McGonigal M D, Cole J, Schwab C W, Kauder D R, Rotondo M F, Angood P B
Department of Surgery, University of Pennsylvania Medical Center, Philadelphia 19104.
J Trauma. 1993 Jun;34(6):863-8; discussion 868-70. doi: 10.1097/00005373-199306000-00018.
This study examined the application of an artificial intelligence technique, the neural network (NET), in predicting probability of survival (Ps) for patients with penetrating trauma. A NET is a computer construct that can detect complex patterns within a data set. A NET must be "trained" by supplying a series of input patterns and the corresponding expected output (e.g., survival). Once trained, the NET can recall the proper outputs for a specific set of inputs. It can also extrapolate correct outputs for patterns never before encountered. A neural network was trained on Revised Trauma Score, Injury Severity Score, age, and survival data contained in 3500 of 8300 state registry records of all patients with penetrating trauma reported in Pennsylvania from 1987 through 1990. The remaining 4800 records were analyzed by TRISS, ASCOT, and the trained NET. Sensitivity (accuracy of predicting death) and specificity (accuracy of predicting survival) were 0.840 and 0.985 for TRISS, 0.842 and 0.985 for ASCOT, and 0.904 and 0.972 for the neural network. This represents a decrease in the number of improperly classified ("unexpected") deaths, from 73 for TRISS and 72 for ASCOT, to 44 for the neural network. The increased sensitivity was statistically significant by Chi-square analysis. The NET for penetrating trauma provided a more sensitive but less specific technique for calculating Ps than did either TRISS or ASCOT. This translated into a 40% reduction in the number of deaths requiring review, and the potential for more efficient use of quality assurance resources.
本研究探讨了一种人工智能技术——神经网络（NET）在预测穿透性创伤患者生存概率（Ps）方面的应用。神经网络是一种计算机结构，能够检测数据集中的复杂模式。必须通过提供一系列输入模式和相应的预期输出（例如生存情况）来“训练”神经网络。一旦经过训练，神经网络就能为特定的一组输入调出正确的输出。它还能对从未遇到过的模式推断出正确的输出。利用1987年至1990年宾夕法尼亚州报告的8300例穿透性创伤患者的州登记记录中的3500例患者的修正创伤评分、损伤严重程度评分、年龄和生存数据对神经网络进行了训练。其余4800条记录由TRISS、ASCOT和经过训练的神经网络进行分析。TRISS的灵敏度（预测死亡的准确性）和特异性（预测生存的准确性）分别为0.840和0.985，ASCOT为0.842和0.985，神经网络为0.904和0.972。这意味着错误分类（“意外”）死亡的数量从TRISS的73例和ASCOT的72例减少到神经网络的44例。通过卡方分析，灵敏度的提高具有统计学意义。与TRISS或ASCOT相比，用于穿透性创伤的神经网络在计算Ps方面提供了一种更灵敏但特异性较低的技术。这使得需要复查的死亡人数减少了40%，并有可能更有效地利用质量保证资源。
