Engineering biomarker representations of vital signs data enhances deep learning mortality prediction

Clinical Snapshot

Key Findings

• P Value: Not reported in available abstract

• Effect Size: Biomarker-engineered (frequency-normalised) representations achieved higher discrimination and lower Brier scores than raw 5-minute and hourly-averaged representations; specific AUROC and Brier score values are not reported in the abstract

• Primary Outcome: Inpatient mortality prediction performance of BiLSTM models trained on three different vital signs representations across HiRID and eICU datasets

• Nnt Or Sensitivity: AUROC (discrimination) and Brier score (calibration) are the primary performance metrics; specific values and confidence intervals are not available from the abstract alone. No sensitivity, specificity, or NNT estimates are reported.

• Confidence Interval: Not reported in available abstract

Clinical Application

The biomarker engineering pipeline extends the PhysioZoo software toolbox and is applied to routinely collected ICU vital signs data, suggesting computational feasibility in settings with electronic health record infrastructure. However, real-time deployment would require integration with ICU information systems, validation on local data, and prospective clinical evaluation. The approach is currently a research methodology rather than a deployable clinical tool. Australian ICUs routinely collect continuous vital signs data through integrated patient monitoring systems and ICU information systems (e.g., ICCA, MetaVision). The methodology is potentially applicable to Australian ICU data, including data held within the Australian and New Zealand Intensive Care Society (ANZICS) Adult Patient Database. However, the models have not been validated on Australian ICU populations, and local validation would be required before any clinical implementation. There are no current TGA-approved or PBS-listed applications of this specific technology. ANZICS and the Australian Commission on Safety and Quality in Health Care (ACSQHC) have emphasised the need for rigorous local validation of clinical prediction models before adoption. The RACGP and CICM (College of Intensive Care Medicine of Australia and New Zealand) would require prospective clinical utility evidence prior to endorsement. Adult ICU patients in general intensive care settings where continuous vital signs monitoring (blood pressure, heart rate, temperature, respiratory rate, SpO2) is routinely performed during the first 24 hours of admission

Abstract

OBJECTIVES: We evaluated bidirectional long short-term memory models for predicting inpatient mortality using different approaches to processing vital signs data collected during the initial 24 h of intensive care unit (ICU) admissions. MATERIALS AND METHODS: We compared 3 vital-sign representations: (1) raw data recorded every 5 min, (2) preprocessed data averaged hourly, and (3) preprocessed data using biomarker representations that extends a digital oximetry biomarker toolbox of PhysioZoo software, applied to blood pressure, heart rate, temperature, respiratory rate, and SpO2. RESULTS: Across 2 large ICU datasets, HiRID and eICU, models trained on the frequency-normalized representation achieved higher discrimination and lower Brier scores than those trained on raw 5-min and hourly averaged data. DISCUSSION: The use of biomarker representations of vital signs yielded the largest improvements in discrimination and overall probabilistic performance reflected by lower Brier scores for predicting inpatient mortality by deep learning. CONCLUSION: Thus, we recommend using a similar approach to vital signs preprocessing for time-series predictive models.

References

1. 1.Mamandipoor, B., Shen, I., Hsu, C.-N., & Gabriel, R. A. (2026). Engineering biomarker representations of vital signs data enhances deep learning mortality prediction. Journal of the American Medical Informatics Association. https://doi.org/10.1093/jamia/ocag066