Roche and IBM have developed an algorithm based on real-word data, including electronic health records, that predicts patients with early risk of diabetes-related chronic kidney disease.
The Roche-IBM predictive algorithm—trained using half a million electronic health records—outperformed published algorithms derived from clinical trial data in a one-to-one comparison, as well as in cohort studies.
Results of the research, published this week in the journal Nature Medicine, were based on a year-long collaboration between Roche (the study lead), IBM, Eli Lilly, the Regenstrief Institute and the Indiana Bioscience Research Institute.
“This study demonstrates the growing importance of real-world data and predictive analytics in diabetes care,” said Mark Davies, chief medical officer (Europe) for IBM Watson Health. “There is a growing need to improve screening performance and the decision-making processes in diabetes care, and this new data suggest that real-world data and analytics can be applied to help in early recognition of risk of CKD.”
Also See: Mount Sinai partners with startup to develop AI for kidney disease
The Indiana Bioscience Research Institute, Eli Lilly and Indiana University School of Medicine—through the Regenstrief Institute—provided Roche with another independent real-world data set originating from almost 100,000 patients with diabetes obtained from the Indiana Network of Patient Care database, which confirmed the findings.
“These extracted datasets constitute the largest RWD basis that has been used thus far to investigate CKD as a long-term complication of diabetes,” according to the study’s authors, who concluded that the “teaching of predictive analytics algorithms using RWD could achieve equivalent or even enhanced accuracy compared with those using clinical trial data.”
At the same time, researchers added that “further testing on additional datasets will be necessary before these conclusions can be generalized.”
Nonetheless, the authors speculated that “it is the diversity of the RWD that makes the prediction algorithm better suited for generalization.”