Secondary Logo

Journal Logo

The Spine Blog

Friday, April 24, 2020

Do Complication Prediction Models Help Us?

Traditional surgical decision-making involves weighing risks and benefits of surgery, with surgeons generally advising patients based on likely outcomes and complication rates for the "average" patient. However, such information may not be particularly helpful for individual patients, most of whom are not "average" and may have predicted outcomes and risk profiles far different than the average patient. In order to address this limitation of traditional shared decision-making, researchers have been working to create models predicting outcomes and complication rates based on individual characteristics. Most of these have looked at specific diagnoses and operations and involve computer-based multivariate models. Unfortunately, most surgeons in busy clinical practice do not use these models as they are cumbersome and time-intensive. In order to improve the clinical utility of a risk prediction model, Andrew Broda and colleagues analyzed the NSQIP database that included over 48,000 cervical fusions and nearly 130,000 lumbar procedures (both decompressions and fusions were included) in order to develop a multivariate model. This model was used to create a 0-12 point scoring system based on the number of comorbidities that predicted overall complication rates for cervical and lumbar surgery. Complications were recorded within 30 days of surgery and included postoperative blood transfusion, surgical site infection, DVT, PE, UTI, MI, reintubation, and stroke, among others. Readmission and reoperation were not included in the list of complications. They identified 16 factors independently associated with complication rate including age over 59, ASA score 3-5, hypertension, renal failure, CHF, BMI < 18.5, cancer, and COPD. Each factor was weighted equally in their scoring system. In the lumbar cohort, the overall complication rate was 12.4%, ranging from 4% for those with a score of 0 (25% of patients), to 11% for those with a score of 3 (22% of patients), and to 63% for those with a score of 12 or more (0.2% of patients). The cervical cohort had a similar distribution across scores and a somewhat lower overall complication rate (8%). The authors found good agreement with an internal validation analysis using the 20% of the dataset not used to create the model. Receiver operator curves showed fair predictive accuracy with 0.77 area under the curve for the cervical model and 0.740 for the lumbar model.

The authors have done a nice job using big data to create a fairly accurate complication rate prediction model that uses a simple scoring algorithm for cervical and lumbar procedures. This study has all of the limitations related to using large administrative database studies, though factors such as data miscoding or missing data likely did not affect outcomes much given the large number of patients involved. The major limitations are more related to the patient characteristics and complications used to build the model. In an effort to simplify the model, all complications, ranging from blood transfusion to cardiac arrest, were combined. Similarly, risk factors including disseminated cancer (with an OR of 4.3 for a complication) and African American race (OR 1.2) had equal weight in the scoring system. Failing to differentiate one level laminectomy patients from multilevel fusion patients also leads to a loss of information in the model that likely limited its accuracy.  The authors made these tradeoffs in order to simplify the scoring system and make it more usable in the clinical arena. However, these oversimplifications may have limited how much the model adds to the decision-making process beyond the overall complication rates from large studies looking at the "average" patient. In the lumbar cohort, 70% of patients had a score of 0-3, with predicted complication rates ranging from 4%-11%. A patient could obtain a score of three by being a 60-year old African American with hypertension and no other comorbidities. Similarly, a cancer patient on dialysis requiring a preoperative transfusion would also score 3, despite having a markedly higher complication risk. This model does serve as a good reminder that increasing the number of comorbidities increases complication risk. How much it adds beyond the surgeon's overall gestalt remains to be seen.

Please read Mr. Broda's article on this topic in the May 1 issue. Do you think this scoring system would be useful in your practice?

Adam Pearson, MD, MS

Associate Web Editor