Characterizing Critical Illness Through Irregular Clinical Time Series and Continuous Multiscale Entropy of Physiological Waveforms

Abstract

A major challenge in critical care research is that clinical data are inherently irregular and heterogeneous. Laboratory values, vital signs, medication administrations, and clinical events are recorded at uneven intervals, reflecting the episodic nature of bedside care rather than continuous physiological processes. Traditional analytical approaches often require resampling or aggregation of these irregular clinical time series, which may obscure important temporal relationships and physiological dynamics. Characterizing critical illness through irregular clinical time series therefore requires computational frameworks capable of integrating asynchronous clinical data with continuously monitored physiological waveforms. By preserving the temporal structure of these data streams, it becomes possible to model evolving physiological trajectories, identify early markers of instability, and better understand how therapeutic interventions interact with underlying disease processes.Sepsis is a life-threatening medical emergency characterized by a dysregulated host response to infection that leads to systemic inflammation, endothelial activation, microcirculatory dysfunction, and neuroendocrine–immune dysregulation. Clinical outcomes in sepsis depend on the balance between injurious inflammatory cascades and the host’s capacity to restore physiological homeostasis. The autonomic nervous system (ANS) plays a central role in this process by mediating cytokine-to-brain communication, regulating cardiovascular responses, and exerting anti-inflammatory effects that may prevent progression to circulatory collapse. Current clinical management of sepsis emphasizes early administration of broad-spectrum antibiotics, aggressive fluid resuscitation, and vasopressor therapy, particularly norepinephrine for septic shock. However, both the pathophysiology of sepsis and the therapeutic administration of catecholamines contribute to a hyper-adrenergic state characterized by sympathetic overactivation and elevated circulating catecholamines. These changes profoundly influence autonomic regulation and cardiovascular dynamics.Heart rate variability (HRV) provides a noninvasive method for assessing autonomic function and has been widely investigated as a prognostic marker in sepsis. Time-domain, frequency-domain, nonlinear, and entropy-based features derived from electrocardiogram (ECG) and photoplethysmography (PPG) waveforms offer quantitative measures of autonomic modulation and physiological adaptability. Sequential evaluation of these waveform-derived features before, during, and after therapeutic interventions enables differentiation between persistent autonomic injury driven by systemic inflammation and rapid, dose-dependent physiological responses induced by exogenous catecholamines. Despite these advances, significant barriers remain for clinical translation of autonomic biomarkers in sepsis, as most prior investigations rely on static HRV metrics that fail to capture the dynamic evolution of autonomic dysfunction during disease progression and clinical management. Autonomic dysregulation and loss of physiological adaptability are not unique to sepsis. Major surgical procedures induce a complex inflammatory and neuroendocrine stress response characterized by endothelial dysfunction, microcirculatory alterations, and metabolic perturbations. Surgical critical illness therefore provides a complementary and mechanistically controlled physiological framework in which the timing and magnitude of therapeutic exposures are well defined. This controlled environment enables rigorous evaluation and validation of autonomic dysfunction signatures derived from continuous physiological waveforms. This dissertation investigates whether discriminative features extracted from physiological waveforms can serve as surrogate markers of autonomic dysfunction in critical illness. The central hypothesis is that waveform-derived measures of physiological complexity, particularly continuous multiscale entropy (MSE), capture early disruptions in autonomic regulation prior to overt hemodynamic deterioration, thereby identifying patients at risk for progression to septic shock and multiorgan dysfunction. Furthermore, these autonomic signatures are hypothesized to generalize to surgical critical illness, where controlled physiological stress allows mechanistic validation of resilience and failure. To address this hypothesis, this work develops novel machine learning methodologies that leverage advanced signal processing of irregular clinical time series and high-resolution physiological waveforms. By integrating continuous multiscale entropy with waveform-derived features from ECG and PPG signals, the proposed framework enables physiologically interpretable quantification of autonomic dysfunction. These methods aim to support personalized risk stratification and outcome prediction across both sepsis and surgical critical illness. Collectively, this dissertation establishes a scalable, noninvasive approach for physiological waveform complexity, providing new insights into the dynamic physiology of critical illness and enabling earlier detection of physiological deterioration.