Towards a viable Artificial Intelligence companion for the healthcare management of Non-Alcoholic Fatty Liver Disease (NAFLD)

NAFLD是指很少或不饮酒的人肝细胞中脂肪的不健康堆积。NAFLD包括多种疾病严重程度，早期脂肪变性主要是良性的，但晚期纤维化和肝硬化与更严重的肝脏疾病和死亡率密切相关。NAFLD的一个主要挑战是从良性到更严重阶段的进展具有很大的可变性；因此，很难评估导致晚期进展的特征以及何时会发生。另一个问题是NAFLD诊断的金标准是通过肝活检，这是一种难以置信的侵入性手术，通常要到疾病的后期，不可逆转的阶段才会进行。该项目的高级目标是开发数据科学和机器学习方法，以支持NAFLD的精确医学解决方案，促进强大的早期诊断，准确的风险分层和精确的预后。研究人员此前曾多次尝试开发NAFLD的人工智能伴侣，重点是利用常规临床程序获得的数据取代目前对NAFLD的侵入性诊断措施，并确定新的生物标志物组合，以取代现有的指示疾病严重程度的替代评分。这些研究通常使用ML或DL算法对通过非侵入性方式（如血清分析）收集的数据进行处理。尽管一些研究人员取得了令人鼓舞的成果，但仍存在一些局限性。几乎所有的研究都集中在NAFLD的诊断上，而忽略了预后和事件时间分析。数据科学问题也限制了这些研究的发现：医疗数据本质上是有缺陷的，数据集的特征是稀缺和稀疏，导致ML算法使用的训练集明显小于原始队列；神经网络的可解释性问题在高度准确但难以理解的算法之间提出了权衡；此外，还缺乏多组学数据的使用，以及在人群中缺乏异质性。考虑到以前作品的优势和局限性，我们目前有4个我们希望实现的广泛目标。首先是整合表型多组学数据和肝损伤的组织学特征，以确定具有特定疾病驱动因素的患者亚群。其次，我们希望通过将传统的有监督的ML和DL分类方法与生存分析相结合，在个体和亚群体的基础上建立疾病预测模型，以确定导致NAFLD纵向进展的因素。第三，我们希望确定患者特异性的临床生物标志物，可以反映脂肪变性、脂肪性肝炎（NASH）的等级和纤维化阶段——通过这种方法，我们的目标是能够通过常规血液检查而不是活检来诊断和预测个体。我们的最终目标是比较拓扑数据分析和机器学习在数字化活检图像分类中的应用，以确定更严重的NASH、纤维化和肝硬化中肝脏的良性脂肪变性。到目前为止，该项目利用了欧洲NAFLD注册表，这是欧洲最大的NAFLD数据集，包括历史NAFLD病例以及自2018年1月1日以来获得的NAFLD病例，该数据集是IMI-2资助的LITMUS项目的一部分，该项目是一项积极的研究，从13个欧洲国家的bbbb25个站点招募。该登记处的患者从基线开始随访10年，以便对分散在NAFLD疾病进展谱上的患者进行纵向分析。收集的数据包括个人的临床信息、肝脏组织病理学、活检样本等。来自登记处的多组学数据的使用也被用于更好地了解患者之间肝损伤的可变性。

NAFLD is the unhealthy build-up of fat in hepatocytes in individuals with little-to-no-alcohol consumption. NAFLD encompasses a large spectrum of disease severity, with early-stage steatosis being largely benign, however latter stage fibrosis and cirrhosis being strongly linked to more severe liver conditions and mortality. A major challenge of NAFLD is the progression from benign to more serious stages are characterised by a substantial level of variability; it is therefore difficult to assess the characteristics that contribute to advanced stage progression and when this will occur. Another issue is the gold standard for NAFLD diagnosis is via liver biopsy, an incredibly invasive procedure, often not performed until latter, irreversible stages of the disease.The high-level aim of this project is to develop Data Science and Machine Learning methods that support a precision medicine solution to NAFLD, fostering robust early diagnosis, accurate risk stratification and precise prognosis. Several attempts have been made previously by researchers to develop an AI companion for NAFLD, with a strong focus on the replacement of current invasive diagnostic measures for NAFLD using data obtained from routine clinical procedures and to identify novel combinations of biomarkers that can replace existing surrogate scores that indicate disease severity. These studies have typically used ML or DL algorithms upon data collected through non-invasive modalities, such as blood serum analysis. Despite some researchers yielding promising results, several limitations exist. Almost all studies have focuses upon diagnosis of NAFLD, with prognosis and time-to-event analysis being omitted. Data Science issues have also limited the findings for these studies: medical data is inherently flawed with datasets being characteristically scarce and sparse resulting in training sets used by ML algorithms being significantly smaller than the original cohort; interpretability issues of neural networks raise a trade-off between highly accurate but poorly understood algorithms; and also a lack of multi-omics data usage as well as a lack of heterogeneity in populations.With the strengths and limitations of previous works in mind, we have at present 4 broad objectives we would like to achieve. The first is to integrate phenotypic multi-omics data and histological features of hepatic damage to identify subpopulations of patients with specific disease drivers. Secondly, we wish to develop predictive disease models on an individual and sub-population basis to identify the factors that contribute to the longitudinal progression of NAFLD, through combining traditional supervised ML and DL classification methods with survival analysis. Thirdly we hope to identify clinical biomarkers that are patient specific that can reflect the grade of steatosis, steatohepatitis (NASH) and stage of fibrosis - through this we aim be able to diagnose and prognose individuals via routine blood tests rather than biopsy. Our final objective is to compare the use of topological data analysis and machine learning for the classification of digitized biopsy images to determine benign steatosis of the liver from more serious NASH, fibrosis, and cirrhosis.The project so far has utilised the European NAFLD registry, the largest NAFLD dataset in Europe which comprises of historic NAFLD cases as well as NAFLD cases acquired since 01/01/18 as part of the IMI-2 funded LITMUS project - an active study recruiting from >25 sites across 13 European countries. Patients within this registry are followed-up to 10 years from baseline to allow for longitudinal analysis of patients that are scattered across the NAFLD disease progression spectrum. The data collected includes an individual's clinical information, liver histopathology, biopsy samples amongst others. The use of multi-omics data from the registry is also applied to better understand inter-patients' variability of hepatic injury.