Tree-based machine learning methods have gained traction in the statistical and data science fields. They have been shown to provide better solutions to various research questions than traditional analysis approaches. To encourage the uptake of tree-based methods in health research, we review the methodological fundamentals of three key tree-based machine learning methods: random forests, extreme gradient boosting and Bayesian additive regression trees. We further conduct a series of case studies to illustrate how these methods can be properly used to solve important health research problems in four domains: variable selection, estimation of causal effects, propensity score weighting and missing data. We exposit that the central idea of using ensemble tree methods for these research questions is accurate prediction via flexible modeling. We applied ensemble trees methods to select important predictors for the presence of postoperative respiratory complication among early stage lung cancer patients with resectable tumors. We then demonstrated how to use these methods to estimate the causal effects of popular surgical approaches on postoperative respiratory complications among lung cancer patients. Using the same data, we further implemented the methods to accurately estimate the inverse probability weights for a propensity score analysis of the comparative effectiveness of the surgical approaches. Finally, we demonstrated how random forests can be used to impute missing data using the Study of Women’s Health Across the Nation data set. To conclude, the tree-based methods are a flexible tool and should be properly used for health investigations.

Introduction: Statistical methods to assess the impact of an intervention are increasingly used in clinical research settings. However, a comprehensive review of the methods geared toward practitioners is not yet available. Methods and Materials: We provide a comprehensive review of three methods to assess the impact of an intervention: difference-in-differences (DID), segmented regression of interrupted time series (ITS), and interventional autoregressive integrated moving average (ARIMA). We also compare the methods, and provide illustration of their use through three important healthcare-related applications. Results: In the first example, the DID estimate of the difference in health insurance coverage rates between expanded states and unexpanded states in the post-Medicaid expansion period compared to the pre-expansion period was 5.93 (95% CI, 3.99 to 7.89) percentage points. In the second example, a comparative segmented regression of ITS analysis showed that the mean imaging order appropriateness score in the emergency department at a tertiary care hospital exceeded that of the inpatient setting with a level change difference of 0.63 (95% CI, 0.53 to 0.73) and a trend change difference of 0.02 (95% CI, 0.01 to 0.03) after the introduction of a clinical decision support tool. In the third example, the results from an interventional ARIMA analysis show that numbers of creatinine clearance tests decreased significantly within months of the start of eGFR reporting, with a magnitude of drop equal to − 0.93 (95% CI, − 1.22 to − 0.64) tests per 100,000 adults and a rate of drop equal to 0.97 (95% CI, 0.95 to 0.99) tests per 100,000 per adults per month. Discussion: When choosing the appropriate method to model the intervention effect, it is necessary to consider the structure of the data, the study design, availability of an appropriate comparison group, sample size requirements, whether other interventions occur during the study window, and patterns in the data.

Background: The Oncology Care Model (OCM) was developed as a payment model to encourage participating practices to provide better-quality care for cancer patients at a lower cost. The risk-adjustment model used in OCM is a Gamma generalized linear model (Gamma GLM) with log-link. The predicted value of expense for the episodes identified for our academic medical center (AMC), based on the model fitted to the national data, did not correlate well with our observed expense. This motivated us to fit the Gamma GLM to our AMC data and compare it with two other flexible modeling methods: Random Forest (RF) and Partially Linear Additive Quantile Regression (PLAQR). We also performed a simulation study to assess comparative performance of these methods and examined the impact of non-linearity and interaction effects, two understudied aspects in the field of cost prediction. Methods: The simulation was designed with an outcome of cost generated from four distributions: Gamma, Weibull, Log-normal with a heteroscedastic error term, and heavy-tailed. Simulation parameters both similar to and different from OCM data were considered. The performance metrics considered were the root mean square error (RMSE), mean absolute prediction error (MAPE), and cost accuracy (CA). Bootstrap resampling was utilized to estimate the operating characteristics of the performance metrics, which were described by boxplots. Results: RF attained the best performance with lowest RMSE, MAPE, and highest CA for most of the scenarios. When the models were misspecified, their performance was further differentiated. Model performance differed more for non-exponential than exponential outcome distributions. Conclusions: RF outperformed Gamma GLM and PLAQR in predicting overall and top decile costs. RF demonstrated improved prediction under various scenarios common in healthcare cost modeling. Additionally, RF did not require prespecification of outcome distribution, nonlinearity effect, or interaction terms. Therefore, RF appears to be the best tool to predict average cost. However, when the goal is to estimate extreme expenses, e.g., high cost episodes, the accuracy gained by RF versus its computational costs may need to be considered.

Background: Prior research suggests clinical effects of glucagon‐like peptide‐1 receptor agonists (GLP‐1RAs) and sodium‐glucose cotransporter‐2 inhibitors (SGLT2is) are mediated by changes in glycated hemoglobin, body weight, systolic blood pressure, hematocrit, and urine albumin‐creatinine ratio. We aimed to confirm these findings using a meta‐analytic approach. Methods and Results: We updated a systematic review of 9 GLP‐1RA and 13 SGLT2i trials and summarized longitudinal mediator data. We obtained hazard ratios (HRs) for cardiovascular, renal, and mortality outcomes. We performed linear mixed‐effects modeling of LogHRs versus changes in potential mediators and investigated differences in meta‐regression associations among drug classes using interaction terms. HRs generally became more protective with greater glycated hemoglobin reduction among GLP‐1RA trials, with average HR improvements of 20% to 30%, reaching statistical significance for major adverse cardiovascular events (ΔHR, 23%; P =0.02). Among SGLT2i trials, associations with HRs were not significant and differed from GLP1‐RA trials for major adverse cardiovascular events ( P _interaction =0.04). HRs for major adverse cardiovascular events, myocardial infarction, and stroke became less efficacious (ΔHR, −15% to −34%), with more weight loss for SGLT2i but not for GLP‐1RA trials (ΔHR, 4%−7%; P _interaction <0.05). Among 5 SGLT2i trials with available data, HRs for stroke became less efficacious with larger increases in hematocrit (ΔHR, 123%; P =0.09). No changes in HRs by systolic blood pressure (ΔHR, −11% to 9%) and urine albumin‐creatinine ratio (ΔHR, −1% to 4%) were found for any outcome. Conclusions: We confirmed increased efficacy findings for major adverse cardiovascular events with reduction in glycated hemoglobin for GLP1‐RAs. Further research is needed on the potential loss of cardiovascular benefits with increased weight loss and hematocrit for SGLT2i.