基于多模态数据融合的乳腺癌风险预警模型的病例队列研究

Breast cancer is the malignant tumor with the highest incidence among global female population, which threaten women's physical and mental health seriously. At present, because of the differences in race, structure of the breast and living habit and other relevant reasons, the breast cancer risk prediction model including Gail model established by domestic and overseas scholars couldn't to be applied for Chinese women appropriately. Although the two existing models in China which have add SNP into Gail model can improve the screening efficacy to some extent, because of geographical features and technical constraints, they should not achieve a wide range of applications..With the advent of the era of big data, the mathematical model using the traditional method of deterministic algorithm solely is difficult to realize the high-efficiency analysis of mass of multi-modal data changing through longitudinal time axis, and it can't intuitively reflect the law of breast cancer pathogenesis either, so it is an urgent need to find new optimization method through breaking through the bondage of the original mathematical and statistical methods..This issue aim to screen factors set influencing the occurrence, development and outcome of breast cancer from the female breast health assessment index system by combining the method of traditional mathematical model with numerical approximation theory, statistical inference theory and the theory of intelligent optimization method, to build dynamic breast cancer risk assessment model with self-learning ability by making full use of machine learning methods to combine multi-modal female breast health data and form the female breast disease follow-up cohort, which provide the theoretical basis for the establishment of the female breast cancer screening and targeted prevention strategies among high-risk population.

乳腺癌是全球女性发病率最高的恶性肿瘤，严重威胁女性身心健康。目前由于存在种族、乳房结构、生活习惯等方面的差异，国内外学者建立的Gail模型等多种乳腺癌风险预测模型不适用于我国女性。国内已有的两个模型是在Gail模型中纳入SNP，筛选功效有所提高，但因地域特征和技术限制无法实现大范围应用。随着大数据时代的到来，单纯使用传统的数学模型方法的确定性算法很难实现对海量多模态数据的高效处理分析，无法真实直观反映乳腺癌发病的规律，迫切需要突破原有数学和统计方法束缚，找到新的优化方法。本课题拟将传统数学模型方法结合数值逼近理论、统计推断理论、智能优化理论等方法，从女性乳腺健康评估指标体系中筛选影响乳腺癌发生发展和转归的因素集；充分利用机器学习的方法融合多模态的女性乳腺健康数据，构建具有自学习能力的动态乳腺癌风险评估模型和女性乳腺疾病随访队列，为女性乳腺癌高危人群的筛查、针对性预防策略制定等提供理论依据。

乳腺癌是全球女性发病率最高的恶性肿瘤，严重威胁女性身心健康。虽然目前我国属乳腺癌低发国家，但我国人口基数巨大，发病率呈快速上升趋势，乳腺癌的防治工作面临严峻的挑战。国外的预防经验证实，有效的预防可以降低乳腺癌的发病率和死亡率。欧美国家经过多年的探索已建立了一套行之有效的预防模式：积极开展健康促进工作—基于危险因素的疾病风险评估—针对性的合理有效干预措施，达到对疾病的有效防控。其中构建适合中国女性特征的乳腺癌风险评估模型，对于及早发现乳腺癌高危个体和制定科学的乳腺癌预防策略具有重大意义。国内外学者建立的Gail模型等多种乳腺癌风险预测模型不适用于我国女性。国内已有的两个模型是在Gail模型中纳入SNP，筛选功效有所提高，但因地域特征和技术限制无法实现大范围应用。本项目在对乳腺健康评估指标体系进行较为系统全面的探索的基础上，采用多种机器学习方法相结合，筛选女性乳腺癌影响因素集，建立具有自学习能力的动态乳腺癌风险预警模型。同时受社会网络启发，对于基于社区人群数据，采用关联规则算法探索乳腺癌发病危险因素，创新性地提出了适合中国女性的基于社交网络的乳腺癌风险评估模型（BCRAM），通过计算风险相关因素（rrf）的相似度，构造了一个加权或未加权的医疗社交网络，根据高危概率值，将人群鉴定为高风险或低风险。我们将社交网络模型与其他已有的乳腺癌风险评估模型，包括Gail模型、Tyrer-Cuzick模型、相对风险模型（Rrm）及本项目组前期传统方法建立的模型进行了比较。BCRAM的ROC曲线下面积为0.785，优于Rrm（AUC = 0.694），Gail (AUC =0.574)、Tyrer-Cuzick（AUC = 0.694）以及我们前期模型（AUC = 0.722），提示BCRAM的精度最优。该模型弥补了传统统计模型的缺陷，其不依赖固定因素选择与乳腺癌风险相关因素的独特设计方法，使得模型具有较好的通用性。为女性乳腺癌高危人群的筛查、针对性预防策略的制定及基础研究提供理论依据。