Dybowski R, Weller P, Chang R, Gant V
Infection and Immunity Laboratory, UMDS, London, UK.
Lancet. 1996 Apr 27;347(9009):1146-50. doi: 10.1016/s0140-6736(96)90609-1.
Decisions about which patients to admit to intensive care and how long to keep them there are difficult. A flexible computer-based mathematical model which is sensitive to the complexity of intensive care medicine, and which accurately models prognosis, seems highly desirable.
We have created, optimised by genetic algorithms, trained, and evaluated the performance of an artificial neural network (ANN) in the clinical setting of systemic inflammatory response syndrome and haemodynamic shock. 258 patients were selected from an intensive care database of 4484 patients at a London teaching hospital and randomised to a network training set (168) and a test set (90). The outcome evaluated was death during that hospital admission and the performance of the neural net was compared (by receiver operating characteristic [ROC] curves and by Brier scores) with that of a logistic regression model.
Artificial neural network performance increased with successive generations; the best-performing ANN was created after 7 generations and predicted outcome more accurately than the logistic regression model (ROC curve area 0.863 vs 0.753).
In this study, ANNs have lent themselves particularly well to modelling a complex clinical situation; we suggest that this relates to their inherently flexible nature which accommodates interactions between the clinical input fields. In addition, we have demonstrated the value of a second computational technique (genetic algorithms) in "tuning" ANN performance. These techniques can potentially be implemented in individual intensive care units; the outcome models which they will generate will be sensitive to local practice. Analysis of such accurate clinical outcome models may empower clinicians with a hitherto unappreciated degree of insight into those elements of their clinical practice which are most relevant to their patients' outcome.
决定哪些患者应收入重症监护病房以及在该病房应留观多长时间是困难的。一个对重症医学的复杂性敏感且能准确模拟预后的灵活的基于计算机的数学模型似乎非常必要。
我们创建了一个人工神经网络（ANN），通过遗传算法对其进行优化、训练，并在全身炎症反应综合征和血流动力学休克的临床环境中评估其性能。从伦敦一家教学医院的4484例患者的重症监护数据库中选取258例患者，随机分为网络训练集（168例）和测试集（90例）。评估的结局是此次住院期间的死亡情况，并将神经网络的性能与逻辑回归模型的性能进行比较（通过受试者工作特征曲线[ROC]和Brier评分）。
人工神经网络的性能随着代数的增加而提高；第7代后创建的表现最佳的人工神经网络比逻辑回归模型更准确地预测了结局（ROC曲线面积分别为0.863和0.753）。
在本研究中，人工神经网络特别适合对复杂的临床情况进行建模；我们认为这与其固有的灵活性有关，这种灵活性能够适应临床输入字段之间的相互作用。此外，我们展示了第二种计算技术（遗传算法）在“调整”人工神经网络性能方面的价值。这些技术有可能在各个重症监护病房中实施；它们生成的结局模型将对当地的实践敏感。对这种准确的临床结局模型进行分析可能会使临床医生对其临床实践中与患者结局最相关的那些因素有前所未有的深入了解。
