Unfortunately, research studies cannot address every human patient on the planet, and machine learning models are no exception. In healthcare, patient populations are what make groups of patients, and therefore, their data and disease characteristics—homogeneous. Examples of patient populations include inpatients, outpatients, emergency room patients, children, adults, and US citizens. Geographically, populations can even be defined at the state, municipal, or local levels.

What would happen if you tried to do modeling across different populations? Data from separate populations hardly ever overlap. First of all, it may be difficult to collect the same set of features across various populations. Some of the data may simply not be collected for certain populations. If you were trying to combine inpatient and outpatient populations, for example, you wouldn't get hourly blood pressure readings or intake/output measurements for your outpatients. In addition, another contributing problem is that data for different populations will most likely come from different sources, and you probably already know that the chance of two different data sources sharing many common features is low. How can you build a model based on patients who don't have the same features? Even if there is a shared lab test, for example, variations in how the lab quantity is measured and the units in which it is expressed make it nearly impossible to produce a homogeneous, coherent dataset.