Machine Learning Prediction of Visual Outcome after Surgical Decompression of Sellar Region Tumors

Introduction: This study aims to develop a machine learning-based model integrating clinical and ophthalmic features to predict visual outcomes after transsphenoidal resection of sellar region tumors. Methods: Adult patients with optic chiasm compression by a sellar region tumor were examined to develop a model, and an independent retrospective cohort and a prospective cohort were used to validate our model. Predictors included demographic information, and ophthalmic and laboratory test results. We defined "recovery" as more than 5% for a p-value in mean deviation compared with the general population in the follow-up. Seven machine learning classifiers were employed, and the best-performing algorithm was selected. A decision curve analysis was used to assess the clinical usefulness of our model by estimating net benefit. We developed a nomogram based on essential features ranked by the SHAP score. Results: We included 159 patients (57.2% male), and the mean age was 42.3 years old. Among them, 96 patients were craniopharyngiomas and 63 patients were pituitary adenomas. Larger tumors (3.3 cm vs. 2.8 cm in tumor height) and craniopharyngiomas (73.6%) were associated with a worse prognosis (p < 0.001). Eyes with better outcomes were those with better visual field and thicker ganglion cell layer before operation. The ensemble model yielded the highest AUC of 0.911 [95% CI, 0.885-0.938], and the corresponding accuracy was 84.3%, with 0.863 in sensitivity and 0.820 in specificity. The model yielded AUCs of 0.861 and 0.843 in the two validation cohorts. Our model provided greater net benefit than the competing extremes of intervening in all or no patients in the decision curve analysis. A model explanation using SHAP score demonstrated that visual field, ganglion cell layer, tumor height, total thyroxine, and diagnosis were the most important features in predicting visual outcome. Conclusion: SHAP score can be a valuable resource for healthcare professionals in identifying patients with a higher risk of persistent visual deficit. The large-scale and prospective application of the proposed model would strengthen its clinical utility and universal applicability in practice.