Targeted-BEHRT: Deep Learning for Observational Causal Inference on Longitudinal Electronic Health Records

Abstract

Observational causal inference is useful for decisionmaking in medicine when randomized clinical trials (RCTs) are infeasible or nongeneralizable. However, traditional approaches do not always deliver unconfounded causal conclusions in practice. The rise of "doubly robust" nonparametric tools coupled with the growth of deep learning for capturing rich representations of multimodal data offers a unique opportunity to develop and test such models for causal inference on comprehensive electronic health records (EHRs). In this article, we investigate causal modeling of an RCT-established causal association: the effect of classes of antihypertensive on incident cancer risk. We develop a transformer-based model, targeted bidirectional EHR transformer (T-BEHRT) coupled with doubly robust estimation to estimate average risk ratio (RR). We compare our model to benchmark statistical and deep learning models for causal inference in multiple experiments on semi-synthetic derivations of our dataset with various types and intensities of confounding. In order to further test the reliability of our approach, we test our model on situations of limited data. We find that our model provides more accurate estimates of relative risk [least sum absolute error (SAE) from ground truth] compared with benchmark estimations. Finally, our model provides an estimate of class-wise antihypertensive effect on cancer risk that is consistent with results derived from RCTs.

Systolic Blood Pressure and Cardiovascular Risk in Patients With Diabetes: A Prospective Cohort Study

“…Furthermore, as previously reported, when such rich data are provided, the unsupervised learning component (masked EHR modeling) of the T-BEHRT model works well in reducing bias in estimation. 9 When data are sparse or limited (finite sample), the doubly-robust estimation, with known benefits for finite-sample estimation, more accurately estimates RR than the variants without utilization of doubly-robust estimation. 20 Furthermore, when positivity (overlap) between exposure groups is limited, the T-BEHRT model fares better than other benchmark models.…”

“…For the deep learning approach, we used Targeted Bidirectional Electronic Health Records Transformer (T-BEHRT) for risk ratio (RR) estimation of association between SBP and cardiovascular outcomes with SBP of <120 mm Hg considered as reference group. 9 For each of these comparisons, T-BEHRT was first trained to jointly predict exposure category (propensity score) and risk of outcome with 5-fold cross-validation implemented for training and testing. 19 The T-BEHRT model incorporated all recorded diagnoses and medications in medical history prior to study entry in addition to baseline smoking status (current, former, and never a smoker)—identified by last known status in the 12 months before baseline—and sex.…”

“…Methods combining deep learning modeling for rich EHR data utilizing automatic feature extraction and confounding adjustment as well as semiparametric statistical modeling for mitigation of selection biases have shown promising results in several works utilizing semisynthetic and routine clinical EHR data. [8][9][10] In this study, we leveraged such novel deep learning methods to re-evaluate the relationship between SBP and cardiovascular events in a sample of 49 000 UK patients with diabetes.…”

“…Methods combining deep learning modeling for rich EHR data utilizing automatic feature extraction and confounding adjustment as well as semiparametric statistical modeling for mitigation of selection biases have shown promising results in several works utilizing semisynthetic and routine clinical EHR data. 8–10…”

Systolic blood pressure, chronic obstructive pulmonary disease and cardiovascular risk

“…2 8-10 With the availability of comprehensive electronic health records (EHR) and the advancement of deep learning (DL) causal modelling, the opportunity for more accurate modelling of associations among subgroups with poorer health has arisen. [10][11][12] While traditional modelling requires manual confounder selection, DL approaches such as Targeted Bidirectional EHR Transformer (T-BEHRT) automatically extract latent features that are confounding the association and more accurately estimate risk ratio (RR) in observational settings. 10 12 In this study, we applied the T-BEHRT model to evaluate the association between SBP and risk of cardiovascular outcomes in a cohort of 39 602 patients with COPD.…”