Das Ananya, Ben-Menachem Tamir, Cooper Gregory S, Chak Amitabh, Sivak Michael V, Gonet Judith A, Wong Richard C K
Division of Gastroenterology, Department of Medicine, University Hospitals of Cleveland, Case Western Reserve University, Cleveland, OH 44106-5066, USA.
Lancet. 2003 Oct 18;362(9392):1261-6. doi: 10.1016/S0140-6736(03)14568-0.
Models based on artificial neural networks (ANN) are useful in predicting outcome of various disorders. There is currently no useful predictive model for risk assessment in acute lower-gastrointestinal haemorrhage. We investigated whether ANN models using information available during triage could predict clinical outcome in patients with this disorder.
ANN and multiple-logistic-regression (MLR) models were constructed from non-endoscopic data of patients admitted with acute lower-gastrointestinal haemorrhage. The performance of ANN in classifying patients into high-risk and low-risk groups was compared with that of another validated scoring system (BLEED), with the outcome variables recurrent bleeding, death, and therapeutic interventions for control of haemorrhage. The ANN models were trained with data from patients admitted to the primary institution during the first 12 months (n=120) and then internally validated with data from patients admitted to the same institution during the next 6 months (n=70). The ANN models were then externally validated and direct comparison made with MLR in patients admitted to an independent institution in another US state (n=142).
Clinical features were similar for training and validation groups. The predictive accuracy of ANN was significantly better than that of BLEED (predictive accuracy in internal validation group for death 87% vs 21%; for recurrent bleeding 89% vs 41%; and for intervention 96% vs 46%) and similar to MLR. During external validation, ANN performed well in predicting death (97%), recurrent bleeding (93%), and need for intervention (94%), and it was superior to MLR (70%, 73%, and 70%, respectively).
ANN can accurately predict the outcome for patients presenting with acute lower-gastrointestinal haemorrhage and may be generally useful for the risk stratification of these patients.
基于人工神经网络（ANN）的模型有助于预测各种疾病的预后。目前尚无用于急性下消化道出血风险评估的有效预测模型。我们研究了利用分诊时可用信息的人工神经网络模型是否能够预测该疾病患者的临床结局。
根据急性下消化道出血患者的非内镜检查数据构建人工神经网络和多因素逻辑回归（MLR）模型。将人工神经网络将患者分为高风险和低风险组的性能与另一个经过验证的评分系统（BLEED）进行比较，结局变量包括再出血、死亡以及控制出血的治疗干预措施。人工神经网络模型采用最初12个月内收治于主要机构的患者数据（n = 120）进行训练，然后采用接下来6个月内收治于同一机构的患者数据（n = 70）进行内部验证。随后，人工神经网络模型在另一个美国州的独立机构收治的患者（n = 142）中进行外部验证，并与多因素逻辑回归模型进行直接比较。
训练组和验证组的临床特征相似。人工神经网络的预测准确性显著优于BLEED（内部验证组中死亡预测准确性为87% 对21%；再出血预测准确性为89% 对41%；干预预测准确性为96% 对46%），且与多因素逻辑回归模型相似。在外部验证中，人工神经网络在预测死亡（97%）、再出血（93%）和干预需求（94%）方面表现良好，且优于多因素逻辑回归模型（分别为70%、73%和70%）。
人工神经网络能够准确预测急性下消化道出血患者的结局，可能普遍适用于这些患者的风险分层。
