We are looking at prediction of immune checkpoint blockade outcomes in a few different advanced cancer types – non-small cell lung cancer, melanoma and renal cell carcinoma. We’re using real-world data from Moffitt’s electronic health records to predict outcomes, hopefully in the future for new patients about to receive immune checkpoint blockade.

What was the study design?

We used machine learning techniques alongside the real-world data. Moffitt partners with a company called Flatiron that does abstraction of our clinical data from the electronic health record. So they take in our data, they create an organised tabular format data, cleaned-up data from the EHR, and then we take that information and we feed it into machine learning models that are specifically trained for survival analysis.

So instead of classical classification or regression models we’re actually training for time to an event. So we try a few different models and compare which ones are best suited for our data and that goes through a traditional machine-learning pipeline with hyperparameter tuning, cross validation, testing and training and then we analyse the performance and the different features involved in what makes the model work.

What were the key findings?

First and foremost we found that the survival models outperformed the classical biomarkers. So usually the FDA-approved biomarkers for immune checkpoint blockade are tumour mutational burden, TMB, or PD-L1 which is done with a tumour proportion score with immunohistochemistry. Our models are substantially outperforming those. Those, I don’t know how widely they’re being used in the clinic because they’ve been shown to not work very well but it turns out with just a few important clinical features and blood biomarkers you can get a much better prediction of progression free survival and overall survival.

We also took a look into the features that drive the predictions and found some pretty important ones, namely the history of therapy that the patient has been on, the albumin within their blood, their cognitive performance on the ECOG test and a few others.

Then lastly we also took a look at whether we can do better prediction if we have early on-treatment data. So instead of just trying to predict before the start of therapy whether or not a therapy will succeed, it turns out we can make better predictions if we continue collecting that information because they’re getting blood biotests every 2-3 weeks anyway. So when we update our predictions with that information we can get even better prognostication.

What is the importance of these results?

At the moment our model is really good at being able to prognosticate for an individual patient. We have an expected progression free survival on this treatment, maybe 50% chance up to this month just based on their clinical information. But, to me, the real clinical utility will come when we can compare these prognostic results potentially with other types of therapy. So at the moment maybe half of patients do not get a good response on immune checkpoint blockade. For the other half it works great but the half that it doesn’t work for, maybe they’re better suited to some other types of treatment. So if we can build multiple models looking at each different kind of potential first-line therapy then we can compare outcomes – ‘Hey, they’re probably better off with this therapy versus the other.’