Fig. 5: Demographic and institutional performance analysis of SpineXtract.
From: AI-driven label-free Raman spectromics for intraoperative spinal tumor assessment
a Sex-specific performance (n = 142) shows balanced accuracy across tumor types for both females and males, with slight variations in ependymoma classification. b Age group-specific performance reveals consistent accuracy across age cohorts (<30, 30-60, >60 years) for most tumor types. c Cross-institutional diagnostic performance compares metrics across three tertiary medical centers (New York, Michigan, Vienna) demonstrating robust generalization with strong ROC curves for all tumor types at each institution and corresponding confusion matrices. d Correlation between tumor patch count and diagnostic performance reveals higher patch counts generally yield more accurate predictions (189.4 patches for correct vs. 103.7 for incorrect predictions, p < 0.05), with contrasting heatmap examples of diagnostic integration shown for correctly identified schwannoma and incorrectly classified myxoid ependymoma, highlighting the importance of diagnostic feature recognition in challenging cases.
