Self-supervised and supervised contrastive learning methods show strong potential in medical imaging by leveraging labeled and unlabeled data to learn transferable and clinically meaningful representations. However, frameworks such as SimCLR, SupCon, and MoCo treat each image independently and ignore clinically relevant metadata, including patient identity and patient-specific variability. Moreover, many commonly used augmentations, originally designed for natural images, can distort modality-specific anatomical or pathological structures, reducing the fidelity and clinical usability of the learned features.To address these limitations, MedPACL (Medical Patient-Aware Contrastive Learning) is introduced as a framework specifically tailored for medical imaging. It employs a contrastive loss that integrates disease labels and patient identity, promoting intra-patient consistency while enhancing inter-class separability. By explicitly modeling variability across patients and accounting for heterogeneous pathological patterns among individuals with the same diagnosis, MedPACL learns stable, interpretable, and robust representations. MedPACL also incorporates modality-specific augmentation pipelines designed to preserve semantic integrity across CT scans, chest X-rays, and MRI images. The framework is evaluated on three heterogeneous datasets and compared with leading contrastive approaches. Results demonstrate that MedPACL produces structured embeddings that capture both patient-level and disease-specific information, improving downstream tasks such as classification, clustering, representation visualization, and patient-aware data splitting. Overall, MedPACL provides a reliable and generalizable strategy for representation learning in medical imaging.