By Brenda Cass, Data Science Strategist, Validic
When we use our data to gain a good understanding of our present, we inevitably start to wonder what it could tell us about the future. It makes sense, then, that machine learning (ML) has an entire subdomain dedicated to predicting what will happen next. Regression is a process that allows us to model the relationship between contributing variables and a dependent variable in order to figure out what its next value might be – and healthcare is full of examples where regression has been used to predict health outcomes. The Trauma Injury Severity Score (TRISS), for example, uses a logistic regression model to predict the mortality of an injured patient based off of their revised trauma score (RTS), injury severity score (ISS), and the patient’s age. And, today, healthcare organizations are using machine learning to predict healthcare costs for patients, length of hospital stays, and how nutrition will affect glycemic responses.
Leveraging regression models to make health predictions
Many of us were taught about linear or polynomial regression in school; spreadsheets and graphing calculators can perform these tasks easily. Unfortunately, these types of regression models struggle to successfully model complex, non-linear relationships. We still have more options, though, building upon these models to derive real value from data.
Validic uses a specialized neural network regression model to predict discrete blood glucose measurements from a person’s historical values. This particular model uses only blood glucose values, some statistics derived from those values, and the time at which the measurement was taken to make a prediction within a 12 mg/dl margin of error – the same margin of error as the blood glucometers themselves. That’s pretty exciting, given that a value predicted sufficiently in advance can give care providers a chance to intervene when someone might run into trouble with their blood glucose.
Neural networks perform very well when fed lots of data, something Validic happens to have in spades. Unfortunately it’s hard to explain exactly how a neural network is coming to its conclusions, which makes some people skeptical. We can see the weights assigned to nodes, but how exactly to translate that into human readable patterns is something that data scientists are still actively working to do effectively. However, we are not left completely in the dark because of a little trick called empirical observation.
Over time we are able to compare predictions to an individual’s actual measurements to assess how accurate it is. Individual predictions that are consistently accurate will build confidence while predictions that are consistently inaccurate tell us that there might be something anomalous going on with that person — both useful outcomes!
A group of people with consistently poor predictions, for example, indicates that the model may need additional training with data that better represents the population. The important part is allowing people to observe the model’s results over time alongside their typical workflow, and let them decide when they are ready to use the insights it surfaces.
Personalizing interventions, improving outcomes
The information derived from this metric predictor can be used enable proactive interventions before a user logs an out-of-range blood glucose reading, helping to prevent adverse outcomes resulting from diabetes. And, this technique can be applied to a variety of data types, improving care for individuals managing a variety of chronic diseases or other health issues.
Data science is most impactful when it works with people, helping them get to more of what matters, and regression models can definitely help us prioritize and be proactive about our work – if we let them.
See you next time where I will be blogging about how Validic is predicting when a user is going to stop using their activity tracker.
View More >