How to Get Research-Grade AI Accuracy… in 3 Minutes

Original article was published by Frederik Bussler on Artificial Intelligence on Medium

Unsurprisingly, blood glucose is the most important factor, as higher fasting blood glucose signifies the body’s inability to break down glucose, a sign of diabetes. BMI is the next most important, as overweight and obese individuals are at a greater risk of diabetes. Next up is age, as health tends to decay over time.

Several models are made in the background that statistically weight each attribute to make predictions. The most accurate model selected is a logistic regression classifier, with a cross-validated Jaccard Score of 0.768. Recall that the research paper achieved an accuracy of 0.763.

We can use AutoML to achieve similar results in a tiny fraction of the time.

At a glance, we can predict the likelihood of any patient having diabetes, and even segment patients according to their risk. For example, patients with blood glucose 152 or higher, a BMI over 32, and who are at least 33 have a 2.4x greater likelihood of having diabetes.

Deployed in the field, this model would enable the prioritization of testing and a greater understanding of the risk of diabetes in patients.


Even a few years ago, using AI was a difficult, time-intensive process. Today, AutoML tools make it easier than ever to deploy AI and find meaningful insights. All you need is a dataset with a KPI and attributes.

In another article, I made an AI that beats doctors in heart failure prediction, but AutoML can be used for any tabular data, so I’ve applied it to tasks like predicting 2020’s instability, predicting customer churn, predicting CO2 emissions, and so much more.