Open Collections

Abstract

Sepsis remains a major cause of preventable pediatric mortality in low- and middle-income countries, where early identification of high-risk children is challenged by limited health worker availability, high patient volumes, and variability in frontline assessments. This thesis aims to develop and validate data-driven tools for early in-hospital risk stratification using routinely obtainable admission data from children under five admitted with suspected sepsis across six Ugandan hospitals (mortality 7% in the 0–6-month cohort; 4.1% in the 6–60-month cohort). We conducted a secondary analysis evaluating three modeling streams: photoplethysmography (PPG)-only models based on ten features extracted from one-minute infrared waveforms, clinical-only models using eight simple clinical and anthropometric variables, and combined clinical+PPG models integrating both data sources. Across all streams, model development employed repeated nested cross-validation, strict separation of internal and external cohorts, and model-agnostic interpretability techniques. PPG-only models demonstrated, for the first time to our knowledge, that brief, low-cost waveforms captured at admission could encode clinically meaningful signatures predictive of mortality, though performance was moderate and sensitive to probe-related domain shift. Clinical-only models achieved stronger and more stable discrimination, retaining ranking ability during temporal external validation despite lower outcome prevalence. Combined clinical+PPG models yielded the highest overall performance, improving discrimination, precision–recall metrics, and decision-analytic utility relative to unimodal approaches. These models also reduced false positives at fixed sensitivity thresholds, offering operational advantages in resource-constrained settings. External validation revealed heterogeneous generalizability: probe heterogeneity impaired PPG-only models in younger children, whereas combined models were more resilient and performed well in cohorts with harmonized data-collection pipelines. Taken together, these findings demonstrate the feasibility and promise of integrating physiological waveforms and simple clinical measures into machine learning models for early pediatric risk stratification in resource-constrained hospitals. By leveraging rapid, low-cost, and objective admission data, the models developed in this thesis provide a foundation for scalable digital decision-support tools that may strengthen triage, reduce resource misallocation, and support timely escalation of care. Prospective evaluation and broader external validation are needed before clinical implementation.