Exercise 1: Choosing a clinical problem and framing it as a machine learning task.

In this exercise, you will practice how to choose a clinical problem to tackle, explore how to research some basic facts about the disease, and frame the clinical problem as a machine learning task.

Part 1: Choosing a clinical case.

Exciting clinical problems abound in the realm of Healthcare AI. To help you narrow down some potential ideas, we will use the American College of Radiology Data Science Institute (ACR DSI) curated list of use cases. Alternatively, you may explore any clinical problem of your choosing without using the database if you already had something in mind. Please note that some use cases in the ACR DSI database can seem overwhelming or obscure without some background medical knowledge. If this is the case, simply choose something that’s more intuitive. A few good starting examples include:

• Acute Appendicitis.
• Aging Brain – Dementia.
• Incidental Pulmonary Nodules on CT.
  Your task for this first part is to choose a case. If you choose to use the ACR DSI database, submit the link that points to your use case. If using any other resources, please submit an image capture or PDF of a medical or AI journal article that explores this problem within the context of machine learning.

Part 2: Background research.

To deliver an algorithm that performs well in a clinical setting, it can be extremely beneficial to have some rudimentary understanding of the disease or situation you are addressing. To do so, engaging in some background research to build your foundation will be crucial. For this part of the exercise, find a review article from the PubMed database describing the problem. Ask yourself, “what is the current gold standard for confirming or ruling out the presence of this disease/state”. Understanding the current gold standard/ground truth test will be a critical benchmark when testing any algorithm you develop. Submit a screen capture or PDF of the article to receive credit for this portion.

Part 3: Framing the problem as a machine learning task.

With some background knowledge of the disease or condition at hand, as well as how it is currently ruled in or out, consider how the task could be framed as a machine learning problem. We will be going over this in much more detail in the respective lesson of this course, but some early exposure will be useful. For example, solving the problem of autonomously detecting one or more lung nodules could be framed as an object detection task. Measuring changes in tumor volume over time may be best framed as a segmentation task. Etc. Submit a short paragraph detailing how you would frame your clinical problem of choice as a machine learning task, and explain your rationale.

Below you will find a workspace with a Jupyter Notebook where you can record the answers to each part of this exercise listed above.