Trustworthy AI for radar vital signs: detecting and mitigating gender bias in healthcare

Abstract Artificial intelligence (AI) has emerged as a fundamental component in modern healthcare, particularly in noninvasive monitoring of vital signs using radar-based systems. However, algorithmic fairness concerns, such as gender bias, can undermine trust in these systems. This study investigates the impact of gender representation in training data on the accuracy and fairness of radar-based vital sign estimation. We trained machine-learning models on 60 dataset configurations–male-only, female-only, balanced, male-dominant and female-dominant–each containing 640 radar-derived samples (3200 total). Models trained on female-dominant data achieved the highest classification accuracy (94.5%) and lowest regression error (RMSE = 0.70), whereas male-only datasets performed worst (accuracy = 78.2%, RMSE = 1.80). Disparate impact analysis revealed up to a 16.5% performance advantage for female-skewed training data, and multiple fairness metrics, including disparate impact ratio and statistical parity difference, were employed to quantify bias across subgroups. To address these disparities, we implemented a multilevel mitigation framework integrating TimeGAN-based data augmentation, fairness-aware learning constraints, and threshold adjustment. This approach reduced the bias score from 0.0261 to 0.00005 ( $$\textit{p} &lt; 0.001$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mi>p</mml:mi> <mml:mo>&lt;</mml:mo> <mml:mn>0.001</mml:mn> </mml:mrow> </mml:math> ) while maintaining a high F1 score ( $$&gt; 0.98$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mo>&gt;</mml:mo> <mml:mn>0.98</mml:mn> </mml:mrow> </mml:math> ). Our results underscore the necessity of balanced and diverse datasets and demonstrate that incorporating fairness-aware strategies can yield equitable and trustworthy AI systems for clinical vital sign monitoring.