The data used for analysis were from the United States Department of Veterans Affairs (VA) centers (compassing VA hospitals and clinics) stored in the Corporate Data Warehouse. A total of 3675 veterans from Corporate Data Warehouse were identified as patients with mCRPC, excluding those with missing values and including only the African American and White populations. We excluded patients with mCRPC who received docetaxel, cabazitaxel, mitoxantrone, or sipuleucel-T before or after starting abiraterone or enzalutamide. This exclusion resulted in a dataset of 2890 veterans who began treatment with abiraterone or enzalutamide between 2014 and 2017. In addition, 4 patients were excluded because they switched to different second-line treatments on the same day. Thus, the final analysis included 2886 patients. The study follow-up concluded in 2020.

We used causal survival forests with 20 covariates to estimate heterogeneous treatment effects in follow-up days [19]. Causal forests and causal survival forests are similar to random forests and random survival forests, respectively, with the primary difference being their data splitting criteria [20]. In a random survival forest, data splitting minimizes differences in prediction errors in follow-up days within each group, considering censoring. In contrast, a causal survival forest splits data to maximize heterogeneity, that is, the difference in the estimated follow-up days between the treated and untreated groups, accounting for censoring. The advantage of using a causal survival forest lies in its robustness to censoring, unlike a causal forest [19]. In addition, compared with a random survival forest, a causal survival forest offers a more comprehensive assessment of heterogeneity. It can outperform linear regression models by measuring heterogeneous treatment effects conditional on a nonlinear function of many covariates, accounting for higher-order interactions [21].

Each observation in the dataset corresponds to a single patient with mCRPC and includes the patient’s covariate data, follow-up days (the outcome), treatment assignment (1 for enzalutamide, 0 for abiraterone), and death (event). Survival duration was analyzed based on either abiraterone or enzalutamide treatment without an untreated control group. Consequently, the dataset does not allow for the determination of the individual treatment effects of each drug.

As this is a retrospective observational study, we must account for the nonrandomization of treatment assignment and use the augmented inverse-propensity weighting (AIPW) estimator, a doubly robust method. To do this, the true treatment assignment was fitted as a function of the observed covariates. The predicted value from this model provides a propensity score, that is, an estimate of the probability of treatment assignment conditioned on a set of covariates for each patient. Then, 2 models that estimate the outcome (follow-up) were fitted, one using enzalutamide and the other using abiraterone. Each outcome was then weighted by the estimated propensity score, which yields the weighted average of the two outcome models [22].

We used a causal survival forest for our analysis [19]. To ensure that the assumptions of finite horizon, ignorability, overlap, ignorable censoring, and positivity are satisfied so that the treatment effect of the drugs is identified when using causal survival forests, we made the following analytical choices. The parameter horizon, referring to restricted mean survival time, was set to a threshold so that the estimated censoring probabilities are not below 0.2, satisfying the finite horizon and positivity assumption. We started with a horizon of 2000 days (approximately close to the maximum follow-up days) and decremented by 100 to 1000.

For ignorability, we applied balanced diagnostics by checking the weighted absolute standardized mean difference (ASMD) of variables between patients treated with enzalutamide and patients treated with abiraterone. A number close to 0 indicates that the propensity scores are well-calibrated. We examined the propensity score distribution for both groups to test for overlap and checked if they clustered at 0 or 1.

The honesty fraction was set to 0.7, meaning 70% of the subsample was used for splitting and 30% for populating the leaf nodes. However, for ranking the observations based on covariates and coming up with subgroups, the model must not be fit using the observations being compared. We used 10-fold cross-fitting for this reason. The conditional average treatment effects (CATE) models were fit for 9 folds, and the “unseen” observations in the left-out fold were ranked based on their predictions and split along the median into 2 groups. The same process was repeated 10 times, with each fold serving as a left-out fold once.

We set the number of trees to 15,000. We generated the policy tree after calculating the AIPW scores. A policy tree is a set of treatment rules, that is, rule-based policies, in the form of a decision tree that physicians can use to prescribe abiraterone and enzalutamide. The policy tree was generated using 70% of the training data, and the policy’s value was determined using the 30% test data. The policy value is defined as the average difference in follow-up days obtained if the patients with mCRPC are administered enzalutamide or abiraterone. R (version 4.4.1; The R Foundation) with packages grf and policytree were used for analyses.

