Understanding disease relationships is vital for uncovering causal mechanisms and improving prevention and treatment strategies. Traditional disease networks, often based on Pearson correlation, are limited by their inability to account for confounding effects, making them unsuitable for causal inference. In this study, we present a novel approach using partial correlation to construct a putative causal disease network from comorbidity data in the Human Disease Network (HuDiNe), which includes over 291,000 associations among 995 diseases derived from Medicare claims. Using a James–Stein shrinkage estimator, we computed partial correlations while controlling for the influence of other diseases. This method identified 7,894 statistically significant associations among 697 diseases, capturing both positive and negative relationships. Compared to Pearson-based networks, the partial correlation network was sparser and less modular, highlighting its specificity for direct associations. We validated key findings, particularly the high connectivity of hypertension, through Mendelian randomization studies. In addition to recovering established links (e.g., hypertension with obesity and cardiovascular diseases), the method uncovered novel associations with conditions such as breast cancer, endometriosis, and glaucoma. Negative correlations—such as between diabetes and aortic aneurysm—further demonstrated the method’s ability to detect inverse relationships. Our results show that partial correlation analysis applied to summary-level data offers a promising tool for causal disease network construction. This approach enables hypothesis generation without the need for individual-level data, laying the groundwork for integrative analyses that incorporate genomic and clinical information to refine disease prediction and intervention strategies.