Using electronic health records to estimate risk of bleeding during surgery

Bleeding is a known complication of all surgical procedures – in fact it is so common that it is considered best practice to document the estimated blood loss (EBL) at the completion of the procedure. Common bleeding can have disastrous effects including hemodynamic instability and even death. The current best practice uses looking at historical averages of procedures and significant coagulation abnormalities for patients, but lacks specificity at the patient level. We hypothesize that it will be possible to use machine learning techniques to leverage data available from the electronic health records (EHR) prior to surgery in order to predict blood loss on the individual case level.

Using retrospective data from UCSF EHR system, we extracted 16276 surgery cases with demographic data including as well as clinical data like vital signs, lab test results, medication taken and estimated blood volume. Missing vital signs and laboratory tests were.

Two different models were trained. First a continuous model using LASSO with the outcome of interest being the difference in hemoglobin levels before and right after surgery. Second, we attempted to predict massive hemorrhage defined as binary trait using logistic regression with regularization. Data was split to 80% training and 20% held-out testing set. 10-fold cross validation was performed to optimize regularization parameter.

We evaluated the models globally and per-procedure and found the RMSE for a test set was 0.91mL, and the binary model had an AUC of 0.85.