Rogers Michael P, Janjua Haroon, DeSantis Anthony J, Grimsley Emily, Pietrobon Ricardo, Kuo Paul C
From the Department of Surgery, University of South Florida Morsani College of Medicine, Tampa, FL (Rogers, Janjua, DeSantis, Grimsley, Kuo).
SporeData Inc., Durham, NC (Pietrobon).
J Am Coll Surg. 2022 Apr 1;234(4):652-659. doi: 10.1097/XCS.0000000000000108.
The American College of Surgeons (ACS) NSQIP risk calculator helps guide operative decision making. In patients with significant surgical risk, it may be unclear whether to proceed with "Hail Mary"-type interventions. To refine predictions, a local interpretable model-agnostic explanations machine (LIME) learning algorithm was explored to determine weighted patient-specific factors' contribution to mortality.
The ACS-NSQIP database was queried for all surgical patients with mortality probability greater than 50% between 2012 and 2019. Preoperative factors (n = 38) were evaluated using stepwise logistic regression; 26 significant factors were used in gradient boosted machine (GBM) modeling. Data were divided into training and testing sets, and model performance was substantiated with 10-fold cross validation. LIME provided individual subject mortality. The GBM-trained model was interpolated to LIME, and predictions were made using the test dataset.
There were 6,483 deaths (53%) among 12,248 admissions. GBM modeling displayed good performance (area under the curve = 0.65, 95% CI 0.636-0.671). The top 5 factors (% contribution) to mortality included: septic shock (27%), elevated International Normalized Ratio (22%), ventilator-dependence (14%), thrombocytopenia (14%), and elevated serum creatinine (5%). LIME modeling subset personalized patients by factors and weights on survival. In the entire cohort, mortality positive predictive value with 2 factor combinations was 53.5% (specificity 0.713), 3 combinations 64.2% (specificity 0.835), 4 combinations 72.1% (specificity 0.943), and all 5 combinations 77.9% (specificity 0.993). Conversely, mortality positive predictive value fell to 34% in the absence of 4 factors.
Through the application of machine learning algorithms (GBM and LIME), our model individualized predicted mortality and contributing factors with substantial ACS-NSQIP predicted mortality. USE of machine learning techniques may better inform operative decisions and family conversations in cases of significant surgical risk.
美国外科医师学会（ACS）国家外科质量改进计划（NSQIP）风险计算器有助于指导手术决策。对于手术风险较高的患者，是否进行“孤注一掷”式的干预可能并不明确。为了优化预测，研究人员探索了一种局部可解释的模型无关解释机器（LIME）学习算法，以确定患者特定因素对死亡率的加权贡献。
查询ACS-NSQIP数据库，获取2012年至2019年间所有死亡概率大于50%的手术患者。使用逐步逻辑回归评估术前因素（n = 38）；26个显著因素用于梯度提升机（GBM）建模。数据分为训练集和测试集，并通过10倍交叉验证来证实模型性能。LIME提供个体受试者的死亡率。将GBM训练的模型内插到LIME中，并使用测试数据集进行预测。
12248例入院患者中有6483例死亡（53%）。GBM建模表现良好（曲线下面积 = 0.65，95%CI 0.636 - 0.671）。对死亡率贡献最大的前5个因素（%贡献）包括：感染性休克（27%）、国际标准化比值升高（22%）、呼吸机依赖（14%）、血小板减少（14%）和血清肌酐升高（5%）。LIME建模子集根据因素和生存权重对患者进行个性化分析。在整个队列中，2种因素组合的死亡率阳性预测值为53.5%（特异性0.713），3种组合为64.2%（特异性0.835），4种组合为72.1%（特异性0.943），所有5种组合为77.9%（特异性0.993）。相反，在没有4个因素的情况下，死亡率阳性预测值降至34%。
通过应用机器学习算法（GBM和LIME），我们的模型个性化预测了死亡率和相关因素，且具有较高的ACS-NSQIP预测死亡率。在手术风险较高的情况下，使用机器学习技术可能会更好地为手术决策和与家属的沟通提供依据。