In this study, we applied a causal survival forest to assess the heterogeneous treatment effects of the mCRPC medications enzalutamide and abiraterone on overall survival. We found that, on average, for 2886 veterans, 59.94 (95% CI 35.60-84.28) more days of survival than abiraterone. The value of our policy is 56.38 (95% CI 8.89-103.87), which is not statistically different from 59.94. This means that the simple, transparent treatment rules using the policy tree method to prescribe abiraterone and enzalutamide can achieve a policy value similar to the conditional average treatment effect. Schoen et al [10] used an adjusted hazard model and found that age, BMI, and PSA level were positively associated with mortality. Our policy tree elicited the decision checks for age, PSA level, and BMI to make prescription decisions on which of the 2 medications is likely to prolong survival.

Interestingly, even though in univariate analyses in previous studies, the median treatment duration was longer in patients with hypertension or diabetes, the policy tree did not choose both hypertension and diabetes and did not contribute to the highest variation between the 2 subgroups. In previous work, increased hemoglobin A_1c, a marker of diabetes severity, was associated with more prolonged survival in veterans prescribed enzalutamide compared with abiraterone [24]. Instead, our policy tree chose peripheral neuropathy. However, peripheral neuropathy is a complication of diabetes (in addition to being a predictor of severe diabetes, it could be caused by factors such as injury and exposure to toxins and is also a common side effect of cancer treatment) [7,25]. Vision loss can be a complication of diabetes and is associated with increased hospitalizations [26], impaired cognition [27], and higher mortality [28]. To delay diabetic neuropathy, apart from good glycemic control, it has also been found beneficial to control blood pressure [29]. All these could have contributed to choosing peripheral neuropathy instead of diabetes and hypertension by our policy tree.

In addition, our policy tree incorporated osteoporosis, which was also one of the top covariates contributing to the variation between the 2 subgroups. Previous studies show an increased risk of falls and fractures among patients with mCRPC taking ARPIs [30,31]. Osteoporosis and abnormal gait may identify impaired functional status or increased frailty, which is associated with survival outcomes in mCRPC [9]. Osteoporosis is a hidden nonmotor symptom of Parkinsons’ disease, which contributed to variation in subgroups and was chosen by the policy tree [32]. Bilirubin levels could contribute to osteoporosis, which is also shown as one of the top variations among the two subgroups [33]. Our policy tree chose race, and existing studies have demonstrated that abiraterone has improved survival when given as the first line of treatment for the African American population [34]. Together, these results show that our policy tree could elicit a variety of comorbidities and suggest therapy.

In addition to identifying the covariates, selecting treatments for patients with mCPRC using the generated policy tree considering their demographics, laboratory measures, and comorbidities may help mitigate complications and prolong overall survival. As no generic version of enzalutamide is available in the US, the drug remains expensive. Therefore, from a cost perspective, it is prudent to administer enzalutamide only to patients most likely to benefit from it in terms of survival duration.

One of the limitations of this study is that the data are only for US veterans, so that the generalization of the prescription rules to the other populations is not advised. Another limitation is that we excluded patients with mCRPC who had docetaxel, cabaxitaxel, mitoxantrone, and sipuleucel-T before or after prescribing abiraterone or enzalutamide. This was performed to identify patients with higher comorbid disease burdens with fewer therapeutic options. However, other treatments are prescribed for patients with mCRPC. A third limitation of this study was the lack of data on time-varying covariates. Even though causal survival forest is robust to censoring, it will be interesting to include time-varying covariates in the future and to check for differences in the generated policy tree. The fourth limitation of this study was that we used only 70% of the training data to generate a policy tree, because the causal survival forest used cross-fitting and honesty approaches. Furthermore, we estimated the overall benefit of the policy tree (ie, the average difference in the follow-up days obtained if the patients with mCRPC are administered enzalutamide or abiraterone by adhering to the policy tree) in this study without considering the benefits to subgroups of patients based on race or other demographics. Future research should estimate the benefit of such policy trees to see if they are equally beneficial to various subgroups of demographics. Similarly, this study used only mCRPC patient data. Future studies should develop treatment selection for earlier stages of prostate cancer, such as the metastatic hormone-sensitive stage.

We used a machine learning-based survival approach, that is, causal survival forest model, to estimate the variations in the survival of patients with mCRPC who were administered either enzalutamide or abiraterone. Our estimation revealed that patients with mCRPC who were administered enzalutamide had longer survival than patients who were administered abiraterone. We were able to use a data-driven approach to identify heterogeneity and subgroups of patients. We then created policy trees to aid physicians administer personalized treatment, that is, abiraterone versus enzalutamide, based on patient characteristics, including demographics, laboratory values, and comorbidities, to improve survival duration.