OR WAIT 15 SECS
(ACR Pediatrics 2014) Pattern-sorting analysis similar to Google ad technology has identified clusters of disease measures in records of children with rheumatic disease that appear more useful than current criteria sets.
Netflix sees patterns in people's movie preferences. Google matches people to their shopping habits. Now, the same kind of human and automated brainpower is serving a crucial clinical need: finding new patterns within the baffling heterogeneity of childhood rheumatic conditions.
Speaking at an American College of Rheumatology symposium in Florida, Rae Yeung MD of Toronto's Hospital for Sick Children described a sophisticated computer-based analysis that has taken a Netflix-like approach to the accepted variables in childhood rheumatic diseases, such as age at diagnosis, standard measures of disease activity, and molecular biomarkers. Aiming at a new and more objective way to categorize rheumatic conditions of childhood, the analysis has defined five unique clusters of variables that best describe the distribution of variables, revealing “quite distinct and ... homogeneous patient groups,” as she put it.
The new set of composite measures out-performs ILAR (International League Against Rheumatism) classification criteria, judging by the p-values of differences between the categories, she said. One particular cluster of variables, never before described, appears to represent a clinical entity that needs early treatment, but would be difficult to recognize with clinical signals alone.
The initial analysis involved 157 treatment-naÃ¯ve Canadian children with signs and symptoms of rheumatic disease, and the results were validated in a different cohort of 102 patients. Researchers used the Kruskal-Wallis analysis of variance, a nonparametric statistical method of determining the degree of difference between a set of samples.
In most cases the computer assessment defined as most critical the same elements that expert pediatric rheumatologists would have identified as important (ESR, CRP, CHAQ, VAS, several interleukins, and clinical parameters such as number of axial and effused joints as well as enthesitis). The principal components identified in the initial cohort "map beautifully” with samples in the validation cohort, Yeung said.
The exception was one cluster that showed a strong discordance between clinical variables, which had low values, and biological variables such as protoinflammatory cytokines, which had high values. This was also the only group in which, as Yeung put it, “the kids were getting worse over the trajectory of the disease, probably because we’re not seeing them clinically and not treating as aggressively as we should.”
Describing the work as “really at the beginning,” Yeung said that much more data needs to be analyzed. “Another problem is that there are things that may be important which we don’t yet know how to measure,” such as certain cytokines, remarked Fabrizio DeBenedetti MD of the Ospedale Pediatrico Bambino Gesu in Rome.