Cecília R. C. Calado

Chronic kidney disease (CKD) is commonly assessed using blood biomarkers such as creatinine, urea and albumin. This study compared a portable six-channel near-infrared (NIR) analyzer (miniNIR; 1300–2100 nm) with a benchtop Fourier-transform near-infrared (FT-NIR) instrument for the discrimination and quantification of fifty-four independent concentration-defined profiles in composed fetal bovine serum solutions spanning six simulated KDIGO categories. SVM-RBF, Random Forest, XGBoost, Logistic regression and partial least-squares discriminant analysis (PLS-DA) were evaluated. For miniNIR binary ‘healthy/disease’ profile discrimination, SVM-RBF yielded a validation AUC of 1.00, accuracy of 0.91, sensitivity of 1.00 and specificity of 0.83. The six-class profile assignment was substantially more difficult: PLS-DA was selected by grouped nested cross-validation, but validation accuracy was 0.18, macro-F1 was 0.11 and macro-AUC was 0.77 across the evaluable classes. For miniNIR regression, validation R² values were 0.48 for creatinine, 0.30 for urea and 0.10 for albumin. The strongest FT-NIR validation results were R² = 0.64 for creatinine, 0.79 for urea and 0.89 for albumin, depending on the spectral window. For FT-NIR binary ‘healthy/disease’ profile discrimination, SVM-RBF yielded a validation AUC of 0.99, accuracy of 0.71 and a specificity of 0.47. Although the high binary AUC in this controlled concentration design for the miniNIR, it should not be interpreted as evidence of CKD-specific recognition or clinical performance. The comparison quantifies the information retained and lost when continuous FT-NIR sampling is reduced to six broad discrete channels; validation in clinically characterized human samples is required before any screening application can be considered.

Background: Pre-transplant serum Fourier-transform infrared spectroscopy (FTIRS) has identified an internally validated biochemical signal associated with subsequent biopsy-proven kidney allograft rejection. Whether this signal retains discrimination across post-transplant time horizons and rejection-phenotype-defined endpoints remains unclear. Methods: This internally conducted extension study reused the same single-center cohort of 79 kidney transplant recipients (40 with biopsy-proven rejection; 39 with five-year rejection-free follow-up). The primary analysis did not retrain the original model; the fixed nested out-of-fold FTIR-derived score was evaluated at cumulative rejection horizons of 30, 90, 180, 365, 730, 1095, and 1825 days. Discrimination was assessed using AUCs with bootstrap confidence intervals and permutation testing. Cox regression, C-index, Kaplan–Meier analysis, exploratory decision curve analysis, and exploratory endpoint-specific models were also performed. Results: The fixed score retained discrimination across horizons, with AUCs of 0.780, 0.829, 0.845, and 0.860 at 30, 90, 180, and 365 days, respectively, and 0.850 at 1825 days. In Cox analysis, the FTIRS score was associated with rejection risk (HR per SD 1.95, 95% CI 1.40–2.73; p = 0.00009; C-index 0.709). Endpoint-specific models showed variable discrimination but involved small subgroups. Conclusions: The fixed pre-transplant FTIRS score retained discriminatory value across cumulative rejection-time horizons, with numerically highest discrimination within the first year. Endpoint-specific findings were hypothesis-generating and did not establish time- or phenotype-specific spectral signatures. Independent external validation is required.

Background: Biotrauma from invasive mechanical ventilation (IMV) and extracorporeal membrane oxygenation (ECMO) drives systemic inflammation, metabolic dysregulation, and organ dysfunction in critically ill patients. Therefore, this study aimed to identify clinical and metabolomic features associated with ICU mortality in patients receiving IMV or ECMO, as these remain incompletely characterized. Methods: The retrospective analysis included 30 ICU patients on IMV and 22 on ECMO. Metabolomic and proteomic profiling were performed using ultra-high-performance liquid chromatography coupled with high-resolution mass spectrometry (UHPLC-HRMS), and serum spectral analysis by Fourier-transform infrared spectroscopy (FTIRS). Significant variables were incorporated into multivariate logistic regression models, ranked by AIC, AUC, and statistical significance. Model performance was evaluated using stratified 5-fold cross-validation. Final models were adjusted for relevant demographic and clinical covariates. Results: The IMV cohort showed discriminatory FTIRS wavenumbers across all preprocessings, and 155 metabolites plus 14 proteins were significantly altered, with unadjusted models achieving mean AUCs above 0.9. The ECMO cohort showed discriminatory FTIRS wavenumbers in one preprocessing, and 15 metabolites plus 3 proteins were highlighted. FTIRS, metabolomic, and proteomic models reached mean AUCs of 0.967, 0.867, and 0.783, respectively, with lower stability during cross-validation. Adjustment for demographic and clinical covariates reduced model robustness. Conclusions: Stronger and more reproducible molecular signatures related to ICU mortality were observed in the IMV cohort, whereas the ECMO cohort showed reduced model stability, likely reflecting increased biological heterogeneity and small sample size. These findings support the utility of integrated omics for characterizing critical illness and outcome stratification, while reinforcing the need for validation in larger and independent cohorts.

