Related Experiment Video
Updated: Aug 8, 2026
Cutoff Value of Phase Angle by Bioelectrical Impedance Analysis at Admission as a Prognostic Factor in Patients with Acute Heart Failure
Published on: June 10, 2025
Admission photoplethysmography-based mortality prediction in hospitalized Ugandan children with suspected or
Mahan Rahimi1,2, Matthew O Wiens2,3,4, Jerome Kabakyenga5,6
1School of Biomedical Engineering, University of British Columbia, Vancouver, BC, Canada.
Insights
Early detection of sepsis in children is vital. Machine learning models using photoplethysmography at admission show potential for identifying high-risk pediatric patients in resource-limited settings.
Area of Science:
- Pediatric critical care
- Biomedical engineering
- Machine learning in healthcare
Background:
- Sepsis is a leading cause of preventable pediatric deaths in developing nations.
- Effective risk identification tools are lacking, hindering timely interventions.
- Early detection is crucial for resource allocation in under-resourced settings.
Purpose of the Study:
- To develop and validate machine learning models for predicting pediatric sepsis mortality.
- To assess the utility of photoplethysmography (PPG) for risk stratification upon hospital admission.
- To identify high-risk children in resource-limited settings for targeted interventions.
Main Methods:
- Secondary analysis of prospectively collected data from the Smart Discharges project.
- Development of machine learning models using features from 1-minute PPG signals at admission.
- Prediction of all-cause in-hospital mortality in children under five, stratified into two age groups (0-6 and 6-60 months).
Main Results:
- Mortality rates were 7% in younger (0-6 months) and 4.1% in older (6-60 months) children.
- The best-performing models achieved an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.70 and 0.67 for younger and older cohorts, respectively.
- Models demonstrated satisfactory calibration and clinical utility via decision curve analysis, using rapid, objective PPG data.
Conclusions:
- Simple, admission-based machine learning models using PPG show moderate predictive capacity for pediatric sepsis mortality.
- These models offer a potential tool for identifying high-risk children in resource-limited settings where advanced diagnostics are unavailable.
- External validation is recommended prior to clinical implementation; models are not a substitute for clinical judgment.
Abstract:
Sepsis remains a major cause of preventable pediatric hospital deaths in developing countries, with progress hindered by the lack of effective risk identification tools. Early detection of children at highest risk upon hospital admission is crucial for guiding clinical care and allocating resources, particularly in resource-limited settings. Photoplethysmography, which measures blood oxygen levels, also provides objective insight into cardiovascular alterations associated with sepsis. We conducted a secondary analysis of prospectively collected data from the Smart Discharges project, involving children under five years hospitalized with suspected or confirmed infection at six Ugandan hospitals, and developed models to predict all-cause in-hospital mortality across two age groups (0-6 and 6-60 months). Mortality was 7% in younger and 4.1% in older children. Machine learning models were trained on features extracted from one-minute photoplethysmograms collected at the time of admission. The best-performing model achieved mean values for the area under the receiver operating characteristic curve of 0.70 (95% CI: 0.62-0.76) in the younger cohort and 0.67 (95% CI: 0.56-0.73) in the older cohort, with corresponding values for the area under the precision-recall curve of 0.18 (95% CI: 0.12-0.27) and 0.14 (95% CI: 0.06-0.22), respectively. Calibration within risk strata was satisfactory (Brier scores 0.06 and 0.04), and decision curve analysis showed clinical utility. Notably, the models' predictive capacity, although moderate, was achieved with a rapid and readily available objective measurement at admission, without the need for extended monitoring. While less accurate than most existing risk scores and not a substitute for clinical judgment, these simple admission-based models may help identify high-risk children and guide targeted interventions where sophisticated diagnostics are unavailable. External validation is needed before adoption.