External validation of a smartphone app model to predict the need for massive transfusion using five different definitions

Abstract

Previously, a model to predict massive transfusion protocol (MTP) (activation) was derived using a single-institution data set. The PRospective, Observational, Multicenter, Major Trauma Transfusion database was used to externally validate this model's ability to predict both MTP activation and massive transfusion (MT) administration using multiple MT definitions. The app model was used to calculate the predicted probability of MTP activation or MT delivery. The five definitions of MT used were: (1) 10 units packed red blood cells (PRBCs) in 24 hours, (2) Resuscitation Intensity score ≥ 4, (3) critical administration threshold, (4) 4 units PRBCs in 4 hours; and (5) 6 units PRBCs in 6 hours. Receiver operating curves were plotted to compare the predicted probability of MT with observed outcomes. Of 1,245 patients in the data set, 297 (24%) met definition 1, 570 (47%) met definition 2, 364 (33%) met definition 3, 599 met definition 4 (49.1%), and 395 met definition 5 (32.4%). Regardless of the outcome (MTP activation or MT administration), the predictive ability of the app model was consistent: when predicting activation of the MTP, the area under the curve for the model was 0.694 and when predicting MT administration, the area under the curve ranged from 0.695 to 0.711. Regardless of the definition of MT used, the app model demonstrates moderate ability to predict the need for MT in an external, homogenous population. Importantly, the app allows the model to be iteratively recalibrated ("machine learning") and thus could improve its predictive capability as additional data are accrued. Diagnostic test study/Prognostic study, level III.