Smartphone-based Early Detection of Pediatric Sepsis Signs Using AI-Powered Triage Tool

Thesis Abstract

This study addresses the critical challenge of early identification of pediatric sepsis signs to improve timely clinical intervention and reduce mortality in acute pediatric care. Despite advances in pediatric sepsis management, delays in recognition remain a leading cause of adverse outcomes, particularly in low-resource or high-patient-volume settings where clinical signs may be subtle or non-specific. The aim is to develop and validate a smartphone-based AI-powered triage tool that analyzes real-time patient data and caregiver-reported symptoms to generate risk scores for pediatric sepsis and prompt clinical action. Specific objectives include (1) to design a mobile application that integrates vital sign inputs, visual and behavioral cues via image and audio capture, and standardized symptom questionnaires; (2) to train and optimize machine learning models (including gradient boosting, random forest, and deep learning classifiers) on a multicenter pediatric dataset to detect early sepsis signals with high sensitivity and acceptable specificity; (3) to evaluate the tool’s diagnostic performance against established pediatric sepsis criteria (pSOFA, PELOD-2) and clinician assessment; (4) to assess usability, acceptance, and impact on clinician workflow through mixed-methods evaluation; and (5) to explore ethical, legal, and data privacy implications of deploying AI-driven triage in pediatric care. The study adopts a pragmatic mixed-methods design grounded in the Technology Acceptance Model (TAM) and the Diffusion of Innovations theory to examine user adoption and integration into routine care. A multicenter cohort comprising 1,200 pediatric patients aged 1 month to 18 years will be recruited from three tertiary pediatric emergency departments over 18 months. Data collection will combine prospective sensor data (heart rate, respiratory rate, oxygen saturation), anthropometric measures, caregiver-reported symptoms via a structured questionnaire, and video-audio cues when consented, stored and processed in compliance with HIPAA-equivalent standards. The primary data analysis will involve developing and validating predictive models using cross-validated penalized logistic regression and gradient boosting methods, with model performance assessed by AUROC, sensitivity, specificity, positive predictive value, and negative predictive value. Calibration will be examined using Hosmer-Lemeshow tests, and decision-curve analysis will quantify clinical usefulness. Secondary analyses will employ time-to-event methods (Cox proportional hazards) to evaluate early intervention impact on time to antibiotic administration and progression to organ dysfunction. A nested qualitative study with 40 semi-structured clinician and caregiver interviews will be analyzed thematically to elucidate acceptability, perceived accuracy, and workflow implications, using NVivo for coding and triangulation with quantitative findings. The expected findings include (i) a robust AI triage model achieving AUROC ?0.90 with sensitivity ?0.85 and specificity ?0.75 in external validation; (ii) evidence that integration of digital biomarkers and caregiver input enhances early detection beyond traditional vital signs alone; (iii) high usability ratings (System Usability Scale ?70) among clinicians and acceptable user experience among caregivers; and (iv) feasible data governance frameworks that satisfy ethical and legal standards for pediatric AI tools. The study contributes to knowledge by demonstrating how smartphone-enabled AI triage can operationalize early sepsis detection at point-of-care, augment clinical judgment, and shorten time to treatment, thereby potentially reducing pediatric sepsis mortality and morbidity. It advances understanding of the practical integration of AI-driven decision support in pediatric emergency workflows, addresses gaps in clinically validated mobile diagnostics for sepsis, and informs policy on data privacy, consent, and governance for mobile health interventions in children. The main conclusion is that a rigorously validated smartphone-based AI triage tool can reliably identify pediatric patients at elevated risk for sepsis in real-time, enabling timely intervention while maintaining acceptable patient privacy and workflow compatibility. Recommendations include scaling validation in diverse geographic regions, continuous model recalibration with new data, integration with electronic health record systems, and ongoing monitoring of equity, bias, and patient outcomes to ensure safe, effective deployment in routine pediatric care.