The chemotherapy docetaxel is widely accepted as a standard therapy for metastatic castration–resistant prostate cancer. But 10–20 percent of patients will have adverse side effects that force discontinuation of treatment. These patients may have been better off with another treatment in the first place, but who’s to know before trying the drug which patients will go on to experience debilitating side effects?

A crowdsourced competition asked this as an open question. In the Journal of Clinical Oncology Clinical Cancer Informatics, competition organizers and participating teams reported their findings: Using open data from four previously conducted clinical trials, teams of international researchers designed mathematical models predicting the likelihood that a patient will discontinue docetaxel treatment due to adverse events.

Specifically, the challenge was to connect any of 129 baseline clinical measurements to the chance of docetaxel discontinuation. In all, 34 international teams submitted 61 models. Seven of these teams submitted models with similarly high predictive ability and so technically “won” the challenge. The five clinical factors that were most predictive were measures of hemoglobin, alkaline phosphatase, aspartate aminotransferase, prostate specific antigen, and ECOG performance status. The seven successful models all integrated these five factors into various computational frameworks.

Interestingly, after the competition officially ended, these top seven teams decided to collaborate outside the framework of the competition, resulting in refinements that led to a combined model that was more predictive than any of the submissions alone.

The combined model stratified patients into groups with low and high risk of discontinuing docetaxel due to adverse events, with the high group having more than double the likelihood of discontinuation as the low group.

“Not only could a model like this help identify patients who might benefit more from a different treatment, it also has the potential to immediately impact future clinical trials by improving patient selection through the use of novel patient selection designs,” says Devin Koestler, PhD, assistant professor of Biostatistics at the University Kansas Medical Center, and one of the first authors of the paper.

The project was overseen as a collaborative effort between 16 institutions, led by academic research institutions including CU Cancer Center, open–data initiatives including Project Data Sphere, Sage Bionetworks, and the DREAM Challenges, and industry and research partners including Sanofi, AstraZeneca, and the Prostate Cancer Foundation.

Because only 10–20 percent of patients discontinue treatment due to adverse events, no single trial has enrolled enough patients to predict with statistical significance who would discontinue docetaxel – commonly, these trials tested the effectiveness of treatments in the population that was able to finish the regimen and were not designed to answer this secondary question of who would be unable to finish. Had clinical trial results remained firewalled by the academic or industry sponsors, this secondary question would have remained unanswered; however, the decision to open these clinical trial data allowed the current researchers to combine the numbers from four previous trials, pooling over 2,000 patients – enough to start identifying statistically significant patterns.