Background <p>Visual cognitive impairment is among the most common postoperative cognitive dysfunctions, significantly impacting recovery and quality of life in elderly patients. However, effective preoperative prediction methods remain lacking. We developed a machine learning model using graph theory analysis of resting-state functional connectivity networks to predict postoperative visual cognitive impairment.</p> Methods <p>In this nested case-control study, 74 elderly patients underwent preoperative rs-fMRI. Postoperative impairment was assessed using Trail Making Test Part A (TMT-A) and Digit Symbol Substitution Test (DSST). We constructed dynamic functional connectivity networks, extracting local (degree, path length, clustering coefficient, efficiency) and global features (modularity, entropy). Sparse representation-based feature selection and classification were applied to build the predictive model.</p> Results <p>16 patients (21.6%) developed postoperative impairment. Analysis revealed 16 significant features (<i>P</i> &lt; 0.05) differentiating groups, with key hubs in a visual-cognition network: Inferior occipital gyrus (ventral visual stream), Medial superior frontal gyrus (default mode/executive control), Cuneus (dorsal attention network), The model achieved an AUC = 0.877, accuracy = 0.840, sensitivity = 0.833, and specificity = 0.842 on independent testing.</p> Conclusion <p>Our graph theory-machine learning framework reliably predicts postoperative visual cognitive impairment by identifying disruptions in a clinically interpretable visual-cognition network. This approach offers potential guidance for perioperative decision-making.</p> Trial registration <p>Chinese Clinical Trial Registry (ChiCTR-DCD-15006096, 16th/March/2015, <a href="https://www.chictr.org.cn/showproj.html?proj=10583">https://www.chictr.org.cn/showproj.html?proj=10583</a>).</p>