Background: Rejection remains a major barrier to long-term kidney allograft survival, and pre-transplant risk stratification remains incomplete. This study evaluated whether pre-transplant serum Fourier-transform infrared (FTIR) spectra, analyzed using machine learning methods, could identify kidney transplant recipients at increased risk of subsequent biopsy-proven rejection. Methods: In this retrospective single-center study, 80 pre-transplant serum samples collected on the day of transplantation were initially evaluated; after spectral quality control, 79 samples were retained for analysis. FTIR spectra were acquired in transmission mode and analyzed in the 600–1900 cm⁻¹ and 2800–3400 cm⁻¹ regions. Multiple preprocessing strategies were assessed, including Rubber Band baseline correction, vector normalization, and first- and second-derivative transformation, with and without normalization. Naïve Bayes classifiers with Leave-One-Out Cross-Validation and Fast Correlation-Based Filter feature selection were applied. Results: Exploratory analysis showed broad overlap between groups, indicating a subtle multivariate spectral signal. In the initial exploratory workflow, classifier performance depended strongly on preprocessing and feature selection. Because non-nested feature selection may produce optimistic estimates, the main supervised analysis was repeated using FCBF nested within each LOOCV training fold. The best-performing nested model was obtained using second derivative transformation followed by normalization in the combined 600–1900 and 2800–3400 cm⁻¹ regions, achieving an AUC of 0.837, accuracy of 0.747, sensitivity of 0.675, specificity of 0.821, balanced accuracy of 0.748, and F1-score of 0.730. Permutation testing with 1000 label-randomized repetitions supported performance above chance expectation, with no permuted model reaching the observed AUC (empirical p = 0.000999). Conclusions: Pre-transplant serum FTIR spectroscopy combined with leakage-aware nested machine learning analysis identified an internally validated spectral signal associated with subsequent biopsy-proven rejection. These findings support FTIR as a promising complementary and hypothesis-generating approach for pre-transplant biochemical risk phenotyping, requiring external multicenter validation before clinical application.

Background: Simultaneous pancreas–kidney (SPK) transplantation can restore renal function and insulin independence, but non-technical pancreas graft dysfunction remains difficult to anticipate. Methods: We conducted an exploratory single-centre retrospective biomarker-modelling study to determine whether day-zero recipient serum Fourier-transform infrared (FTIR) spectra are associated with subsequent loss of insulin independence after SPK transplantation. Results: Among 104 screened recipients, 51 met predefined sample-availability, spectral-quality, data-linkage and endpoint-adjudication criteria; 30 maintained pancreas graft function and 21 developed dysfunction. Cases dominated by early technical surgical failure were excluded. Clinical-only, FTIR-only and FTIR–clinical Naïve Bayes models were evaluated using leave-one-out cross-validation with Fast Correlation-Based Filter feature selection. In locked-feature internal validation, the best FTIR-only model used second-derivative spectra with vector normalization and nine selected wavenumbers, achieving AUC 0.997 (95% CI 0.985–1.000) and accuracy 0.961 (95% CI 0.902–1.000). A fixed-feature permutation analysis exceeded label-randomized performance (empirical p = 0.001). The secondary Group 1 versus Group 3 analysis suggested discrimination of pancreas dysfunction despite preserved kidney function (AUC 0.992; accuracy 0.930). Conclusions: Given the small cohort, high-dimensional input, non-nested feature selection, selection-bias risk and absence of external validation, serum FTIR should be considered a candidate risk-enrichment platform requiring prospective multicentre validation.

