External Validation of AI for Rotator Cuff Calcific Tendinopathy Shows Diminishing Returns Beyond 75% Training Volume

In external validation of a VGG19 model for rotator cuff calcific tendinopathy on shoulder radiographs, AUC plateaued at 75% training volume (0.945) vs 100% (0.953; P=.161) and models were equivalent within a ±0.02 AUC margin.

• External AUC increased from 0.835 with 25% training data to 0.903 (50%), 0.945 (75%), and 0.953 (100%); the gain from 75% to 100% was not significant (Δ AUC=0.008; P=.161).
• Across three independent 75% training subset realizations, external AUC remained highly stable (0.945–0.948), and all were equivalent to full-data training within ±0.02 AUC margin (TOST P=.016).
• Training with 75% data reduced training time by 11.4% and measured energy by 15.4% compared to using all data.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.