Design and applicability of clinical prediction models for serious infection in children with fever: a scoping review using causal directed acyclic graphs

<title>Abstract</title> <p>Background Clinical prediction models (CPMs) have been developed to estimate the risk of serious or invasive bacterial infection (SBI/IBI) in febrile children and support clinical decision-making, however routine clinical uptake remains limited. Differences in study design and model development may hinder comparison of available tools and assessment of suitability for specific contexts. This scoping review aimed to characterise the key design and development features of available CPMs and map them within causal directed acyclic graphs (DAGs) to examine how they may influence transportability and clinical applicability across clinical settings. Methods A scoping review was conducted per the JBI framework. Medline, Embase and relevant bibliographies were systematically searched to identify CPMs predicting SBI/IBI risk in febrile children (&lt; 18y) presenting to emergency and/or ambulatory care. Extracted data included study design, derivation cohort characteristics, outcome definitions, predictors, modelling approaches and output types. DAGs were constructed to depict hypothesised causal dependencies between key concepts. Results Of 4010 studies identified, 56 met inclusion criteria. Substantial heterogeneity was observed across all aspects of study design and model development. Most CPMs were developed in populations aged under 3 months (n = 35), however age-ranges and cohort selection criteria were diverse. SBI was the predominant target outcome (n = 36), however definitions differed and reported prevalence varied substantially (0.8%-37.4%). Developed models varied in algorithm complexity, predictor selection and output format, with potential implications for interpretability and usability. Algorithms included straightforward rule-based, decision-tree, and score-based models, and more complex statistical and machine-learning approaches. Predictor requirements differed in number (2–28 predictors/model), type, definition and handling of missing data. Prediction outputs were predominantly categorical risk-groups intended to guide clinical management. Models generating probabilistic risk estimates often demonstrated better discrimination but did not consistently define clinically-actionable decision thresholds. Constructed DAGs illustrated pathways through which cohort selection and outcome definitions may influence case ascertainment, predictor-outcome relationships, and model transportability. Conclusions CPMs for predicting SBI/IBI in febrile children demonstrate substantial heterogeneity extending beyond predictive performance alone. Causal DAGs can provide a transparent framework for comparing and evaluating the transportability and applicability of available CPMs, highlighting opportunities to align core design elements to improve comparability, implementation and clinical utility. Clinical Trial Registration (if any) : Not applicable</p>