Decision Tree-based Modelling for Identification of Predictors of Blood Loss and Transfusion Requirement After Adult Spinal Deformity Surgery

INTRODUCTION

Blood loss requiring significant transfusion requirement is often associated with multilevel fusion for adult spinal deformity (ASD). Surgical factors including number of vertebral levels incorporated in the fusion, duration of exposure, patient position, blood pressure, and spinal osteotomies and complex realignment procedures performed, are recognized risk factors for increased intraoperative bleeding.^1–4 Patient factors that have been elucidated include classification and severity of the spinal deformity, as well as patient age, height, weight, and preoperative comorbidities.^5,6

Although red blood cell (RBC) transfusion is perhaps safer than it ever has been, it still may be associated with adverse events including transfusion reaction, transfusion-related lung injury, postoperative complications, and infection.^7–10 There is growing evidence that blood transfusion may increase intensive care unit and overall length of hospital stay, as well as increase mortality.^11,12 Although the associations described in the literature do not constitute causality, most deformity surgeons universally agree that minimizing blood loss and transfusion requirement is a key tenet of successful surgery. In an era of increased focus on healthcare expenditures, it is also impossible to overlook the direct cost of 1 unit of RBC in the United States, which ranges from 700 to 1200 dollars, and the resulting economic impact.¹³ Anticipating blood loss a priori in patients at known high risk for blood loss is critical for optimal preoperative transfusion planning, and potential employment of surgical strategies such as antifibrinolytic use and intraoperative cell salvage devices to reduce utilization of allogeneic blood.^14–18

Although studies have evaluated the adult spine surgical population as a whole, with regards to predictors for blood loss and transfusion, such a population is relatively heterogenous in its composition of patient characteristics and surgical procedures formed. Previous studies have looked at only parts of the puzzle, assessing a subset of variables that may be related to blood loss in adult deformity surgery, or utilizing cohorts of patients with incomplete data. Although regression models have been the benchmark to detect main effects, proper assessment of the interactions between risk factors can significantly increase the number of potential models, making model selection more problematic.¹⁹ Further, it is our experience that surgeons often intuitively want to classify patients into high or low risk, rather than deploy risk ratios generated by regression methods.

To that end, we sought to predict intraoperative blood loss and perioperative RBC transfusion using machine learning methods, specifically conditional inference tree analysis.^20,21 The primary objective of this study is to determine predictive factors associated with blood loss and transfusion in a large and relatively homogenous cohort of patients with ASD from a single institution, multisurgeon database. Secondarily, we used conditional inference tree analysis to identify and characterize combinations of variables that best predict highest risk for intraoperative blood loss and perioperative RBC transfusion, and to define cutoff values for these variables when possible.

MATERIALS AND METHODS

We retrospectively reviewed our institution's prospectively collected database to identify patients who were aged 18 years or older when they underwent posterior spinal fusion (PSF) for ASD, with a minimum of 4 levels fused, from January 2012 through December 2017. We identified 909 such patients with complete surgical and inpatient admission data.

RESULTS

Predictors of Percent Blood Volume Lost and Perioperative RBC Transfusion by Conditional Interference Tree Analysis

Conditional inference tree analysis was used to identify which variables, or combination of variables, were best suited to predict increased intraoperative blood loss and perioperative RBC transfusion. Of the predictors that were examined for percent blood volume lost, the analysis identified number of levels fused (splitting at 6 and 13 levels), pelvic fixation, ASA score (splitting at 1), 3-column osteotomy, posterior column osteotomy, and preoperative hemoglobin (splitting at 13.6 g/dL) as significant risk factors for increased blood loss. See Figure 1 for full description of conditional inference tree nodes as well as incremental percent blood volume lost added by each risk factor.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Classification and decision tree model for intraoperative percent blood volume (BV) lost in different risk groups of patients who underwent surgery for ASD. Performance measures: ≤ 6 levels: R^2-Adj = 0.58; 6 to 13 levels: R^2-Adj = 0.51; > 13 levels: R^2-Adj = 0.64.

