Finding combinatorial drug repositioning therapy for Alzheimer's disease and related dementias

摘要 人工智能和机器学习(ML)模型在 临床应用。如果我们允许这些“自主的”ML模型为 对于临床决策，重要的是确保它们确实引入了算法不公平(例如， 不同人群的疾病负担或治疗机会的差异)。我们 提出新的技术解决方案来缓解算法不公平。我们将解决两个问题 主要类型的数据偏差(子组和表示)，以减少其对ML的负面影响 模特们。基于上下文信息和新的因果推理技术，我们将确定 潜在的异常值和与任务无关的混杂因素，并通过定制的缓解措施解决这些问题 避免学习错误信息的策略(例如，下采样和因子减少) 这可能会导致健康差距。此外，我们还将提出FairAUC(一种新的优化方案 机制)以最大限度地提高预测精度，同时通过设计考虑公平性。与之相反 对于事后的公平纠正方法，我们的方法将自动考虑两者 培训阶段的目标是在准确性和公正性之间取得最佳平衡。

Summary Artificial intelligence and machine learning (ML) models are becoming increasingly popular in clinical applications. If we allow these “autonomous” ML models to make recommendations for clinical decisions, it is important to ensure that they do introduce algorithmic unfairness (e.g., differences in the burden of disease or opportunities of treatment for different populations). We propose novel technological solutions to mitigate algorithmic unfairness. We will address two major types of data biases (subgroup and representation) to reduce their negative impact on ML models. Based on contextual information and novel causal inference techniques, we will identify potential outliers and task-irrelevant confounders and address them with customized mitigation strategies (e.g., down-sampling and factor reduction) to avoid learning erroneous information that might lead to health disparities. In addition, we will propose FairAUC (a new optimization mechanism) to maximize prediction accuracy while considering fairness by design. As opposed to post-hoc fairness rectification approaches, our method will automatically consider both objectives in the training phase to strike the optimal balance between accuracy and fairness.