Fluid management in critically ill trauma patients remains highly heterogeneous, particularly in those with pelvic fractures, where resuscitation requirements vary widely by injury severity and physiological response. This study aimed to develop a Dynamic Time Warping (DTW)-based adaptive framework to identify clinically interpretable severity phenotypes using 24-hour fluid trajectories as temporal evidence of physiological dysfunction rather than as causal determinants. The objective was to establish a data-driven, severity-centered phenotyping system capable of supporting precision fluid therapy and real-time clinical decision-making in intensive care settings.

A total of 1,306 ICU patients with pelvic fractures were enrolled from multiple sources, including MIMIC-IV, eICU-CRD, and a prospective Chinese multicenter cohort. All patient data were harmonized through a unified feature mapping process and processed using intelligent, machine-learning–based imputation under clinical constraints. Temporal trajectories of 24-hour fluid input were constructed for each patient and analyzed using a weighted and constrained DTW algorithm that captured nonlinear temporal similarities between individuals.

To ensure robust and interpretable clustering, a Clinical Severity Phenotype Classifier was designed to define subgroups based on trauma severity, hemodynamic stability, metabolic disturbance, and organ function, rather than observed treatment intensity—thus avoiding reverse causality. The classification was implemented within an Adaptive Balanced Framework, which dynamically adjusted group boundaries to maintain both clinical coherence and statistical balance across phenotypes. Validation included mortality gradient assessment, trajectory consistency, and balance evaluation. Kaplan–Meier survival analysis and Partial Dependence Plot (PDP) modeling were used to confirm the clinical interpretability and prognostic discrimination of the identified phenotypes.

Four clinically meaningful phenotypes were identified, each demonstrating distinct pathophysiological characteristics, fluid management patterns, and outcomes.

Phenotype 0 (Mild Trauma – Stable) consisted of 261 patients (20.0%) with a mean 24-hour fluid intake of 3,979 mL and ICU mortality of 1.9%. Patients exhibited higher hemoglobin (12.0 ± 1.3 g/dL), lower lactate (1.9 ± 0.6 mmol/L), and stable hemodynamics (MAP 88 mmHg). Their fluid trajectory showed a de-escalating pattern (Day 1: 3,979 mL → Day 2: 2,582 mL, ratio 0.65), reflecting preserved perfusion and adequate physiological reserve. Conservative management was generally sufficient.

Phenotype 1 (Moderate Trauma – Compensated) included 425 patients (32.5%) with a 24-hour fluid intake of 3,891 mL and ICU mortality of 1.9%. Patients presented with mild anemia (Hb 11.4 ± 1.4 g/dL), normal renal function (Cr 0.9 ± 0.3 mg/dL), and maintained compensation. Their fluid trajectory (Day 1 → Day 2 ratio 0.68) indicated controlled resuscitation tapering, and the majority required only moderate ventilatory support (42%).

Phenotype 2 (Moderate-Severe Trauma – Decompensating) comprised 405 patients (31.0%) with ICU mortality of 4.4% and mean 24-hour fluid intake of 4,063 mL. These patients showed declining hemoglobin (10.6 ± 1.4 g/dL), elevated lactate (2.6 ± 0.9 mmol/L), and early metabolic decompensation. Their trajectory (Day 1: 4,063 mL → Day 2: 2,447 mL, ratio 0.60) indicated ongoing resuscitation demand, and vasopressors were required in 21% of cases.

Phenotype 3 (Severe Trauma – Critical) represented the most severe group (n = 215, 16.5%) with a mean 24-hour fluid volume of 9,010 mL and ICU mortality of 19.1%. These patients exhibited profound shock physiology with low hemoglobin (9.7 ± 2.0 g/dL), high lactate (4.1 ± 2.5 mmol/L), hypotension (MAP 77 mmHg), renal impairment (Cr 1.3 mg/dL), and tachycardia (HR 100 bpm). Their high-volume, rapidly de-escalating fluid pattern (ratio 0.52) reflected aggressive initial resuscitation followed by cautious volume restriction. Over half required vasopressors (52%) and 73% mechanical ventilation.

Overall, the mortality–severity correlation was 0.846 (excellent), the severity gradient correlation was 0.999, and the group balance score was 0.783. The mortality range across phenotypes (17.2%) confirmed substantial clinical separation. Fluid–severity correlations were modest (total r = 0.240), supporting the interpretation of fluid input as a physiological marker rather than a determinant of outcome. Kaplan–Meier curves showed clear survival separation (log-rank p < 0.001), while PDP analysis confirmed lactate, hemoglobin, and baseline MAP as dominant predictors of mortality.

This study presents a DTW-based adaptive clinical severity framework that integrates dynamic trajectory modeling with pathophysiological stratification to identify four distinct and interpretable phenotypes in pelvic fracture ICU patients. By using 24-hour fluid patterns as temporal evidence of disease progression, the approach avoids reverse causality and enables more accurate representation of clinical severity. The identified phenotypes form a progressive continuum from stability to critical illness, each associated with unique physiological profiles, fluid utilization patterns, and outcome gradients.

The adaptive balancing mechanism within the framework ensures both statistical robustness and clinical realism, allowing for scalable integration into multi-center datasets. The system’s embedded decision support module translates phenotype recognition into actionable treatment guidance, such as conservative management for stable cases and early invasive hemodynamic monitoring with vasopressor initiation for critical patients.

These findings establish a new methodological paradigm for precision critical care, where data-driven temporal analysis enhances pathophysiological understanding and therapeutic decision-making. The DTW-based adaptive phenotyping framework provides a foundation for real-time ICU monitoring, risk stratification, and phenotype-guided interventional research aimed at optimizing individualized fluid therapy and improving outcomes in severe trauma populations.