A machine learning framework for trustworthy bio-medical risk factor identification – robust, explainable, and human-centred detection of endo- and phenotypes in lung cancer

肺癌是世界范围内死亡的主要原因，非小细胞肺癌（NSCLC）占所有病例的85%，总体5年生存率为17.8%。NSCLC的早期发现和靶向患者特异性治疗对于提高患者生存率至关重要。随着医疗数据量的不断增长，可以通过识别生物标志物来改善肺癌的诊断和治疗，这些生物标志物可用于识别患病风险增加或需要不同类型治疗的患者。最近，越来越多的机器学习方法被开发出来，以促进识别这些风险因素。然而，像蛋白质组学和电子健康记录这样的数据非常具有挑战性，因为这些数据是高维的、嘈杂的，并且包含各种数据偏差来源。此外，它需要对临床和生物学领域有深刻的理解才能正确地解释结果。在这个项目中，我们的目标是开发一个ML管道来识别非小细胞肺癌中值得信赖的医疗风险因素和生物标志物。考虑到数据的性质和领域内的复杂性，我们提出了一种健壮的、可解释的、以人为中心的方法。一方面，我们希望通过使使用的ML分析更加鲁棒来减少噪声和已知数据偏差的影响。另一方面，我们希望通过增加任何给定分析的可解释性，并通过为临床专家提供包含其领域知识的工具，来简化对结果的评估。我们相信，在这个项目中开发的技术可以很容易地应用于其他疾病领域，并将加速个性化医疗的发展。

Lung cancer is a leading cause of death worldwide, with non-small cell lung cancer (NSCLC) accounting for up to 85% of all cases and overall a 5-year survival rate of 17.8%. Both early detection and targeted patient-specific therapies of NSCLC are crucial to improve patient survival. As the amount of healthcare data is continuously growing, the diagnosis and treatment of lung cancer can be improved by identifying biomarkers which can be used to identify patients with an increased risk to develop the disease or which require a different type of therapy. Recently, an increasing number of machine learning approaches have been developed to facilitate the identification of such risk-factors.However, data like proteomics and electronic health records are very challenging to work with as the data is high dimensional, noisy, and contain various sources of data bias.Additionally, it requires a deep understanding of the clinical and biological domain to interpret results correctly.In this project we aim to develop a ML pipeline to identify trustworthy medical risk factors and biomarkers in NSCLC. Given the nature of the data and the complexity within the domain, we propose a robust, explainable, and human-centred approach. On the one side, we want to reduce the impact of noise and known data bias, by making the used ML analysis more robust. On the other side, we want to simplify the evaluation of results for clinical experts by increasing the explainability of any given analysis and by providing them with tools to include their domain knowledge.We believe that techniques developed within this project can easily be applied to other disease areas and will accelerate the development of personalised medicine.