Robust machine learning for healthcare

人工智能在医疗保健领域的巨大前景需要时间才能实现。除了由于媒体炒作而导致的获取数据的困难和对人工智能不切实际的期望之外，要实现人工智能在医疗保健领域的广泛应用，还需要取得许多基本的机器学习进展。目前的一些缺点包括但不限于：1)预测罕见关键事件（如心脏骤停）的错误率高；Ii)对于由于政策和实践的变化而引起的数据的潜在变化，模型缺乏鲁棒性；Iii)缺乏模型的可解释性，这使得使用者（如临床医生）很难根据偏离直觉的预测采取行动。在我的实验室里，我们已经开始解决其中的一些基本问题。在这笔拨款中，我提出了一项研究计划，在未来5年内在解决这些问题方面取得实质性进展。许多关键事件（如心脏骤停）的发生频率通常低于5%。这些问题往往被错误地视为分类。我们工作的目标之一是扩展生成模型的类别，创建适合医疗保健中异常值检测目的的模型。处理数据的移位。患者群体、护理实践和数据库系统随着时间的推移而发展，但迄今为止，很少有工作明确地将数据转移和发展作为病历建模的一部分。我们发现有一种简单而有效的缓解策略：将原始特征聚合到专家定义的临床概念中。随着时间的推移，我们将开发新的鲁棒特征表示技术和自适应学习，以提高机器学习（ML）在未来医疗保健部署中的准备程度。Explainability。一旦开发出来，将机器学习模型有效地转化为实践需要建立用户的信任。我们调查了临床医生，了解他们对可解释模型的看法。下一步是将这些可解释性的感知集成到ML模型开发中。在这个项目中，我将探索反事实模型来实现这一目标。培训。我的工作涉及并影响了许多学科，从计算机科学到医学再到伦理学。在过去的八年里，我一直在培训来自不同背景的HQP，从计算机科学到生物学，从工程到经济学，以及医疗专业人员，使用和开发新的医疗保健人工智能模型。我的目标是监督和帮助这个新兴领域的发展，从而培养出一支技能非常高的劳动力队伍。的影响。大量可用的健康数据清楚地表明，机器学习在未来的医学中扮演着重要的角色。本基金提出的工作将有助于为负责任的人工智能+医疗奠定基础，确保人工智能在这一重要领域的潜力得以实现。

The great promise of AI in healthcare is taking time to materialize. Besides difficulties with access to the data and unrealistic expectations of the AI due to the hype fueled by the media, there are many fundamental machine learning advances that need to be made to achieve the widespread use of AI in healthcare. Some of the current shortcomings include but are not limited to 1) high error rates in predicting rare critical events (such as cardiac arrest); ii) lack of model robustness to the underlying changes in the data arising due to changes in the policy and practice; iii) lack of the model explainability that makes it hard for the users, e.g. clinicians, to act upon predictions that diverge from their intuition. In my lab, we have started addressing some of these fundamental issues. In this grant I am proposing a research program to make substantial improvements in addressing these questions over the next 5 years. Classification vs outlier detection The frequency of many critical events, such as cardiac arrest, is usually below 5%. These problems are often incorrectly treated as classification. One of the goals of our work will be to expand the class of generative models, creating models fit for the purpose of outlier detection in healthcare. Dealing with shifts in the data. Patient populations, care practices, and database systems evolve over time, and yet few works to date have explicitly addressed data shift and evolution as part of the medical record modeling. We showed that there is a simple yet effective mitigation strategy: aggregation of raw features into expert defined clinical concepts. We will develop new robust feature representation techniques and adaptive learning over time to make increase readiness of machine learning (ML) for healthcare deployment in the future. Explainability. Once developed, translating ML models effectively to practice requires establishing users' trust. We surveyed clinicians to understand their perception of interpretable models. The next step is to integrate these perceptions of explainability into the ML model development. In the program proposed in this grant I will explore counterfactual models to achieve this goal. Training. My work draws on and affects many disciplines, from computer science to medicine to ethics. Over the last eight years, I have been training HQP coming from a variety of backgrounds, from computer science to biology, from engineering to economics as well as medical professionals in using and developing new AI models for healthcare. It is my goal to supervise and help grow this new and burgeoning field to result in a very highly skilled workforce. Impact. The amount of available health data has made it clear that machine learning has a great role to play in the future of medicine. The work proposed in this grant will help to build the foundation for responsible AI+healthcare to ensure the realization of the AI potential in this important field.