Table 2 Examples of mitigation strategies that reduced AI bias in CVD prediction or detection
From: Artificial intelligence bias in the prediction and detection of cardiovascular disease
Author (year) Study aim | Unit of analysis and sample size | AI application | How the outcome was established | Bias mitigation strategies and outcome |
|---|---|---|---|---|
Dupulthys et al. (2024)48 To determine if an AI-enhanced ECG with EHR-extracted risk factors can be used to identify subclinical AF during SR in a screening scenario | 173,537 ECGs (from 68,880 patients who may have had AF risk factors) from Roeselare, Belgium SR ECGs from patients with and without AF and with or without AF risk factors were analyzed | AI algorithm trained on ECGs with or without confirmed AF and with or without the inclusion of AF risk factors (e.g., previous CVD, obesity, smoking) to detect AF in patients presenting with SR | The diagnostic label of AF or SR was automatically assigned by the GE MUSE Cardiology Information System | Dataset balancing (age and sex-matched data). Dataset balancing showed performance consistency across test-sets and when risk factors were included in the model. In contrast, the model trained without age or sex-matching resulted in age bias and unstable model performance across test-sets. |
Meng et al. (2022)50 To introduce a novel clinical knowledge-enhanced ML pipeline to support timely and cost-effective IHD prediction | Cleveland Clinic Foundation IHD dataset of 303 patients with and without IHD that may have cardiac risk factors | ML-based models to diagnose IHD | IHD was defined as ≥50% narrowing of at least 1 of the coronary arteries on coronary angiography | Clinical input during model development and variable selection. The model with clinical input resulted in superior accuracy compared to the ML models without clinical input. |
