Improved propensity matching for heart failure using neural gas and self-organizing maps

We studied heart failure mortality and hospitalization of N = 7, 788 subjects in the Digitalis Intervention Group (DIG) clinical trial. Cases were defined as subjects with New York Heart Association (NYHA) class III - IV symptoms, while controls were defined as subjects with NYHA class I - II symptomatology. Controls were propensity matched with cases using logits from logistic regression, best winning nodes for neural gas and self-organizing maps, and k-means cluster analysis. Cox proportional hazards (PH) regression models were ran to determine the all-cause mortality and hospitalization hazard ratio (HR) for having NYHA functional class III - IV. Unmatched data resulted in a mortality HR of 1.28 (95% CI, 1.17 - 1.41), while logit-based propensity matching resulted in a mortality HR of 1.29 (95% CI, 1.15 - 1.44). When neural gas (NG) was used for propensity matching with normalized and standardized features, the mortality HR was 1.34 (95% CI, 1.19 - 1.50) and 1.05 (95% CI, 0.94 - 1.17), respectively. Propensity matching with self-organized maps (SOM) and normalized and standardized features yielded mortality HRs of 1.31 (95% CI, 1.16 - 1.46) and 1.05 (95% CI, 0.94 - 1.17), respectively. Crisp K-means cluster-based matching performed worse and biased the HRs towards the null value of HR = 1. The strongest influence of matching was observed for NG when normalized features were used.