Automated detection of intravenous fluid contamination using longitudinal routine laboratory data: development and external validation of a multitask learning framework.

Autores de IIS La Fe

• Jordi Tortosa Carreres

  Autor

• Nahum Martí Montoro

  Autor

• Jonnathan Andres Acevedo Galvis

  Autor

• Ricardo Miguel Alonso Díaz

  Autor

• Oscar Fuster Lluch

  Autor

• María Begoña Laiz Marro

  Autor

Participantes ajenos a IIS La Fe

• Pascual-Escrivà C
• Sahuquillo-Frias L

Grupos

• Grupo de Investigación en Medicina de Laboratorio (GIMLAB)

Abstract

OBJECTIVES: To develop and externally validate an automated framework for intravenous fluid contamination (IFC) screening using routinely available longitudinal laboratory data. METHODS: A bicentric retrospective study included an internal cohort (IC; Hospital La Fe, València) for model development and an external validation cohort (EVC; Hospital General Universitari de València) composed of hospitalized patients with a previous determination within 5 days. Normalized relative change variables were calculated from paired measurements. Controls were retrospectively selected in the IC after removal of outliers using an autoencoder; contaminated samples were generated through controlled in vitro admixture with normal saline and glucose-containing solutions. A multitask neural network provided contamination probability, likely contaminant type, and severity-oriented estimation. Performance was evaluated using repeated internal 80:20 partitions (IRP; n = 30), the EVC, and an exploratory real-world dataset of retrospectively suspected IFC cases (RW-IFC). Model outputs were used to derive a stratified risk framework. RESULTS: The IC included 1760 samples (50% IFC), the EVC 1371 samples (5.4% IFC), and RW-IFC 37 samples. Discriminative performance remained high (ROC-AUC 0.94-0.96), with sensitivities and specificities =85%. Most false-negatives corresponded to contamination =5%, and the mean predicted contamination level among false-positive cases was 2.12 ± 1.06%. Multiclass accuracy ranged 0.87-0.90. Severity estimation showed strong internal correlation (? = 0.83; p < 0.0001) and good external agreement (0.84 within ±1 level), although correlation decreased in the EVC. 97% of RW-IFC cases were correctly identified. Risk stratification in the EVC achieved a sensitivity of 99.1% while reducing unnecessary alerts by 17%. CONCLUSIONS: This framework showed potential to support automated IFC screening.

Copyright © 2026 Elsevier B.V. All rights reserved.