Eris van Twist

Critically ill children admitted to the paediatric intensive care unit (PICU) generate vast amounts of continuous, high-frequency multimodal monitoring data. While these data convey insight into the patient’s status, human cognition precludes a swift translation into actionable information at the bedside. Translating data to information is inherently complex due to its volume, heterogeneity and real-world imperfections such as noise and missingness. This complexity, alongside the urgency of critical care, necessitates algorithms with analytical and interpretive capacity to translate data into actionable information. The resultant process, wherein data is actively used to guide clinical care and decision-making, is known as data-driven healthcare. While research on data-driven algorithms in healthcare has expanded, there is a persistent gap towards implementation, limiting the impact on clinical practice. Underlying the gap are systemic issues including methodological inconsistencies, poor (data) quality control, limited transparency, lack of trust among clinicians and workflow barriers. Adding to the chaos, real‑world monitoring data are often fragmented across hospital information systems, differ in resolution and structure, and contain missingness or noise. Clinicians may not fully grasp the immediate value of integrating data in the care process, while engineers and data scientists lack knowledge of the clinical context. Thus, to order the chaos, a methodological foundation is needed, grounded in transdisciplinary expertise at the intersection of medicine, engineering and data science. This thesis establishes a methodological foundation for data‑driven (paediatric) critical care through the lens of the data pipeline, the conceptual and practical sequence where raw input data flow towards an interpretable output. The data pipeline offers a standardized framework for algorithm development, reporting and reviewing. Important steps are data collection and processing, algorithm development and validation and output (interpretation). Transparent reporting across these stages is essential to ensure algorithm reproducibility. Robust and reproducible algorithm development is challenging amidst the variation in objectives, analytical approaches and clinical contexts. In Section II, we review existing quality assessment tools and reporting guidelines applicable to algorithms. Quality standards were fragmented, limited in scope or insufficiently established. Therefore, we introduce DATA-CARE, a quality assessment tool that offers critical appraisal of the data pipeline by integrating clinical, epidemiological and data science domains. DATA-CARE allows practical and comprehensive critical appraisal and indirectly sets a standard for future algorithm development. While further validation is necessary, DATA-CARE achieved high reliability. Importantly, DATA-CARE showed that current algorithms in paediatric critical care literature were predominantly of low quality, due to deficiencies in data quality and pipeline transparency. Furthermore, many studies neglected fundamental statistical principles during reporting. The implementation gap is inevitably associated with these quality deficiencies, as clinical implementation requires robust, reproducible and reliable algorithms. Building on the foundation, Sections III through IV explore algorithm development across three PICU use-cases with distinct clinical contexts and analytical strategies. In Section III, two machine learning (ML) models were developed for automated sleep classification using electroencephalography (EEG) and electrocardiography (ECG) data of noncritically ill children. In Section IV, a neuromonitoring algorithm was developed deriving targets for cerebral autoregulation (PRx), optimal cerebral perfusion pressure (ICP) and cumulative intracranial pressure (ICPcum) in children with traumatic brain injury (TBI). Section V presents a proofofconcept model predicting pulmonary hypertension and cardiac dysfunction in neonates with congenital diaphragmatic hernia using routine monitoring data including vital signs, ECGderived heart rate variability (HRV) and oxygen saturation index. Across all use-cases, data pipelines were constructed with ingestion of bedside monitoring data. It was illustrated that real-world data cannot be automatically presumed to be correct and require (pre-)processing to mitigate heterogeneity, noise and missingness. A minimal artefact detection and filtering protocol was applied, analyses were adapted to missing data patterns and population heterogeneity was managed through sampling techniques and class-weight calibration. Feature selection, distinctive characteristics of the data, and labelling, mapping outcome measures to data samples, emerged as critical components. In this thesis, features were selected using combined clinical insight and data characteristics, while labels were derived from gold standards or best-available surrogates. Future work should improve labelling strategies, especially in the absence of a gold standard, and implement data standardization and validation. Collectively, the use cases illustrated how continuous monitoring data can generate insights without additional burden on patients and/or caregivers. Extensive algorithm validation is needed to bridge the implementation gap. In the literature, such validation using complementary evaluation metrics is frequently omitted. In this thesis, algorithms were internally validated using data withheld from development and assessed using complementary metrics of discrimination and calibration, as well as their ability to generate new insight. In Section III, sleep classification models were externally validated in a cohort of critically ill children. It was shown that models did not generalize to the PICU population. Altered sleep physiology and abnormal data patterns possibly contributed to this, emphasizing the importance of thorough understanding of the data across the data pipeline for robust analysis and validation. Future (external) validation should move beyond binary notions of performance and incorporate clinical utility and long-term outcomes. Explainability, insight into algorithm decision-making, is another barrier to clinical implementation. Explainability can be enhanced through open source code or transparent reporting. In Section IV, the neuromonitoring algorithm was published open source to facilitate external validation by third parties. However, successful use of open source code critically hinges on data formats, emphasizing the need for (inter)national data standardization. Explainability further hinges on successful interpretation of algorithm-derived outputs. The algorithm-derived neuromonitoring targets in Section IV were statistically analysed, revealing significant associations with unfavourable outcome. Although this finding should be confirmed in future studies, it illustrates how traditional statistics can complement algorithm output. Ultimately, improved explainability provides a basis for clinician trust and supports clinical implementation. This thesis underscores the urgent need to establish a methodological foundation for data-driven algorithms in healthcare. Bottom-up improvements are needed in data quality, standardized data pipelines and validation that includes explainability and clinical utility. At the same time, top-down action is needed to establish interoperable data infrastructures and structurally embed inhouse transdisciplinary expertise. Together, these efforts can improve algorithm quality, build clinician trust, bridge the implementation gap and ensure clinical benefit. Data-driven healthcare holds the promise of improved and timely care through patient-tailored monitoring, early warning systems and targeted therapy, ultimately contributing to recovery and outcomes of critically ill children. The presented studies demonstrate the value of transdisciplinary expertise and the data pipeline to create robust and reproducible clinical algorithms in this vulnerable population.