New Pan-Cancer Prognostic Score Outperforms Existing Tools

By Deborah Borfitz

March 15, 2022 | Roche has developed a pan-cancer prognostic scoring system useful for predicting patient survival in clinical trials and monitoring treatment progress over time. The score comprises 27 routinely collected clinical parameters and outperforms all other existing scores in every study phase, according to Anna Bauer-Mehren, data science lead for Roche Pharmaceutical Research and Early Development in Germany.

Bauer-Mehren, speaking at the recent Summit for Clinical Ops Executives (SCOPE), says ROPRO—short for Real-World Data Prognostic Score—was derived from overall survival data on cancer patients mined from the database platform of Flatiron Health (acquired by Roche in 2018). That baseline score is being used to inform patient selection for phase 1 oncology trials where 12-week life expectancy is routinely an inclusion criterion.  

The most widely used performance status scale was developed by the Eastern Cooperative Oncology Group in 1982, she says. Others have since been developed, including the Royal Marsden Index that made its debut over a decade ago.

ROPRO combines many more parameters into one score, using artificial intelligence methods to wrangle all the data during the training phase. Among the inputs are lifestyle factors such as smoking and body mass index, albumin level, a few vital signs, demographic information, and a mix of protective and risk factors.

The pan-cancer score trained on data from 120,000 patients across 17 cancer indications and has been validated in more than 20 studies and in the UK Biobank as being reflective of real-world outcomes, Bauer-Mehren says. Data on advanced non-small-cell lung cancer was most heavily represented. The same validation exercise will next be done in an Asian dataset.

More advanced deep learning does not improve ROPRO’s prediction of overall survival when applied to clinical trial data, says Bauer-Mehren. It’s a robust model that’s easy to apply either retrospectively or prospectively, without any additional overhead, because the 27 variables are already being routinely collected.

ROPRO scores can be computed at baseline to predict overall survival, or for “quantitative and unbiased selection” of patients for phase 1 trials, she continues. The 27 parameters can also be collected at different time points in a clinical trial, making “delta ROPRO” useful for detecting treatment benefits or disease progression early.

In phase 1 oncology trials, 17% of patients would drop out within 12 weeks, typically because they did not survive that long, says Bauer-Mehren. Patient deaths can be lowered to 15% using standard inclusion/exclusion criteria, and down to 11% using ROPRO.

Roche is now implementing ROPRO as an exclusion criterion in its phase 1 studies, but as an “add on” giving investigators the option to overrule ROPRO’s calculation of a patient’s fitness for a trial, she continues. No such overrides have yet to happen.

In a study on 10,000 patients looking retrospectively at complete responders relative to partial responders and those who were stable or whose disease had progressed, delta ROPRO was shown to be an early predictor of best overall response, notes Bauer-Mehren.

Among the group on the chemotherapy drug docetaxel, significant differences were seen across patient groups from the third visit on, she reports. The retrospective analysis was subsequently repeated on other trials, resulting in a paper that has been submitted for publication.

That ROPRO is a pan-cancer score is important, says Bauer-Mehren, since phase 1 oncology studies often enroll a mixture of patients with different indications. “We have trained indication-specific models (e.g., for lung cancer) and see a small performance increase usually.”

Load more comments
comment-avatar