Building Equitable AI in Heart Failure Care: What Our Data Reveals

by Elinor Laws & Maxine Mackintosh

Heart failure affects millions globally, and (of course) we think AI could hold significant promise for enhancing its diagnosis and management. Yet, the effectiveness of AI is inherently tied to the quality and inclusiveness of the underlying data. We recently published a systematic review of 72 datasets used for developing AI tools in heart failure care where we uncovered critical shortcomings in demographic representation and data transparency- issues that have profound implications for clinical practice and health equity.

What we found

Our review covered over 2 million individuals from 23 countries. Although most datasets (86%) reported age and 85% included sex or gender information, only 29% provided any breakdown by race or ethnicity, and just 11% documented socioeconomic status. Even where race or ethnicity was mentioned, about 89% of the individuals were categorised as “White” or “Caucasian.” This imbalance raises serious concerns about how well these datasets represent the diverse populations affected by heart failure.

Data accessibility was another big issue. Only 28% of the datasets were fully accessible, the rest were either under managed access (requiring formal application or even payment) or completely private. This lack of openness makes it hard to conduct independent validations and fully understand the quality and composition of the data.

Most of the datasets came from clinical research studies or trials. While this data is often high quality, the strict inclusion criteria can exclude groups that are underrepresented in everyday clinical practice. As a result, AI systems trained on this data might perform well in controlled settings but may not generalise effectively to the broader, more varied patient population we see every day.

What does this mean for AI in heart failure care

The consequences of these findings could be significant. AI algorithms learn from the data they’re fed, and if that data isn’t diverse, the tools we develop could be inherently biased. For instance, when AI systems are trained on data that predominantly represents one demographic, they might not accurately predict or diagnose conditions in patients from underrepresented backgrounds, leading to misdiagnoses or inappropriate treatment recommendations that directly affect patient outcomes. Furthermore, since heart failure disproportionately impacts minoritised ethnic groups and those with lower socioeconomic status, using datasets that don’t fully capture these populations risks exacerbating existing health disparities by inadvertently disadvantaging the very groups that need the most support. Finally, the heavy reliance on clinical trial data, which often features narrow inclusion criteria, means these datasets may not reflect the complex reality of everyday clinical practice, potentially compromising the external validity of AI models when they are applied in diverse healthcare settings.

Recommendations for advancing data equity

Based on our review, we’ve come up with a few key strategies to boost the quality and representativeness of datasets used in AI development for heart failure care:

Enhanced demographic reporting: We recommend systematic reporting of demographic variables, including race, ethnicity, sex, gender, and socioeconomic status within health datasets. Comprehensive and consistent reporting is essential for understanding the full context of the data and ensuring that AI models are trained on datasets that accurately reflect the target population.

Improved data accessibility: Data sharing is crucial. With only 28% of the datasets fully accessible, there’s a clear need for more open data practices. Whether through open-access repositories or well-defined managed access protocols, making datasets available for independent validation is key to identifying and mitigating potential biases, and creating AI tools that have applicability to a variety of settings.

Diversification of data sources: While clinical trial data is invaluable, it should be complemented with data from routine clinical care and real-world evidence. These sources can offer a more accurate picture of the patient population, capturing a broader range of experiences and health outcomes. This approach is essential for developing AI tools that are robust and effective in everyday settings.

Transparent dataset documentation: Beyond demographic details, we need clear documentation of dataset characteristics. This includes thorough descriptions of data collection methods, inclusion criteria, and any modifications made during data processing. Transparent documentation helps us critically assess data quality and identify potential sources of bias that could affect AI performance. The process of transparently reporting health datasets can be challenging but standards like the STANdards for data Diversity, INclusivity, and Generalisability (STANDING Together) recommendations can be used as a guide. 

---

Our review adds another example to the growing body of evidence showing that gaps in demographic representation and data transparency in datasets used for AI in healthcare can lead to biased tools that may worsen existing health disparities…and heart failure is no exception!Check out the full paper here and let us know what you think. We’d love to hear from you, whether our work has influenced your research or practice, or if you have ideas on how to improve data equity in heart failure care.