Tennant, Ryan2026-04-132026-04-132026-04-132026-03-19https://hdl.handle.net/10012/22997In paediatrics, sepsis is a high-stakes, safety-critical challenge for clinicians to recognize and respond to, where a sick child can look stable until they are not, and where diagnostic definitions, symptom baselines, and care pathways vary across children and healthcare settings. This variability creates uncertainty in the clinical environment and limits the extent to which artificial intelligence (AI) and machine learning-based prediction tools can support clinical decision-making when the goal is to classify sepsis. In this dissertation, we position paediatric sepsis as a structurally uncertain domain and argue for an ecologically inspired, constraint-based approach to AI system design that learns and visualizes (1) the boundaries of physiological functioning and (2) the boundaries of the socio-technical system, to support clinically justified decisions under uncertainty in such non-specific contexts. To ground this work, we first synthesize the paediatric sepsis prediction literature through a scoping review. We find substantial heterogeneity in endpoint definitions, datasets, validation practices, prediction timing, and performance reporting. We also find limited attention to human factors considerations, such as workflow integration, interface design, and interaction design, despite their potential for clinical decision support, which fundamentally motivates the research of this dissertation. We then examine how clinicians experience and manage uncertainty about paediatric sepsis. Through semi-structured interviews with registered nurses, respiratory therapists, pharmacists, nurse practitioners and physicians in Canada about their experiences with paediatric sepsis recognition and response, we develop a domain-grounded account of uncertainty conceptualization and a sensemaking-action cycle model. This work extends the established uncertainty constructs from Naturalistic Decision Making by identifying two sources specific to paediatric sepsis: indeterminate clinical trajectories and operational constraints, including how emotional & intuitive anchoring shapes constructing readiness to act. We also describe AI-associated uncertainty concerns, including how and when predictions may reshape clinical judgement. Next, we establish the paediatric sepsis work domain constraints for recognition and response through Cognitive Work Analysis (CWA), conducting a tri-model Work Domain Analysis of the biological, clinical, and AI-augmented clinical work domains, and a Control Task Analysis of decision-making. We also apply this modelling to compare classification- and constraint-based prediction architectures and translate the constraints into an interactive Ecological Interface Design (EID) concept. Our resulting EID includes a configurable display of a baseline-relative trajectory forecast and AI-based uncertainty-aware cues to support the “gut feeling” of early paediatric sepsis recognition and escalation, and how and when to use model predictions at the bedside, respectively. Finally, we formatively evaluate the constraint-based approach in simulated paediatric sepsis scenarios with clinicians in individual and team-based contexts. Across outcomes, including clinical concern, confidence, trust, and sensemaking, our results suggest the trajectory forecast most strongly influences interpretation and action planning and is more validating of preparatory clinical action when the prediction is concordant with clinical reasoning. While AI-based uncertainty elements do not generally yield quantitative differences in outcomes in our evaluation format, they are perceived as potentially valuable for ongoing AI system use; however, mixed perceptions indicate the need for further research to improve their interpretability and usability at the bedside. Taken together, our findings support the promise of further investigating the constraint-based approach toward AI system design in supporting resilience and clinical judgement in paediatric sepsis. Overall, this dissertation demonstrates why the safe integration of AI into high-stakes healthcare cannot be purely data-driven. By conceptualizing uncertainty before introducing predictive algorithms and by applying CWA to anticipate how predictions may reshape clinical judgement, this research provides a framework for more responsible system design. Our approach supports a thorough examination of domain complexity and clinical judgement, informs decisions about whether and how an AI system should be developed and how prediction information should be introduced, and grounds clinical AI in human factors, ensuring these systems respond to genuine clinical needs.enartificial intelligencecognitive work analysisecological interface designhuman factorspaediatric sepsisDoctoral Thesis