Machine learning-based early risk stratification for pulmonary thromboembolism in critically ill patients undergoing chest CT: a retrospective cohort study using CT-confirmed outcomes

Abstract

Pulmonary thromboembolism is a preventable cause of death in critically ill patients, and thromboprophylaxis decisions in the intensive care unit are made under diagnostic uncertainty. Risk scores derived from general inpatients discriminate poorly in intensive care. We developed and internally validated a machine learning model for pulmonary thromboembolism risk at 24 h after intensive care unit admission, and measured how much of its discrimination came from treatment information already recorded by then. Data came from the Medical Information Mart for Intensive Care IV database. All patients had contrast-enhanced chest computed tomography. The outcome was pulmonary thromboembolism confirmed by computed tomography between 24 h and 30 days after admission. Forty configurations (eight algorithms × five imbalance-handling strategies) were evaluated on one outer 5-fold split without tuning. The model and three comparison scores were placed on one probability scale by within-fold logistic recalibration, and differences in the area under the receiver operating characteristic curve were estimated by paired bootstrap. Death and discharge were treated as competing events. Among 4,470 patients (144 events; 3.22%), the model gave an area under the curve of 0.740 (95% confidence interval 0.700 to 0.780), above the Modified Padua score (0.548) and the Core (0.649) and Extended (0.622) intensive care unit venous thromboembolism scores (differences 0.192, 0.091 and 0.118; all P < .001). Incidence rose from 0.56% to 8.28% across risk quintiles. Removing anticoagulant therapy status lowered the area under the curve to 0.660 (difference 0.080, 95% confidence interval 0.041 to 0.120); the reduced model stayed above Modified Padua ( P < .001) but not above the Core ( P = .62) or Extended ( P = .25) score. The four Brier scores were within 0.001 of one another. Discrimination fell from 0.742 at a 1-day to 0.650 at a 7-day landmark. In this imaging-enriched cohort the model discriminated pulmonary thromboembolism better than three comparison scores. Most of that difference was carried by anticoagulant therapy status, a recorded clinician decision rather than a patient measurement; without it the model was indistinguishable from the intensive care unit venous thromboembolism scores. External validation is required before use.