AI could help cardiologists predict bleeding, stroke risks in AFib patients on DOACs

Chaudhary et al. aimed to see if AI could help clinicians predict the risk of major bleeding events in these patients better than conventional risk scores that already exist. They explored electronic health record (EHR) data from more than 20,000 adult patients with non-valvular AFib who were given DOACs from January 2010 to November 2022. All data came from a single U.S. health system. The median patient age was 73 years old, and 45% were women. Outcomes were examined after one year, two years and five years.

The group used 70% of the data for training purposes and the remaining 30% to test the effectiveness of their algorithms. They experimented with several different machine learning approaches.

Three existing prediction models were used for benchmarking purposes: HAS-BLED, ORBIT and ATRIA.

Overall, the researchers found that their AI models “demonstrated modest improvement in discriminative power, overall performance, risk stratification and calibration” compared to the three traditional risk scores when it came to predicting bleeding events after one year. Similar improvements were seen with bleeding events after two years and five years.

In addition, the group noted, their AI models were able to anticipate hemorrhagic strokes better than traditional risks cores after one year, two years and five years.

It is worth noting, however, that the area under the ROC curve (AUC-ROC) for the team’s AI models were still not as high as clinicians may have seen when predicting other important patient outcomes. While some of the most effective AI models have achieved an AUC-ROC above 0.90, for example, the most successful algorithms in this study still landed in the range of 0.69 to 0.76. 

“This may be attributed to the complexity of the underlying biological processes, unknown or unmeasured confounders, and data limitations, highlighting the need for further refinement and validation of these models,” the authors wrote.

Chaudhary and colleagues also noted that AI was better at performing these risks in some patient groups than others. For instance, one random forest model achieved an AUC-ROC of 0.76 when used to evaluate the full patient population, but it only achieved an AUC-ROC of 0.66 when used on a cohort of low-comorbidity patients.

“Our study demonstrates the consistent improvement in the performance of machine learning models compared to conventional risk scores in predicting bleeding among nonvalvular AFib patients on DOACs at up to 5-years of follow-up,” the group concluded. “The future work entails refining these models, external validation, improving explainability, defining the utility framework, and translation to clinical practice for supporting informed decision-making.”

Click here to read the full study.