Of the predictors that were examined for intraoperative RBC transfusion, the analysis identified number of levels fused (splitting at 13 levels), pelvic fixation, ASA score (splitting at 1), 3-column osteotomy, posterior column osteotomy, hypertension, preoperative hemoglobin (splitting at 13.6 g/dL), and surgical duration (splitting at 8 hours) as significant risk factors for increased volume of intraoperative RBC units transfused. See Figure 2 for full description of conditional inference tree nodes as well as incremental units of intraoperative RBCs added by each risk factor.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Classification and decision tree model for intraoperative RBCs transfused in different risk groups of patients who underwent surgery for ASD. Performance measures: ≤ 13 levels: R^2-Adj = 0.43; > 13 levels: R^2-Adj = 0.59.

Of the predictors that were examined for postoperative RBC transfusion, the analysis identified number of levels fused (splitting at 13 levels), pelvic fixation, ASA score (splitting at 2), age (splitting at 65 years), and surgical duration (splitting at 6 hours) as significant risk factors for increased volume of postoperative RBC units transfused. See Figure 3 for full description of conditional inference tree nodes as well as incremental units of postoperative RBCs added by each risk factor.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

Classification and decision tree model for units of postoperative RBCs transfused in different risk groups of patients who underwent surgery for ASD. Performance measures: ≤ 13 levels: R^2-Adj = 0.41; > 13 levels: R^2-Adj = 0.45.

DISCUSSION

Significant blood loss in ASD surgery is not an infrequent occurrence and has the potential to cause irreversible end-organ damage, such a renal failure, cardiovascular events including stroke, and mortality. Perhaps a less important point in the mind of the surgeon, but nevertheless unavoidable in an era of healthcare cost containment, is the additional cost of care for a transfused patient including the cost of RBCs, FFP, platelets, and use of antifibrinolytics and cell salvage systems.^26,27 Effectively predicting which patients are at an increased risk for sustaining greater intraoperative blood loss and requiring perioperative RBC transfusion arms surgeons with the knowledge needed for more optimized preoperative surgical planning, and enhanced utilization of healthcare resources.^28,29 In this regard, we queried a large, single institution, multisurgeon database to determine the subset of variables that best predict intraoperative blood loss and transfusion requirement in patients undergoing long PSF for ASD. Using machine learning methodology, we found that number of levels fused, pelvic fixation, ASA score > 1, 3-column osteotomy, and posterior column osteotomy were predictive of increased intraoperative blood loss, with the highest additional risk conferred by greater than 13 levels fused during surgery, and a preoperative hemoglobin < 13.6 g/dL. Greater than 13 levels fused, pelvic fixation, ASA score > 1, 3-column osteotomy, and a history of hypertension were associated with the highest risk for requiring increased intraoperative RBC transfusion, whereas greater than 13 levels fused, pelvic fixation, and an ASA score > 2 conferred the highest risk for requiring increased postoperative RBC transfusion.

Our results are consistent with previous studies using conventional regression methodology, which have reported associations between number of levels fused, pelvic fixation, 3-column osteotomy, and perioperative transfusion requirement. Basques et al³⁰ looked at 4223 spine surgery patients using the American College of Surgeons National Surgical Quality Improvement Program (NSQIP) database and found that age > 60 years, ASA score > 3, greater than 3 levels fused, and operative time > 310 minutes were predictive of increased RBC transfusion. The population of patients queried had undergone posterior primary lumbar fusion only. Berenholtz et al looked at 786 patients who underwent primary spine surgery, and found that age > 66 years, and diabetes were independently associated with allogenic transfusion.³¹ Importantly, the patient population was relatively heterogeneous, as it included any case with a primary procedure code for spine surgery. In 112 patients who underwent revision lumbar decompression and posterior fusion, 1 study found that number of levels fused and preoperative hemoglobin were significant predictors of intraoperative blood loss.⁴