The development of new technologies enabling rapid, frequent, and reagent-free monitoring of kidney function is recognized as being of paramount importance. In this work, mid-(MIR) and near-infrared (NIR) spectroscopy were compared for the prediction of key renal biomarkers—creatinine, urea and albumin—using 54 serum solutions mimicking the biochemical profiles of five stages of chronic kidney disease (CKD). MIR spectra were acquired in a high-throughput microplate platform after a simple dehydration step, while the NIR spectra were obtained directly from liquid serum using a fiber optic probe. After evaluating several spectral pre-processing methods and targeted spectral regions, excellent regression models (R² > 0.9 for the best models) were obtained for the three biomarkers. MIR provided highly accurate urea predictions, whereas optimized NIR sub-regions enabled excellent estimation of creatinine and albumin. Both MIR and NIR, associated with supervised classification methods, enabled us to successfully distinguish healthy from diseased profiles and to identify the diseases state with AUC > 0.93. These findings highlight the complementary value of MIR and NIR spectroscopy for kidney disease assessment and their potential integration into point-of-care diagnostic systems.

Traditional and COVID-19-specific severity scores are applied in intensive care units (ICUs) to guide decision-making and predict mortality. Since traditional severity scores (APACHE II, SAPS II, SAPS 3, and SOFA) were not originally designed for SARS-CoV-2, this study compared their performance with COVID-19–specific models (Shang-COVID and SEIMC), including a novel distinction between early (≤7 days) and late (>7 days) ICU mortality. Adult ICU COVID-19 patients from the first two pandemic waves in Portugal were included (n = 286). Six scores were calculated, and four outcomes assessed: hospital, ICU, early ICU, and late ICU mortality. Discrimination was assessed using ROC curves with AUCs, 95% CIs, and p-values. AUCs were compared using the Delong test (early vs. late ICU mortality and across scores within each wave) and the Hanley & McNeil test (between waves for each score). Traditional scores demonstrated robust mortality prediction. SEIMC performed well for hospital (AUC_wave1 = 0.808; AUC_wave2 = 0.724) and ICU mortality (AUC_wave1 = 0.805; AUC_wave2 = 0.706). SEIMC (AUC_wave1 = 0.786; AUC_wave2 = 0.800) and Shang-COVID (AUC_wave1 = 0.617; AUC_wave2 = 0.736) showed potential for early mortality prediction but require further validation and recalibration. Overall performance was superior during the first wave, likely reflecting differences in patient characteristics, viral variants, and health measures. Traditional severity scores demonstrated stable robust prediction of ICU and hospital mortality in COVID-19 cases. Disease-specific scores did not significantly outperform established models, though also showed good predictive ability in some contexts, particularly early ICU mortality. These findings highlight the need for continuous validation and recalibration of predictive tools as clinical contexts evolve.

Background: Despite advanced analytical methods and increasing data availability, most intensive care unit (ICU) prediction models rely on static measurements. However, longitudinal monitoring of biomarkers may better capture disease progression and support timely, individualized interventions within the framework of predictive, preventive, and personalized medicine (PPPM). Since the COVID-19 pandemic, interest in both static and dynamic modelling has expanded. Therefore, this review aimed to summarize current evidence on the use of longitudinal blood biomarker data in ICU prediction models, assess how the pandemic shaped this research, and report validation strategies. Methods: This scoping review followed the PRISMA-ScR guidelines. PubMed and Google Scholar were searched for studies on blood biomarker trajectory analysis in the ICU published between 2014 and 2025, covering five years before and after the onset of the COVID-19 pandemic. Results: Forty-seven studies were included, mainly from North America (47%), Europe (45%), and Asia (34%). ICU and hospital mortality were the predominant outcomes. Although 53% of studies used pre-pandemic data, 94% were published afterwards. The most frequent biomarker categories were immune response (74.5%) and metabolic/organ function (66.0%). Common biomarkers included platelets and lactate ( n = 9), lymphocytes and mHLA-DR ( n = 6), and creatinine and interleukin-6 ( n = 5). Modelling approaches integrated longitudinal regression-based models (31.9%), latent class-based models (44.7%), and machine-learning/data-driven clustering (27.7%). Trajectory patterns varied depending on both biomarker type and modelling technique. Cox regression, Kaplan-Meier, and logistic regression were commonly applied to assess associations with outcomes. Notably, only 21% of studies reported any form of validation, highlighting a major limitation for clinical applicability. Conclusion: Blood biomarker trajectories have potential to improve dynamic risk prediction and stratification, supporting targeted prevention through early identification of high-risk patterns, and enable more personalized treatment via adaptive, patient-specific approaches. Nevertheless, substantial methodological heterogeneity and the low proportion of validated models limit clinical applicability. Greater standardization and robust validation are essential to facilitate translation into PPPM-oriented intensive care.