基于相似度学习的糖尿病个性化诊疗方法研究

Personalized medicine is an effective way to improve the level of diabetes diagnosis and treatment resulting from complicated nosogenesis of diabetes and individual different clinical features of patients. At present, the related research results are mainly carried out by genomic and other omic researches, which are only enough to support the personalized diagnosis and treatment of a small number of diabetes patients in practical clinical environment unfortunately. With the continuous accumulation data produced in the process of medical treatment and the personal home healthcare, the project plans to carry out a data-driven disease personalized diagnosis and treatment methods through patient similarity and drug similarity calculation and learning. First, due to the long and flexible characteristics of diabetes mellitus clinical course, changes in the patient's condition during the clinical course are brought into the patient similarity learning of patients by constructing patients clinical trajectory, to make personalized patient subtype classification more accurate. Secondly, drug chemistry structure and drug target information are fusion with patient similarity network, to provide patients with personalized treatment recommendations, which could avoid the homogenization problem caused by solutions based on only similarity learning of patient characteristics and make the treatment recommendations more personalized. The results of this project will on the one hand help support the promotion of personalized diagnosis and treatments to a large scale in the practical clinical environment at the present stage, on the other hand serve as the basis of further genetic variation study of diabetes, thus promoting the development of personalized medicine of diabetes.

糖尿病致病原因复杂，其患者临床表现个体化差异很大，对其进行个性化诊疗在糖尿病诊疗中具有重大意义。然而目前相关研究主要利用基因组学等技术展开，其研究成果尚不能在临床上普及应用。随着患者就医全过程以及个人家庭护理过程中的数据不断积累，本项目拟从患者相似性和药物相似性学习入手，探寻数据驱动的疾病个性化诊疗方法解决方案。首先，针对糖尿病病程长、病情变化多变的特性，通过构建患者临床轨迹将患者病情变化规律引入患者相似性的构建当中，使个性化患者分型更加精准；其次，本研究将包含药物化学结构和药物靶点信息的相似性网络与患者相似性网络融合，在为患者提供个性化治疗方案推荐时，避免了单纯从患者相似性出发的治疗方式同质化问题。研究结果一方面可辅助现阶段大规模糖尿病个性化诊疗工作的开展，另一方面可为进一步遗传变异研究提供依据，从而推动糖尿病个性化诊疗的发展。

随着信息学技术在生物医学领域应用的不断深入，数据驱动的方法不断涌现来解决生物医学领域的关键问题。本项目基于电子病历大数据和药物大数据，从患者相似性与药物相似性两方面入手，探寻数据驱动的疾病个性化诊疗方法研究。本项目提出了一系列基于患者临床轨迹的时序体征特征提取方法，构建患者相似度网络，并在此基础上融合药物综合相似度网络，构建患者个性化用药推荐模型。此外，本项目还在原有研究内容上进行了一定拓展，构建了面向多中心的相似病例检索系统，并初步构建了药物深度知识图谱。患者相似性分析是对高年资医生思维的模拟，从相似性分析入手的数据驱动方法较其他机器学习方法相比具有更高的临床接受度；药物相似性分析是制药领域的重要手段，而本项目将其与患者临床数据结合起来，为真实世界研究带了新的方法和思路。本项目的研究成果一方面可以辅助临床，帮助医生进行临床决策，另一方面可以为药物重定位研究提供新的启发和思路，对真实世界新药研发和药物评价都具有重要意义。