Predictive Modelling: How Accurate is Accurate Enough?
What is predictive modelling?
You’ve had a long, stressful day at work. You get in your car to drive home. All of a sudden, you feel tight in your chest. You’re unsure whether you need to see a doctor immediately. So you consult your phone, which uses predictive analytics to tell you whether you need to see a doctor. Your phone tells you to go to the hospital. You do so immediately. Crisis averted.
Wouldn’t this make life so much easier? Your health no longer has to be a matter of guesswork. You can play an active role in your own wellbeing by using predictive modelling.
So, what is predictive modelling? Predictive modelling uses past data to learn patterns and trends, and predict future outcomes.
Predictive analytics and diabetes
Many Type 1 Diabetics use continuous glucose monitoring (CGM) devices, which measure glucose levels under your skin. The CGM alerts the user of their levels in real time so that they can make adjustments. However, CGMs are only able to tell a user what their past levels were. They also have inherent delays.
A team of researchers at Penn State noticed this and developed a mathematical model. It predicts the blood glucose levels of Type 1 Diabetics 30 minutes in advance, with a 90% accuracy. It accounts for fluctuations in blood glucose levels, resulting from exercise, stress, and eating. The researchers tested the accuracy of this model and found that the fluctuation of blood glucose levels after meal intake differ for and among patients. Thus, this model personalizes patient care and allows patients to know their future levels.
Similarly, researchers at Ohio University developed a software program that uses current and past patient data to learn and make future predictions. The program gets smarter with time and is able to run hypothetical situations. It predicts what would happen if you ate a chocolate bar or went on a 10-mile run.
However, this software requires you to constantly log in your data.
Inspired by both of these technologies, we developed Docto, an app that uses machine learning to predict blood glucose levels one hour in advance. It does this by learning trends in users’ levels and predicting future outcomes. The best part is, you don’t need to constantly log in your information. You just need to pick tags at the very beginning, so Docto can get to know you better. After that, Docto will be able to predict your future blood glucose levels.
Predictive modelling: How accurate is accurate enough?
So, how accurate is accurate enough? Given that it is difficult to achieve 100% accuracy, how do we know we can trust apps like Docto?
As we mentioned earlier, Docto uses machine learning to make future predictions. Docto goes through a learning phase where it learns as much as it can about a user. It’s similar to when you visit your new doctor for the first time, and they ask you questions to learn your medical history.
After the learning phase, Docto makes predictions about your future blood glucose levels, based on your past data. Predictive analytics (machine learning), improves the certainty of a prediction.
There are other factors that increase accuracy for predictive models, such as sample size and optimization. The more that you use Docto, the more it is able to learn and make more accurate predictions.