Does insulin therapy affect all-cause mortality? machine learning complements propensity score analysis in a pharmacoepidemiologic study of adult diabetic females in Barranquilla, Colombia

Abstract

Aims: To investigate all-cause mortality (ACM) attributable to insulin treated diabetes mellitus through propensity score (PS)-weighting with and without novel confounders identified by Random Survival Forest (a machine learning approach). Methods: Prospective clinic encounter data was obtained from 1517 females with Type 2 diabetes (mean age 63±12 years) from Barranquilla, Colombia (2003 – 2016, censored August 2017) for a median 10-year mortality follow-up. Risk variables of importance for ACM were identified on RSF screening. Survival was compared in retrospective cohorts, identified by baseline treatment with glucose-lowering therapy, and balanced for confounders through PS-weighting with and without RSF variables using multivariable Cox regression. Results: RSF screening identified new risk variables (e.g., recruitment year, parity, reproductive lifespan) for ACM in women receiving insulin. The unweighted risk estimate showed a nonsignificant increased risk for ACM [HR 1.32 (.9, 2), p=0.2] compared to noninsulin treated women. After balancing for risk covariates in the compared cohorts, PS showed no significant effect of insulin on all-cause mortality [HR 95% CI 0.83 (0.5, 1.4) p=0.5] whereas PS-weighted analyses incorporating RSF novel variables approached conservative ACM estimates [HR 95% CI 0.56 (0.3, 1.0) p=0.07)]. The estimated ACM risk from active smoking was also more conservative with RSF weighting. Conclusion: In this observational study, insulin treatment appeared to be a surrogate for higher-risk women with diabetes mellitus. RSF-augmented PS analysis showed that insulin treatment may potentially be associated with a survival advantage compared to non-insulin treatment in older female diabetics.