Background: "Critical weight loss" (CWL) is clinically important in cancer, yet thresholds are inconsistently defined and difficult to operationalise at scale. The common 5-10% cut-point is arbitrary [1], Body Mass Index (BMI)-adjusted "clinically important" thresholds range 5-20% across studies [2], and reported prevalence depends heavily on measurement window. Most adverse-event reporting and trials use %weight loss alone, without accounting for baseline BMI.
Methods: We modelled alternative weight-loss (WL) definitions using the Observational Medical Outcomes Partnership (OMOP) episode structure for comparison at scale across time windows. We computed %WL, timeframe-windowed %WL, and Martin et al.'s survival-validated, BMI-adjusted %WL grade across disease- and treatment-episode levels, in a head and neck cancer (HNC) cohort and a mixed-cancer comparison sample (real-world data from South Western Sydney). Descriptive statistics summarised WL distributions and grade prevalence and a chi-square test assessed height availability by treatment pattern.
Results: Apparent prevalence of ≥5% WL varied more than ten-fold according to measurement window (30-day: 5.5%; 90-day: 16.9%; 180-day: 33.3%; baseline-to-latest: 58.1%). HNC patients showed greater WL than the mixed-cancer sample (n=1,558 vs n=2,000; median %WL -7.4% vs -0.5%; 58.1% vs 25.4% meeting ≥ 5%). Among 1,148 episodes with both a %WL-only and BMI-adjusted grade computable, only 24% agreed: 76% shifted to a more severe grade once BMI was considered, and none improved. Height availability varied by treatment modality, from 89% (systemic therapy plus radiotherapy) to 68% (no treatment), consistent with height capture for body-surface-area systemic therapy dosing needs rather than nutritional surveillance (p<0.001).
Conclusion: Episode-level modelling demonstrated apparent CWL prevalence and severity vary substantially by measurement timeframe and whether BMI is incorporated. This framework offers a transparent and scalable approach to operationalising CWL definitions in routinely collected cancer data, while identifying data-quality issues relevant to future BMI-adjusted implementation.