Meguid Robert A, Bronsert Michael R, Juarez-Colunga Elizabeth, Hammermeister Karl E, Henderson William G
*Surgical Outcomes and Applied Research program, University of Colorado School of Medicine, Aurora, CO†Department of Surgery, University of Colorado School of Medicine, Aurora, CO‡Adult and Child Center for Health Outcomes Research and Delivery Science, University of Colorado School of Medicine, Aurora, CO§Department of Biostatistics and Informatics, Colorado School of Public Health, Aurora, CO¶Division of Cardiology, Department of Medicine, University of Colorado School of Medicine, Aurora, CO.
Ann Surg. 2016 Jul;264(1):10-22. doi: 10.1097/SLA.0000000000001677.
To develop parsimonious prediction models for postoperative mortality, overall morbidity, and 6 complication clusters applicable to a broad range of surgical operations in adult patients.
Quantitative risk assessment tools are not routinely used for preoperative patient assessment, shared decision making, informed consent, and preoperative patient optimization, likely due in part to the burden of data collection and the complexity of incorporation into routine surgical practice.
Multivariable forward selection stepwise logistic regression analyses were used to develop predictive models for 30-day mortality, overall morbidity, and 6 postoperative complication clusters, using 40 preoperative variables from 2,275,240 surgical cases in the American College of Surgeons National Surgical Quality Improvement Program data set, 2005 to 2012. For the mortality and overall morbidity outcomes, prediction models were compared with and without preoperative laboratory variables, and generic models (based on all of the data from 9 surgical specialties) were compared with specialty-specific models. In each model, the cumulative c-index was used to examine the contribution of each added predictor variable. C-indexes, Hosmer-Lemeshow analyses, and Brier scores were used to compare discrimination and calibration between models.
For the mortality and overall morbidity outcomes, the prediction models without the preoperative laboratory variables performed as well as the models with the laboratory variables, and the generic models performed as well as the specialty-specific models. The c-indexes were 0.938 for mortality, 0.810 for overall morbidity, and for the 6 complication clusters ranged from 0.757 for infectious to 0.897 for pulmonary complications. Across the 8 prediction models, the first 7 to 11 variables entered accounted for at least 99% of the c-index of the full model (using up to 28 nonlaboratory predictor variables).
Our results suggest that it will be possible to develop parsimonious models to predict 8 important postoperative outcomes for a broad surgical population, without the need for surgeon specialty-specific models or inclusion of laboratory variables.
开发适用于成年患者广泛外科手术的术后死亡率、总体发病率及6种并发症集群的简约预测模型。
定量风险评估工具未常规用于术前患者评估、共同决策、知情同意及术前患者优化，部分原因可能是数据收集负担以及纳入常规外科实践的复杂性。
采用多变量向前选择逐步逻辑回归分析，利用美国外科医师学会国家外科质量改进计划数据集中2005年至2012年2275240例手术病例的40个术前变量，开发30天死亡率、总体发病率及6种术后并发症集群的预测模型。对于死亡率和总体发病率结局，比较包含和不包含术前实验室变量的预测模型，以及通用模型（基于9个外科专科的所有数据）与专科特异性模型。在每个模型中，使用累积c指数检查每个添加的预测变量的贡献。使用c指数、Hosmer-Lemeshow分析和Brier评分比较模型之间的区分度和校准度。
对于死亡率和总体发病率结局，不包含术前实验室变量的预测模型与包含实验室变量的模型表现相当，通用模型与专科特异性模型表现相当。死亡率的c指数为0.938，总体发病率的c指数为0.810，6种并发症集群的c指数范围从感染性并发症的0.757到肺部并发症 的0.897。在8个预测模型中，最初纳入的7至11个变量占完整模型c指数的至少99%（使用多达28个非实验室预测变量）。
我们的结果表明，有可能开发简约模型来预测广泛外科人群的8种重要术后结局，而无需外科医生专科特异性模型或纳入实验室变量。
