Identifying treatment heterogeneity in atrial fibrillation using a novel causal machine learning method

Abstract

Background: Lifelong oral anticoagulation is recommended in patients with atrial fibrillation (AF) to prevent stroke. Over the last decade, multiple new oral anticoagulants (OACs) have expanded the number of treatment options for these patients. While population-level effectiveness of OACs has been compared, it is unclear if there is variability in benefit and risk across patient subgroups.

Methods: We analyzed claims and medical data for 34,569 patients who initiated a nonvitamin K antagonist oral anticoagulant (non-vitamin K antagonist oral anticoagulant (NOAC); apixaban, dabigatran, and rivaroxaban) or warfarin for nonvalvular AF between 08/01/2010 and 11/29/2017 from the OptumLabs Data Warehouse. A machine learning (ML) method was applied to match different OAC groups on several baseline variables including, age, sex, race, renal function, and CHA₂DS₂ -VASC score. A causal ML method was then used to discover patient subgroups characterizing the head-to-head treatment effects of the OACs on a primary composite outcome of ischemic stroke, intracranial hemorrhage, and all-cause mortality.

Results: The mean age, number of females and white race in the entire cohort of 34,569 patients were 71.2 (SD, 10.7) years, 14,916 (43.1%), and 25,051 (72.5%) respectively. During a mean follow-up of 8.3 (SD, 9.0) months, 2,110 (6.1%) of patients experienced the composite outcome, of whom 1,675 (4.8%) died. The causal ML method identified 5 subgroups with variables favoring apixaban over dabigatran; 2 subgroups favoring apixaban over rivaroxaban; 1 subgroup favoring dabigatran over rivaroxaban; and 1 subgroup favoring rivaroxaban over dabigatran in terms of risk reduction of the primary endpoint. No subgroup favored warfarin and most dabigatran vs warfarin users favored neither drug. The variables that most influenced favoring one subgroup over another included Age, history of ischemic stroke, thromboembolism, estimated glomerular filtration rate, Race, and myocardial infarction.

Conclusions: Among patients with AF treated with a NOAC or warfarin, a causal ML method identified patient subgroups with differences in outcomes associated with OAC use. The findings suggest that the effects of OACs are heterogeneous across subgroups of AF patients, which could help personalize the choice of OAC. Future prospective studies are needed to better understand the clinical impact of the subgroups with respect to OAC selection.

Figure 1

Training and validation of causal machine learning model. Workflow of the repeated training and validation procedure for hybrid causal machine model consisting of the CT, TMLE, and net benefit modules. Each matched head-to-head OAC comparison data are randomly divided into 2 parts: 80% used for developing the clustering and estimating ATE and 20% used to validate the clusters. ATE, average treatment effect; CT, causal tree; OAC, oral anticoagulants; TMLE, targeted maximum likelihood estimation.

Figure 2

Matched event rates of major bleeding, All-cause mortality, and the composite endpoint. All numbers were calculated using matched head-to-head OAC comparison groups. Event rates were calculated as number of events per 1,000 person-year. OAC, oral anticoagulants.

Figure 3

Population ATE. Population level average treatment effects of the OACs, apixaban, dabigatran, rivaroxaban, and warfarin on the endpoints, major bleeding, all-cause mortality, and the composite endpoint. All values were calculated using matched head-to-head OAC comparison groups. OAC, oral anticoagulants.

Figure 4

Subgroups of apixaban vs dabigatran users with respect to the primary composite endpoint. The subgroups are the terminal nodes of the optimal causal ML model. The green subgroups favor the use of apixaban. All the values were estimated based on the matched sample of apixaban and dabigatran users. ATE indicates average treatment effect; ERPO, events rate per 1,000.

Figure 5

Subgroups of apixaban vs rivaroxaban users with respect to the primary composite endpoint. The subgroups are the terminal nodes of the optimal causal forest model. The green subgroups favor the use of apixaban. All the values were estimated based on the matched sample of apixaban and rivaroxaban users. ATE, average treatment effect; ERPO, events rate per 1,000.

Figure 6

Subgroups of dabigatran vs rivaroxaban users with respect to the primary composite endpoint. The subgroups are the terminal nodes of the optimal causal forest model. The green subgroup favors the use of dabigatran while the red subgroup favors rivaroxaban. All the values were estimated based on the matched sample of dabigatran and rivaroxaban users. ATE, average treatment effect; ERPO, events rate per 1,000.

Figure 7

Subgroups of apixaban vs warfarin users with respect to the primary composite endpoint. The subgroups are the terminal nodes of the optimal causal forest model. The green subgroups favor the use of apixaban. All the values were estimated based on the matched sample of apixaban and warfarin users. ATE, average treatment effect; ERPO, events rate per 1,000.

Figure 8

Subgroups of dabigatran vs warfarin users with respect to the primary composite endpoint. The subgroups are the terminal nodes of the optimal causal ML model. The green subgroup favors the use of dabigatran while the red subgroup favors warfarin. All the values were estimated based on the matched sample of dabigatran and warfarin users. ATE, average treatment effect; ERPO, events rate per 1,000; ML, machine learning.

Figure 9

Subgroups of rivaroxaban vs warfarin users with respect to the primary composite endpoint. The subgroups are the terminal nodes of the optimal causal ML model. The green subgroups favor the use of rivaroxaban. All the values were estimated based on the matched sample of rivaroxaban and warfarin users. ATE, average treatment effect; ERPO, events rate per 1,000; ML, machine learning.