Machine Learning-Based models in prediction of Ki-67 in pituitary adenoma: A systematic review and Meta-Analysis.

BACKGROUND

The Ki-67 labeling index is a crucial marker of proliferative activity in pituitary adenomas (PAs) and is associated with tumor aggressiveness and recurrence. Machine learning (ML)-based models can assess Ki-67 non-invasively by incorporating clinical and radiological data. This study aims to evaluate the predictive accuracy of ML-based models in predicting Ki-67 status in PAs.

METHODS

Four electronic databases were searched through August 3, 2025. Studies that developed ML-based models to predict Ki-67 status in PAs and reported area under the curve (AUC), accuracy (ACC), sensitivity (SEN), or specificity (SPE) were included.

RESULTS

Six studies involving 2,746 PAs were included. For all ML-based models, the pooled AUC was 0.82 (95% CI: 0.64–0.99), ACC was 0.83 (95% CI: 0.79–0.88), SEN was 0.75 (95% CI: 0.64–0.84), SPE was 0.86 (95% CI: 0.79–0.91), and the diagnostic odds ratio (DOR) was 17.7 (95% CI: 8.4–37.5). In radiomics-only models, performance was higher with a pooled AUC of 0.89 (95% CI: 0.85–0.92), ACC of 0.82 (95% CI: 0.73–0.91), SEN of 0.67 (95% CI: 0.45–0.83), SPE of 0.91 (95% CI: 0.79–0.96), and DOR of 15.5 (95% CI: 6.9–34.8). Not all included studies reported all performance metrics; therefore, pooled estimates were calculated using the subset of studies reporting each outcome.

CONCLUSION

ML-based models have demonstrated promising diagnostic accuracy for non-invasively predicting Ki-67 status in PAs. However, these findings should be interpreted cautiously given the limited number of included studies and substantial between-study heterogeneity.