Estimate variable importance for recurrent event outcomes with an application to identify hypoglycemia risk factors

Abstract

Recurrent event data are an important data type for medical research. In particular, many safety endpoints are recurrent outcomes, such as hypoglycemic events. For such a situation, it is important to identify the factors causing these events and rank these factors by their importance. Traditional model selection methods are not able to provide variable importance in this context. Methods that are able to evaluate the variable importance, such as gradient boosting and random forest algorithms, cannot directly be applied to recurrent events data. In this paper, we propose a two-step method that enables us to evaluate the variable importance for recurrent events data. We evaluated the performance of our proposed method by simulations and applied it to a data set from a diabetes study.

Baseline factors associated with better response to insulin lispro low mixture or insulin glargine: A post hoc analysis of the DURABLE study

“…It is important to note that our objective was not to merely quantify the differences in effect between these insulin treatments but in large part, was to test if the innovative statistical method used in this analysis could identify variables that would help guide or direct the clinician toward a treatment choice that may be more suitable for a patient. The gradient-boosting method has been demonstrated to be more accurate in selecting the most influential covariates than parametric and other nonparametric models [12][13][14][15]. Prior analyses [18][19][20][21][22] primarily described the direction of a baseline characteristic in association with an outcome when holding other characteristics constant; whereas, our study identified baseline characteristics combinations by analyzing combined baseline characteristics subpopulations.…”

“…More evidence is needed to assist clinicians with tailoring diabetes therapy to meet the specific goals for their patients [1]. The use of new statistical techniques that allow for more accurate identification of patient variables associated with the greatest influence on treatment outcomes when compared with methods commonly used in clinical research [12][13][14][15] may improve understanding of individualized response to diabetes treatments to help identify patients who respond best to one treatment over another. With this aim, we performed a post hoc analysis using an innovative statistical approach to assess data from more than 2000 patients from the DURAbility of Basal versus Lispro Mix 75/25 Insulin Efficacy (DURABLE) study [16,17] to evaluate baseline characteristics associated with better efficacy and safety outcomes in insulin-naïve patients with T2D treated with either twice-daily insulin lispro low mixture 75/25 (LM) or once-daily insulin glargine (IG).…”

Risk Factors for Nocturnal Hypoglycemia in Insulin-treated Patients With Type 2 Diabetes: A Secondary Analysis of Observational Data Derived From an Integrated Clinical Trial Database

“…As a nonparametric regression, gradient boosting identified the variables that were most influential in predicting the outcome (NH). 16,17 Second, those variables, as well as several additional variables that were judged to be clinically relevant to NH based on the literature, were included in a Cox proportional hazards model, with a random effect approach. The results were reported as hazard ratios (HRs) with 95% CIs.…”

A Two-Step Algorithm to Estimate Variable Importance for Multi-State Data: An Application to COVID-19

“…Penalized methods neglect the interaction between variables and their complex relationships or their unknown functional form. Most importantly, these techniques were unable to provide a quantitative assessment of their significance [6]. While some of the features measured in a study may be associated with the sojourn times in the entry states of each transition, they may have different degrees of importance, and some of them may be more important for some transitions than others [7].…”

“…They found that there is no single model selection criterion with uniformly superior performance, and they proposed a two-stage approach for model selection with promising performance in variable selection. Recently, Duan et al [6] proposed a machine learningbased approach to estimate variable importance using martingale and deviance residuals and their standardized counterparts which resulted in promising performances based on simulation studies. In the present study, we proposed a two-step algorithm to estimate the importance of variables for multistate data based on three different machine learning approaches: random forest, gradient boosting, and neural network as the most widely used method.…”