Background
Risk assessment models (RAMs) have been developed to identify children at high risk of hospital‐acquired venous thromboembolism (HA‐VTE). None have been externally validated nor compared.
Objectives
The objective was to compare performance of these RAMs by externally validating them using the Children's Hospital‐Acquired Thrombosis (CHAT) Registry, ie, a multicenter database of children with radiographic‐confirmed HA‐VTE and corresponding controls.
Patients/Methods
Risk assessment models were included if the full logistic regression equation was available and all RAM variables were collected in the CHAT Registry. A random sample of 200 cases and 200 controls was selected. The performance of the RAMs was assessed for discrimination using area under the receiver operating characteristic curves (AUROC), and calibration using plots, slopes, and intercepts, and the Hosmer‐Lemeshow test.
Results
Three RAMs were included. Each had excellent discrimination with AUROC ≥ 0.85. However, calibration was generally poor, with calibration slopes significantly different from 1 (0.71, P < .001; 1.44, P = .002; 0.68, P < .001), intercepts significantly different from 0 (−1.64, P < .001; −0.62, P < .001; 0.78, P < .001), and Hosmer‐Lemeshow test P < .001 for each. Exceptions included the Arlikar et al and Atchison et al RAMs for pediatric HA‐VTE in non‐intensive care unit (ICU) patients and ICU patients, respectively, despite derivation from ICU and non‐ICU patients, respectively. In these subpopulations, both showed excellent discrimination and good calibration.
Conclusion
Given the lack of adequate calibration for evaluated RAMs, further investigation and refinement of RAMs for pediatric HA‐VTE is needed prior to application of a RAM in a clinical setting or risk‐stratified clinical trial of primary thromboprophylaxis against HA‐VTE in children.