P-022 Laboratory medicine Poster

Making time count: extracting prognostic signals from serial procalcitonin measurements in critical care

Audience Choice Award

To rate this poster, scan the QR code on the poster itself — voting only works on site. Every attendee has 3 stars and can give at most one per poster. A star is final and cannot be taken back.

Authors

A. Fataar presenting

Division of Chemical Pathology, Stellenbosch University, Cape Town, South Africa

#0
A. Zemlin

Division of Chemical Pathology, Stellenbosch University, Cape Town, South Africa

#1
E. Kruger

Division of Chemical Pathology, Stellenbosch University, Cape Town, South Africa

#2

Abstract

Background Routine laboratory testing generates large volumes of longitudinal data that remain underutilised in clinical decision-making. Current practice relies on static threshold-based interpretation, overlooking dynamic information in serial measurements. In critical care, approaches that translate routine laboratory data into interpretable temporal risk signals may improve prognostic stratification.

Objectives Using procalcitonin (PCT) as a case study, we evaluated whether simple computational summaries derived from early serial measurements improve prediction of 28-day mortality in ICU patients with suspected sepsis.

Methods We analysed a retrospective ICU cohort using routinely collected laboratory data, reflecting real-world workflows. Serial PCT measurements were aligned to day 0 and assessed over days 0–5. Log-transformed values [log(1+PCT)] were used to derive slope and area under the curve (AUC). Bayesian logistic regression models were fitted to predict 28-day mortality using (i) baseline PCT, (ii) kinetic features, and (iii) kinetic features with selected clinical covariates. Performance was assessed using leave-one-out cross-validation.

Results The cohort included 128 ICU patients with a 28-day mortality of 32.8%. Serial PCT trajectories showed substantial inter-individual variability. Higher AUC of log(1+PCT) was associated with increased mortality (odds ratio [OR] 1.83, 95% credible interval [CrI] 1.25–2.74), while slope showed greater uncertainty (OR 1.42, 95% CrI 0.98–2.12). After adjustment, slope remained associated with mortality (OR 1.65, 95% CrI 1.03–2.70). Kinetics-based models outperformed baseline PCT, with minimal added value from clinical covariates.

Conclusion Simple computational features derived from routine serial biomarker measurements capture prognostic information not evident from single values. This work demonstrates that clinically meaningful risk signals can be extracted from existing laboratory data without additional testing. Key lessons include the value of aligning measurements to clinically relevant timepoints, using simple interpretable features, and validating models in real-world datasets. Because these features are derived from routinely reported results, they are readily implementable within laboratory information systems or clinical dashboards to support risk stratification in critical care.