Seeing the Future: Predicting a Patient’s Need for Shoulder Surgery before the First Encounter

OBJECTIVES: The aim of this study was to determine the likelihood of shoulder surgery based on a pre-visit branching questionnaire implemented prospectively at the time of initial visit scheduling. METHODS: Patients calling a large regional sports health institution with shoulder complaints between Jan 2015 and June 2016 were asked a series of questions according to a branching logic algorithm at the time of initial appointment scheduling (Fig. 1). All patients had appointments scheduled regardless of their responses. In July 2016, a retrospective chart review was conducted to determine which patients were recommended for shoulder surgery. Multivariate regression models were constructed to determine the combination of questions that were asked, or could be asked, that would lead to the highest and most accurate predictive value of recommended surgery. Patient records were excluded if the patients were younger than 13 or over 75, if the appointment was cancelled or scheduled after April 2015, and if the treatment was not yet determined at the time of chart review. RESULTS: After chart review of included patients, 760 records were available for analysis. The multivariate regression model that best matched the data and produced the highest predictive value for surgery had a concordance index of 0.688, representing the rate at which the model correctly assigned a higher surgical risk to patients that were ultimately recommended for surgery against those who were not. Significant variables in this model were if a previous provider ordered an MRI for the patient, injury status, and patient sex. The odds ratios for a patient requiring surgery based on their status in those areas are shown in Table 1. Having an MRI ordered by a previous provider (OR=4.45) and male sex (OR=1.6) were both positive predictors of needing surgery. Indication of injury with a primary complaint of weakness or instability carried the strongest predictive effect of surgery. (OR=1, reference) The odds of surgery decreased if the patient’s primary complaint was pain or if the patient followed the answer pathway: Pain—Not Crushing Pain—Injury—No ER Visit—No Pain Raising Arm. The model can predict between a 7.5% and 95% chance of needing surgery (20% of the population required surgery). A nomogram was constructed from the model such that a patient’s response to each question correlated to a point value, and the total of those points correlated to a probability of needing surgery. CONCLUSION: Based on patient’s response to the questionnaire, we have constructed a model that can both quickly and easily estimate the probability that the patient will require surgery. Our model can predict up to a 95% likelihood of needing surgery and down to a 7.5% likelihood of needing surgery. We believe that this information can guide and improve future scheduling practices and will help patients see the appropriate provider sooner, reduce cost, and improve patient and physician satisfaction.