Artificial Intelligence (AI) is revolutionising healthcare by enhancing diagnostics, enabling personalised treatment, and improving patient outcome predictions. However, these systems also pose significant risks of perpetuating and amplifying biases, which could lead to inequitable healthcare delivery. This paper presents a synthesis of 25 peer-reviewed studies, narrowed down from an initial pool of 498 using eight screening criteria, focused on identifying sources of bias and evaluating mitigation strategies in medical AI. Sources of bias include imbalanced training data, algorithmic design flaws, deployment bias, and clinical implementation inequities, which often disproportionately impact marginalised populations. Notably, 22 out of the 25 studies emphasised data bias as the primary source, underscoring a strong research focus on dataset representativeness and quality. Common bias types identified include racial, gender, age-related, and socio-economic disparities. Mitigation approaches encompass pre-processing and data-centric techniques (e.g., resampling, disaggregation, data augmentation using GAN), model-centric strategies (e.g., multi-calibration and smoothing), fairness evaluation and transparency monitoring (e.g., SHAP, LIME, fairness metrics), and participatory approaches that promote transparency and stakeholder engagement. Notable gaps in the literature include limited attention to temporal fairness, mental health AI, and paediatric or geriatric specific biases. Among the selected studies, medical imaging emerged as the most frequently addressed application area. Some of these techniques, reported in single studies, indicate a need for further validation and integration into standard practice. These findings underscore the need for more comprehensive, equitable, and domain-specific approaches to responsible AI implementation in healthcare.