We identified preoperatively 83 anticoagulated patients (86 knees) who had a TKA between 2000 and 2008 (when the switch was made to electronic medical records and at the beginning of the anticoagulation service at our institution). The anticoagulation service database was queried for Current Procedural Terminology codes of TKA and revision TKA, which was then cross-referenced for International Classification of Diseases, 9th Revision (ICD-9) codes of DVT, PE, atrial fibrillation, heart valve, thrombophilia, cerebrovascular accident, and congestive heart failure. Using ICD-9 codes to search the database allowed us to ascertain why the patients in this study were on warfarin. Nine patients (9 knees) from this initial cohort were excluded from further analysis because of variance regarding intraoperative management with respect to tourniquet letdown around the time of wound closure and drain usage. The remaining 74 patients who underwent TKA and were also on warfarin for an unrelated medical comorbidity were included in this data analysis. Thirty-seven patients (38 knees) were maintained on their warfarin through surgery and 39 patients (39 knees) had their warfarin withheld. The anticoagulation service queried patients as to whether or not their warfarin was held preoperatively. For uniformity of comparison, we established an INR of greater than or equal to 1.5 on the day of surgery as verification that warfarin therapy was continued. In the continued warfarin cohort, there was no preoperative adjustment of patients’ warfarin regimen. We had prior Institutional Review Board approval.
All surgeries were performed through a standard medial parapatellar approach under tourniquet control. The tourniquet was released and hemostasis obtained before wound closure in all cases and the drain was removed on postoperative day one. All patients wore venous foot pumps bilaterally, TED hose on the nonoperative leg, and participated in the standard physical therapy protocol for knee reconstruction. For the discontinued warfarin cohort, warfarin was stopped 1 week before surgery and they were bridged with low-molecular-weight heparin (LMWH) through the evening before the procedure. This was then resumed on the night of surgery along with their regular dose of warfarin. The INR was checked preoperatively and each day postoperatively until entering the therapeutic range. LMWH was discontinued when the INR was above 1.5.
Once these two groups were established, key demographic data and the primary end points for comparison were collected. The protocol on the adult reconstruction service at the authors’ institution is for the patient to receive a red blood cell transfusion if the hematocrit falls below 27% and is accompanied by changes in vital signs or urine output. To ensure completeness of data collection, the entire 90-day global period was reviewed for each medical record regardless of outcomes identified. Importantly, there was a risk associated with the use of general anesthetic that these patients were subjected to because they could not receive a spinal anesthetic. This was addressed preoperatively with each patient during the informed consent period. The senior authors (AAH, HKD) believed the risk of general anesthetic was justified when considering the thrombotic risk posed to these high-risk patients. We thus avoided spinal anesthesia in both groups thereby mitigating the potential complication of epidural hemorrhage.
For categorical patient characteristic variables, we compared the study arms with the chi square test or Fisher’s exact test, as appropriate. For categorical variables with many categories, the Fisher-Freeman-Halton test was used, which is the Fisher’s exact test extended to greater than 2 × 2 crosstabulation tables [7
]. For continuous variables, we compared the study arms using the independent samples t test. For change in INR from preoperatively to postoperatively, the groups were compared postoperatively while controlling for preoperatively in an analysis of covariance fashion. We recorded gender, age, procedure, diagnosis necessitating warfarin therapy, and pre- and postoperative INR levels (Table ).
Three patients in the study required reoperation. To assess if this number of repeated measurements required mixed effect models to account for lack of independence in the observations, mixed effect models were fitted to the outcome variables, which were complications and transfusions. The intraclass correlation coefficients were zero, suggesting there was no need for a mixed effects modeling approach, because all 77 surgery outcomes were sufficiently independent. Therefore, we used standard statistical tests and models for all comparisons.
To model the complication outcome, all complications were combined into a binary variable (1 = any complication occurred, 0 = no complication occurred). The severity of the complications was not accounted for with this approach, so the approach is conservative. That is, any other approach would make continuing warfarin appear even more protective against complications, so no bias was introduced in favor of the study hypothesis by taking this binary approach. To compare the outcomes of presence of a complication (binary) and need for blood transfusion (binary) between the two study groups, an exact Poisson regression model was fitted controlling for age at surgery as a covariate. The modified Poisson regression approach permits the direct calculation of a risk ratio in followup studies and so is preferred to logistic regression, which estimates odds ratios. To compare the outcome of the number of units of blood transfused, we again used exact Poisson regression. In this form of the Poisson model, the number of rare events, or units transfused, was modeled, and the risk ratio represented the ratio of the mean number of units required for each study group. Because most patients required zero units, linear regression was not appropriate, because the data were too nonnormally distributed, even if transformation was applied. The exact version of Poisson regression is not affected by overfitting, where unreliable associations can be introduced from having too many predictors for the number of outcomes. These models were appropriate, then, even with the small number of outcome events observed and the inclusion of age as a covariate.