Anderson Geoffrey A, Bohnen Jordan, Spence Richard, Ilcisin Lenka, Ladha Karim, Chang David
Massachusetts General Hospital, GRB 425, 55 Fruit St, Boston, MA, 02114, USA.
University of Cape Town, Cape Town, South Africa.
World J Surg. 2018 Sep;42(9):2725-2731. doi: 10.1007/s00268-018-4535-8.
The focus of many data collection efforts centers on creation of more granular data. The assumption is that more complex data are better able to predict outcomes. We hypothesized that data are often needlessly complex. We sought to demonstrate this concept by examination of the American Society of Anesthesiologists (ASA) scoring system.
First, we created every possible consecutive two, three and four category combinations of the current five category ASA score. This resulted in 14 combinations of simplified ASA. We compared the predictive ability of these simplified scores for postoperative outcomes for 2.3 million patients in the NSQIP database. Individual model performance was assessed by comparing receiver operator characteristic (ROC) curves for each model with the standard ASA.
Two of our 4-category models and one of our 3-category models had ability to predict all outcomes equivalent to standard ASA. These results held for all outcomes and on all subgroups tested. The performance of the three best performing simplified ASA scores were also equivalent to the standard ASA score in the univariate analysis and when included in a multivariate model.
It is assumed that the most granular data and use of the largest number of variables for risk-adjusted predictions will increase accuracy. This complexity is often at the expense of utility. Using the single best predictor in surgical outcomes research, we have shown this is not the case. In this example, we demonstrate that one can simplify ASA into a 3-category variable without losing any ability to predict outcomes.
许多数据收集工作的重点都集中在创建更细化的数据上。其假设是，更复杂的数据能更好地预测结果。我们推测，数据往往不必要地复杂。我们试图通过研究美国麻醉医师协会（ASA）评分系统来证明这一概念。
首先，我们创建了当前五类ASA评分中每一种可能的连续两类、三类和四类组合。这产生了14种简化的ASA组合。我们在NSQIP数据库中比较了这些简化评分对230万例患者术后结果的预测能力。通过比较每个模型与标准ASA的受试者操作特征（ROC）曲线来评估个体模型的性能。
我们的4类模型中有2个和3类模型中有1个预测所有结果的能力与标准ASA相当。这些结果在所有测试结果和所有亚组中均成立。在单变量分析中以及纳入多变量模型时，表现最佳的三种简化ASA评分的性能也与标准ASA评分相当。
人们认为，最细化的数据以及使用最多数量的变量进行风险调整预测会提高准确性。这种复杂性往往是以实用性为代价的。在外科手术结果研究中使用单一最佳预测指标，我们发现情况并非如此。在这个例子中，我们证明可以将ASA简化为一个3类变量，而不会丧失任何预测结果的能力。
