Can a simple urine test detect one of the deadliest cancers?

About Dataset

This is a brand-new (!) dataset from an open-access paper published December 10, 2020. The paper and the full dataset are open-access (CC-BY), so please give attribution to the original authors in your work.

Background

Pancreatic cancer is an extremely deadly type of cancer. Once diagnosed, the five-year survival rate is less than 10%. However, if pancreatic cancer is caught early, the odds of surviving are much better. Unfortunately, many cases of pancreatic cancer show no symptoms until the cancer has spread throughout the body. A diagnostic test to identify people with pancreatic cancer could be enormously helpful.

The paper

In a paper by Silvana Debernardi and colleagues, published this year in the journal PLOS Medicine, a multi-national team of researchers sought to develop an accurate diagnostic test for the most common type of pancreatic cancer, called pancreatic ductal adenocarcinoma or PDAC. They gathered a series of biomarkers from the urine of three groups of patients:

Healthy controls

Patients with non-cancerous pancreatic conditions, like chronic pancreatitis Patients with pancreatic ductal adenocarcinoma When possible, these patients were age- and sex-matched. The goal was to develop an accurate way to identify patients with pancreatic cancer.

The data

The key features are four urinary biomarkers: creatinine, LYVE1, REG1B, and TFF1.

Creatinine is a protein that is often used as an indicator of kidney function. YVLE1 is lymphatic vessel endothelial hyaluronan receptor 1, a protein that may play a role in tumor metastasis REG1B is a protein that may be associated with pancreas regeneration TFF1 is trefoil factor 1, which may be related to regeneration and repair of the urinary tract Age and sex, both included in the dataset, may also play a role in who gets pancreatic cancer. The dataset includes a few other biomarkers as well, but these were not measured in all patients (they were collected partly to measure how various blood biomarkers compared to urine biomarkers).

I have not changed any of the data from the paper, other than renaming the columns for easy importing and use. The file Debernardi et al 2020 data.csv contains the raw data, while the file Debernardi et al 2020 documentation.csv contains a detailed documentation of what each column represents (as well as the original column names from the paper).

Prediction task

The goal in this dataset is predicting diagnosis, and more specifically, differentiating between 3 (pancreatic cancer) versus 2 (non-cancerous pancreas condition) and 1 (healthy). The dataset includes information on stage of pancreatic cancer, and diagnosis for non-cancerous patients, but remember—these won't be available to a predictive model. The goal, after all, is to predict the presence of disease before it's diagnosed, not after!

Acknowledgements

I would like to thank the authors of this paper, for graciously sharing their raw data with the research community.