LAUSR

BACKGROUND: The Swenson model was developed as a linear regression model to estimate postoperative day 1 hematocrit levels after benign hysterectomy. Predictive variables included preoperative hematocrit level, patient weight, estimated blood loss, intraoperative crystalloid volume, preoperative platelet count, and hysterectomy route that predicted postoperative day 1 hematocrit level at ±5% for 100% of patients who used an internal validation set. OBJECTIVE: We aimed to validate externally the Swenson model using our hysterectomy population and to further validate the model in a cohort that included robotic‐assisted cases. STUDY DESIGN: In a retrospective cohort, data were collected from benign hysterectomies from April 2014 through May 2016. Predicted postoperative day 1 hematocrit level was calculated with the use of the Swenson equation and compared with measured hematocrit values. We compared our results to the Swenson model results using chi‐square or Fisher's exact test. We then restricted our analysis to those with actual postoperative day 1 hematocrit level of ≤30%, to determine whether the model performed accurately in this subgroup that may need intervention. We generated a receiver‐operating characteristic curve with Louden Index to determine the best cut‐point from the Swenson hematocrit level projections for the prediction of an actual hematocrit level of ≤30%. Furthermore, we stratified the Swenson model predicted hematocrit level into 4 ranges: <32%, 32–35%, 35–38%, and >38%. This stratification allowed us to assess the differential accuracy of the Swenson model across hematocrit level ranges. RESULTS: Of 602 hysterectomies, 478 patients had all the variables that were needed for the Swenson model and postoperative day 1 hematocrit level for comparison. The Swenson model was significantly less accurate in our data compared with their validation set with fewer patients whose predicted hematocrit level was accurate at different thresholds from ±1% through ±5% of actual hematocrit level (all P<.001). Only 76.8% of the predicted hematocrits were accurate within ±5%. Analysis of variance showed accuracy that was similar among different surgical routes (P=.193). A quadratic best‐fit curve showed accuracy was maximized when hematocrit level of 36.2%. Projected hematocrit level was ±2.5% of actual but deteriorated at higher and lower hematocrit level values. When the analysis was restricted to those patients with postoperative day 1 hematocrit levels of ≤30%, accuracy was worse, with only 55.3% of predicted hematocrit level values within ±5%. In this subset, the Swenson equation was more likely to overestimate hematocrit level and give false reassurance. A receiver‐operating characteristic curve analysis showed that the best Swenson cut‐point for the prediction of an actual hematocrit level of ≤30% was 31.9% (sensitivity, 75.5%; specificity, 64.0%). Finally, predicted hematocrit level was divided into 4 groups: <32%, 32–35%, 35–38%, and >38%. When predicted hematocrit level was <32% (n=164), the model was more likely to under‐predict hematocrit level and was least accurate in the subset of patients who were most likely to need intervention. When predicted hematocrit level was 32–35% (n=192), 17.2% of the patients (approximately 1 in 6) had an actual postoperative day 1 hematocrit level of ≤30%. When a predicted hematocrit level of ≥35% was used as a cut‐off point, the percentage of patients who had an actual postoperative day 1 hematocrit level of ≤30% dropped to 8.2%. No patients had an actual hematocrit level of <24%, which makes this a reasonable choice for screening for anemia. CONCLUSION: Although the Swenson model was reliable for the prediction of postoperative day 1 hematocrit level in their internal validation set, it did not perform as well in our hysterectomy population. It may have utility as a screening tool if the projected hematocrit level was ≥35%. Further research is needed to create a model for the prediction of postoperative day 1 hematocrit level that can be incorporated into standard practice.