Machine learning hints at complex treatment effects in Look AHEAD
medwireNews: A machine-learning analysis suggests the existence of heterogeneous treatment effects in the Look AHEAD trial, which could account for the overall neutral effect of the study intervention on cardiovascular outcomes.
Aaron Baum (Icahn School of Medicine at Mount Sinai, New York, USA) and colleagues used machine learning, rather than traditional subgroup analysis, because it requires no initial hypotheses and can assess the effect of multiple combinations of variables. They note that it also avoids some statistical pitfalls, such as multiple testing errors.
As described in The Lancet Diabetes & Endocrinology, the analysis used data from 2450 Look AHEAD participants (patients with type 2 diabetes and overweight/obesity) for model development and the data from the remaining 2451 to test the models.
This identified two key subgroups, both with low glycated hemoglobin (HbA1c <6.8%) at baseline. One subgroup, representing 16% of all patients, also had poor general health, with a Short Form Health Survey-36 score below 48. These patients did not benefit from being in the intervention group (weight loss through healthy eating and increased physical activity), in fact having an absolute risk increase of 7.41% for the primary cardiovascular outcome relative to receiving usual care.
Writing in an accompanying editorial, Edward Gregg (Centers for Disease Control and Prevention, Atlanta, Georgia, USA) and Rena Wing (Alpert Medical School of Brown University, Providence, Rhode Island, USA) describe this as “intuitively puzzling” and ask if it could be a chance finding.
By contrast, the other subgroup, representing 24% of the patients, had better general health, and those assigned to the weight loss intervention had an 8.22% absolute reduction in the risk for the primary outcome. Gregg and Wing say that this at least is consistent with previous findings from the trial, with low HbA1c being associated with a greater chance of partial remission from diabetes and with relatively low healthcare utilization.
More broadly, patients not in the first subgroup (ie, those with higher HbA1c or with low HbA1c but good health status) had an absolute risk reduction of 3.46% and a number needed to treat of 28.9 to prevent one primary outcome event over 9.6 years.
Overall, the commentators say the findings “are difficult to interpret,” but stress that “novel and rigorous study of the heterogeneity in intervention response is important to guide personalised care, risk stratification, and translation of prevention studies to clinical and community settings.”
They say: “As we enter an era characterised by increasing lifespans but persistent high prevalence of chronic disease and multi-morbidity, more diverse menus of lifestyle interventions will be necessary to improve population health.”
medwireNews is an independent medical news service provided by Springer Healthcare. © 2017 Springer Healthcare part of the Springer Nature group