Machine Learning Augmented Causal Inference To Estimate The Treatment Effect of Tranexamic Acid In Traumatic Brain Injury

Ines Meyer, Jean-Denis Moyer, Aliénor Dreyfus, Boutonnet Mathieu, Pierre‐Julien Cungi, Arnaud Foucrier, Anatole Harrois, Arthur James, Jean‐Pierre Nadal, Julie Josse, Tobias GaussResearch 2021 · DOI

Abstract

Background: The CRASH-3 trial provides high-level evidence for administering Tranexamic Acid (TXA) in Traumatic Brain Injury (TBI). However, many clinical questions rely solely on observational data. This study explores the feasibility of using advanced causal inference methods on incomplete observational registry data to replicate the findings of the CRASH-3 benchmark trial.

Methods: Two causal inference techniques—inverse propensity weighting (IPW) and the doubly robust (DR) method—were combined with machine learning to handle missing data. The study analyzed 8,269 TBI cases from a trauma registry (683 received TXA, 7,565 did not). The primary outcome was 30-day head injury-related mortality, expressed as Average Treatment Effect (ATE).

Results: Raw mortality was 30% in the TXA group versus 15% in the control group (p<0.001), reflecting severe selection bias. IPW incorrectly indicated higher mortality for TXA (ATE 0.09–0.10). However, the DR method successfully neutralized bias, showing no significant effect on mortality (ATE -0.01), aligning with CRASH-3 results.

Conclusions: Doubly robust estimation, supported by machine learning for missing data, can effectively mitigate confounding in observational studies, providing a robust framework for generating real-world evidence.

Introduction: The Challenge of Confounding by Indication

In clinical research, Randomized Controlled Trials (RCTs) like CRASH-3 are the gold standard because they eliminate bias through random assignment. However, RCTs are costly and time-consuming. Observational registries offer abundant data but suffer from "confounding by indication": in practice, clinicians often reserve treatments like Tranexamic Acid (TXA) for the most severely injured patients.

Table 1: Baseline characteristics showing that patients in the TXA group were significantly more severely injured (e.g., lower GCS, higher ISS) than the control group.

As shown in the baseline characteristics, the TXA group had a median Glasgow Coma Scale (GCS) of 7 compared to 13 in the control group, and a median Injury Severity Score (ISS) of 38 versus 21. Without sophisticated adjustment, TXA appears to "cause" higher mortality simply because it was given to patients who were already at a much higher risk of death.

Methods: Augmented Causal Inference

To overcome these biases, researchers applied two primary causal inference techniques integrated with machine learning:

  1. Inverse Propensity Weighting (IPW): This method re-weights the population to create a synthetic cohort where treatment assignment appears random based on observed symptoms.
  2. Doubly Robust (DR) Estimation: This approach combines a treatment assignment model (propensity) with an outcome prediction model. It is "doubly robust" because the final estimate remains unbiased if either model is correctly specified.
  3. Handling Missing Data: The study utilized Multiple Imputation by Chained Equations (MICE) and Missing Incorporated in Attributes (MIA) via Random Forest algorithms to ensure that incomplete registry entries did not compromise the analysis.

Results: Reconciling Observational Data with Clinical Trials

The raw data showed a stark difference in mortality: 30% for TXA patients versus 15% for controls.

  • IPW Failure: Standard weighting failed to fully correct the selection bias, yielding an ATE of 0.09 to 0.10 (indicating a 9-10% increase in mortality associated with TXA).
  • DR Success: The Doubly Robust models successfully neutralized the confounding factors. The resulting ATE was approximately -0.01 (95% CI [-0.05, 0.03]). Because the confidence interval crosses zero, the result indicates no significant effect on 30-day mortality, which directly replicates the primary findings of the CRASH-3 trial.

While CRASH-3 identified benefits in specific subgroups (mild-to-moderate TBI), this observational analysis did not find significant subgroup effects, likely due to the smaller sample size of TXA recipients in the registry compared to the massive scale of the original RCT.

Discussion and Conclusion

This study serves as a successful "proof of concept" for Target Trial Emulation. By using machine learning to capture complex, non-linear relationships and handle missing data, the researchers demonstrated that observational data can yield reliable insights when analyzed with sufficiently robust causal frameworks.

While limitations remain—such as the potential for unmeasured confounders not captured in the registry—these augmented techniques provide a powerful tool for supplementing clinical trials and exploring treatment effects in real-world populations where RCTs may be impractical or unethical.

Summary updated

Abstract

Abstract Background : The CRASH-3 trial provides a high level of evidence on the question whether to administer Tranexamic Acid (TXA) for Traumatic brain injury (TBI). For numerous other research questions, the available evidence will not correspond to such a level of evidence and will rely on observational evidence only. The development of methodological alternatives to analyze observational data is necessary. The Crash-3trial provided the opportunity to explore the effect of TXA on TBI mortality with two distinct causal inference methods using incomplete observational data. Methods : Two causal inference techniques, inverse propensity weighting (IPW) and doubly robust method (DR), associated with machine learning method techniques to handle missing data, explored the effect of TXA administration on 30-day head injury related death expressed in registry data. The effect was expressed as Average Treatment Effect (ATE). TBI was defined as a head Abbreviated Injury Score &gt;2. The hypothesis expected the results to concur with the results obtained with the CRASH-3 benchmark trial. Results: Between September 2010 and February 2019, from a total of 20037 registry cases 8269 corresponded to the definition of TBI. A total of 683 received TXA and 7565 did not. The observed head-injury related 30-day hospital mortality rate in the group TXA was 30% (205/683) compared to 15% in the group no-TXA (1102/7565, p&lt;0.001). Causal inference with the IPW approach indicates an ATE with a higher mortality after TXA independently of the approach applied to manage missing data (ATE 0.10 (95% IC [0.06, 0.14]) or 0.09 (95% IC [0.03, 0.15])). ATE obtained with DR did not show any effect on mortality independently of the approach applied to missing data (ATE -0.01 (95% IC [-0.05, 0.03]) or -0.01 (95% IC [-0.07, 0.05])). No effect was observed in predefined subgroups. Conclusions: This study demonstrated the feasibility to apply causal inference techniques in incomplete observational data. DR based on a stronger theoretical background compared to IPW, did not show a significant association of TXA administration with in-hospital mortality. This result provides a strong incentive to explore augmented causal inference techniques on incomplete observational data coupled with techniques to handle missing values.
BibTeX
@inproceedings{meyer2021machine,
  author = {Ines Meyer and Jean-Denis Moyer and Aliénor Dreyfus and Boutonnet Mathieu and Pierre‐Julien Cungi and Arnaud Foucrier and Anatole Harrois and Arthur James and Jean‐Pierre Nadal and Julie Josse and Tobias Gauss},
  title = {{Machine Learning Augmented Causal Inference To Estimate The Treatment Effect of Tranexamic Acid In Traumatic Brain Injury}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-600886/v1},
}
Figure