Predictors of blood loss and perioperative transfusion in patients undergoing ASD surgery has been less well studied, and there remains a paucity of data. White et al³² used NSQIP to evaluate 5805 patients who underwent long PSF and found that ASA > 3, increasing construct length, fusion to the pelvis, and performance of osteotomy were independently associated with increased perioperative blood transfusion. From another analysis or the NSQIP, the same group found that patients with ASD with higher frailty scores had higher rates of blood transfusion.³³ It stands to reason that the operative variables identified in this study to be predictive of increased blood loss and perioperative transfusion are those that signify a greater level of surgical complexity. Fusion constructs spanning a greater number of levels require wide exposure, and prolonged surgical duration. Similarly, performance of 3-column osteotomies is known to carry the potential for extensive blood loss.^34,35

Although the question posed in our study is not a new one, the strength of this analysis is the use of conditional inference tree analysis, which allowed us to model complex interactions between variables. The 3 subpopulations generated to predict risk of intraoperative blood loss, by levels fused and types of procedures done, may assist surgeons in stratifying patients into groups of increasing risk. Surgical variables that characterized the highest risk group for intraoperative blood loss were patients who had over 13 levels fused, with performance of 3-column and/or posterior column osteotomies, and pelvic fixation. Further, we were able to characterize the incremental blood loss added by these procedures during surgery, with incorporation of pelvic fixation predicting an additional 20% blood volume lost, when performed for fusions > 13 levels. Pelvic fixation has been shown to be independently associated with increased intraoperative and postoperative RBC transfusion in other studies in patients with ASD.³⁶ Importantly, in the patient population studied, pelvic fixation denoted only iliac or S2-alar-iliac fixation, with the majority being the former. However, even for iliac fixation multiple techniques may be utilized including placing dual iliac screws, and a kickstand or flying buttress construct. These more complex constructs may be associated with increased operative time, and blood loss, which was not evaluated in the present study.

In terms of modifiable patient characteristics, we observed that a preoperative hemoglobin < 13.6 g/dL was predictive of a 14.6% increase in blood loss in cases where over 13 levels were fused, and an additional 0.25 U of intraoperative RBC transfusion in cases where less than 6 levels were fused. Previous studies have demonstrated that decreased preoperative hemoglobin was predictive of increased length of hospital stay, and postoperative delirium.³⁷ As the ASD surgical population is becoming progressively older, and surgeries are more commonly done tailored to age-appropriate alignment objectives, increased understanding of patient comorbidities and nutritional deficiencies that may be related to or causing the preoperative anemia are warranted. As most instances of preoperative anemia are not related to the spinal deformity being treated, it will most likely be present and persist after surgery. These data support the recommendation that preoperative anemia be investigated prior to surgery for its cause and appropriate treatment, to render the surgical patient “fit to fly.”

The strengths of this study are the relatively homogenous population of patients with ASD with long fusion constructs and equivalent case complexity. The retrospective nature of our study could be perceived as a weakness; however, most data were collected prospectively. All patients were from a single institution and treated by surgeons who treat a large volume of adults with spinal deformity, which may limit the generalizability of our results. However, the exclusivity also conferred uniformity, and one might expect these types of surgical procedures to be performed at a tertiary care center by surgeons with similar experience.

CONCLUSION

Fusion spanning 13 vertebral levels or greater, ASA score > 1, 3-column osteotomy, and pelvic fixation are consistent risk factors for increased percent blood volume lost and perioperative RBC transfusion. The best predictors for the subgroup at highest risk for intraoperative blood loss can be identified using conditional inference tree analysis; these included greater than 13 levels fused, pelvic fixation, ASA score > 1, 3-column osteotomy, posterior column osteotomy, and preoperative hemoglobin < 13.6 g/dL. Patients who received major blood transfusion intraoperatively, as defined by > 4 U RBCs (10.8% of all patients), were significantly older (60.2 vs 41.8 years), had higher ASA (2.6 vs 2.4), underwent greater number of levels fused (13.2 vs 9.6), and had lower preoperative hemoglobin (13.1 vs 13.7) (P < .0001 for all). By multivariate regression analysis, patients with a preoperative hemoglobin > 13.0 g/dL were less likely to require major blood transfusion (odds ratio: 0.52, P = .046). An understanding of risk factors for blood loss and transfusion requirement for ASD surgery can facilitate preoperative surgical planning and perioperative management.