华法林是心脏瓣膜置换术后最常用的口服抗凝药物，因其治疗窗窄、个体差异大，极易导致抗凝不充分的血栓栓塞或过度抗凝引起出血甚至死亡，因此，精准个体化华法林用药是保证其药物疗效和用药安全的关键。华法林的量效关系受多种因素影响，各因素影响多呈高度非线性，传统研究方法无法准确发现其用药规律，迫切需要采用智能化方法明确精准用药量与影响因素之间的复杂关系，发现其用药规律。本课题基于前期近3万例心脏瓣膜置换术后抗凝治疗完整临床大数据，运用循征医学和大数据挖掘理论，建立华法林个体化治疗动态信息数据库；基于人工神经网络、模糊神经网络和遗传算法等理论，建立华法林治疗的BP神经网络、自适应神经模糊推理、遗传进化等三种智能化模型；据此数据库样本，采用数据库内样本验证和数据库外样本临床对照相结合的分析方法，优选准确度最高模型作为华法林治疗模型用于临床实践，实现华法林智能化个体化精准治疗。

Warfarin is the most commonly prescribed oral anticoagulant for patients after heart valve replacement. However, the fact that warfarin has narrow therapeutic window and obvious individual difference, which can easily cause the phenomenon that inadequate anticogulation lead to thromboembolism or excessive anticoagulation cause bleeding and even death in clinical practice. Therefore, precise individual warfarin dosage is the key to ensure the safety usage and therapeutic effect during the anticoagulation therapy. It is widely accepted that the dose-effect relationship of warfarin is influenced by many factors and highly nonlinear exists among those factors, which makes it difficult for traditional research method to find the principle accurately. Hence, intelligent methods must be used to make the complex relationship between the influencing factors and principle clear. This study is based on previous nearly 30000 complete clinical big data of anticoagulation therapy after heart valves replacement, using the theory of evidence-based medicine and big data mining to establish a dynamic information database of warfarin individualized treatment; Based on the theory of artificial neural network, fuzzy neural network and genetic algorithm, we establish three kinds of intelligent model of warfarin anticoagulant therapy including BP neural network, adaptive neural fuzzy inference and genetic evolution; We combine sample validation from the database with clinical sample validation outside the database to determine the highest accuracy model as warfarin treatment model which can be applied to clinical practice to realized intelligent and precise warfarin individualized treatment.

华法林是心脏瓣膜置换术后抗凝治疗推荐首选抗凝药，由于其治疗窗窄，个体差异性大，以及用药剂量易受饮食、药物等众多因素影响的特殊性，有效、安全、合理的给药难度极大，稍有不慎，极易因抗凝不充分导致血栓栓塞或过度抗凝引发出血甚至死亡。因此，华法林用药被形象比作“钢丝上跳舞”。个体化精准华法林抗凝治疗一直是临床亟待解决和极具挑战性的重要问题。本项目以前期近3万例心脏瓣膜置换术后抗凝治疗完整临床大数据为基础，采用协方差分析与临床强制纳入变量相结合的方法，最终纳入12个变量形成华法林个体化指标变量池；基于最新人工智能机器学习理论和方法，建立华法林治疗2种智能化个体用药预测模型，即反向传播神经网络（BPNN）和遗传算法优化人工神经网络（BP-GA），和1种传统非智能化多元线性回归（MLR）预测模型；并采用①预测剂量与实际剂量比较，②理想预测百分比，③绝对误差均值（MAE）和均方误差（MSE），④不同剂量亚组敏感性分析方法，通过内部验证和外部验证，优选准确度最高模型。结果显示：1）预测剂量与实际剂量的比较：无论是内部还是外部验证组，BPNN和BP-GA模型预测剂量与实际用药剂量无差异（P>0.05），而MLR模型预测剂量与实际用药剂量有差异（P<0.05）；2）总体MAE与MSE值比较：无论在内部验证组还是外部验证组，BP-GA模型MAE和MSE值均小于其他两种模型，但三者间无统计学差异（P>0.05）；3）总体理想预测百分比：在外部验证组，BP-GA模型的理想预测百分比高于MLR模型（P<0.05）；4）剂量亚组敏感性分析，无论内部验证组还是外部验证组，3种模型中剂量组理想预测百分比均最高，均大于77%。在外部验证组，BP-GA在中等剂量组理想预测百分比高于其他两种模型（P<0.05）。综上，智能化BP-GA模型预测效果最优。本项目研究成果为华法林个体化精准抗凝治疗和类似临床问题的解决提供了新思路和新方法；智能化的方法丰富了临床药物治疗研究理论和实践，能有效改善和提高华法林个体化抗凝治疗效果，有望实现华法林智能化个体化精准治疗。