Predicting Dynamically Evolving New-Onset Venous Thromboembolic (VTE) Event Risk in Hospitalized PatientsPellathy, Tiffany (2021) Predicting Dynamically Evolving New-Onset Venous Thromboembolic (VTE) Event Risk in Hospitalized Patients. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractBackground. Hospital acquired (HA) venous thromboembolism (VTE) is the leading cause of preventable hospital death. VTE pathology and symptoms evolve slowly over hours to days. No current VTE risk assessment models incorporate the progressive accrual of dynamic patient data over time of hospitalization. Classification algorithms which incorporate prediction time windows hold promise for closing this gap. Methods. An observational, retrospective, cohort study was conducted to develop predictive models to classify patients (n=2370) at risk for HA-VTE during SDU admission. Binary logistic regression (BLR), naïve Bayes (NB), Random Forest (RF), and Gradient Boosted Decision Tree (GBDT) algorithms were used to train models for two prediction time windows. Performance was evaluated with 10-fold stratified cross-validation. Models (S+/-) were developed to differentiate patients suspected of HA-VTE who underwent diagnostic radiology evaluation (n=760) from those not suspected/not tested (n=1614). A second set of models (C+/-) were then developed to differentiate between confirmed positive (n-47) and negative (n-713) diagnostic test results. Models were built using a stage-wise process that increased data granularity with each stage: 1) present-on-admission data; 2) low frequency (LF) medication and laboratory data added; and 3) addition of high frequency (HF) vital sign data, collected at a rate of every 20 seconds. Performance was evaluated at each stage using metrics robust to class imbalance and prioritizing recall (TPR). Results. All models demonstrated improved precision-recall performance with progressive addition of dynamic clinical data. Using dynamic LF and HF data, at a prediction time 24 hours in advance of HA-VTE event, the S+/- NB model TPR was 76% (AUPRC .52, PPV 46%, AUROC .60) and RF and GBDT models identified true negatives with a specificity of 80%, and the C+/- NB model had a 91% TPR (AUPRC .77, PPV 53%, AUROC .68). Dynamic hematologic labs, BP, HR, and RR values were identified as important predictors of HA-VTE event outcomes, with importance varying by time prediction window. Conclusion. Classification algorithms applied to routinely collected dynamic clinical data can produce models with improved HA-VTE risk prediction ability over static data models and have the potential to improve detection of at-risk patients. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}