New machine learning model predicts pancreatic cancer risk in diabetic patients

Machine learning models can predict patients with new-onset diabetes who are at greater risk for developing pancreatic cancer, a study has shown.

A team of investigators analysed a retrospective cohort of patients with new-onset diabetes from multiple healthcare networks in the US. They designed an XGBoost machine learning model from a portion of this cohort (training set) and tested it on the rest of the participants (test set).

The features of the XGBoost model were explained using Shapley values, while its performance was compared with two contemporary models designed to predict pancreatic cancer among patients with new-onset diabetes.

The XGBoost models had an area under the curve of 0.80 (95 percent confidence interval, 0.76‒0.85) compared with 0.63 and 0.68 for the other two models in the test set.

With cutoffs based on the Youden index, the XGBoost model showed a sensitivity of 75 percent, a specificity of 70 percent, an accuracy of 70 percent, a positive predictive value (PPV) of 1.2 percent, and a negative predictive value of >99 percent. It also obtained a PPV of at least 2.5 percent, with a sensitivity of 38 percent.

“The XGBoost model was the only model that detected at least 50 percent of patients with cancer 1 year after the onset of diabetes,” said the investigators.

All three models exhibited similar features that predicted pancreatic cancer. These included older age, weight loss, and the rapid destabilization of glucose homeostasis.