS as a platform capable of providing a set of clinically significant biomarkers to predict patients’ outcome.

2024Conference object

Open access

Early mortality prediction in intensive care unit patients based on serum metabolomic

Publication . Araújo, Rúben; Ramalhete, Luís; Von Rekowski, Cristiana; Fonseca, Tiago AH; Bento, Luís; Calado, Cecília

Predicting mortality in intensive care units (ICUs) is essential for timely interventions and efficient resource use, especially during pandemics like COVID-19, where high mortality persisted even after the state of emergency ended. Current mortality prediction methods remain limited, especially for critically ill ICU patients, due to their dynamic metabolic changes and heterogeneous pathophysiological processes. This study evaluated how the serum metabolomic fingerprint, acquired through Fourier-Transform Infrared (FTIR) spectroscopy, could support mortality prediction models in COVID-19 ICU patients. A preliminary univariate analysis of serum FTIR spectra revealed significant spectral differences between 21 discharged and 23 deceased patients; however, the most significant spectral bands did not yield high-performing predictive models. By applying a Fast Correlation-Based Filter (FCBF) for feature selection of the spectra, a set of spectral bands spanning a broader range of molecular functional groups was identified, which enabled Naïve Bayes models with AUCs of 0.79, 0.97, and 0.98 for the first 48 h of ICU admission, seven days prior, and the day of the outcome, respectively, which are, in turn, defined as either death or discharge from the ICU. These findings suggest FTIR spectroscopy as a rapid, economical, and minimally invasive diagnostic tool, but further validation is needed in larger, more diverse cohorts.

2024-12-19Journal article

Open access