. 2024 Apr 15;5(4):416–426. doi: 10.1093/ehjdh/ztae029

Table 2.

Artificial intelligence electrocardiogram model performance metrics by site

Site	AUC	Optimal probability threshold	Accuracy	Sensitivity	Specificity
Overall	0.922 (0.910, 0.934)	0.11^a	86.9% (85.8%, 87.8%) 4030/4640	82.8% (79.9%, 85.4%) 640/773	87.7% (86.6%, 88.7%) 3390/3867
Overall	0.922 (0.910, 0.934)	0.18^b	89.2% (88.3%, 90.1%) 4139/4640	80.6% (77.6%, 83.3%) 623/773	90.9% (90.0%, 91.8%) 3516/3867
Bern	0.835 (0.782, 0.887)	0.11^a	87.2% (86.0%, 88.3%) 2978/3416	62.1% (49.3%, 73.8%) 41/66	87.7% (86.5%, 88.8%) 2937/3350
Bern	0.835 (0.782, 0.887)	0.07^b	83.3% (82.0%, 84.5%) 2844/3416	68.2% (55.6%, 79.1%) 45/66	83.6% (82.3%, 84.8%) 2799/3350
Oxford	0.900 (0.869, 0.931)	0.11^a	78.4% (74.1%, 82.2%) 330/421	73.7% (68.4%, 78.5%) 224/304	90.6% (83.8%, 95.2%) 106/117
Oxford	0.900 (0.869, 0.931)	0.04^b	84.6% (80.7%, 87.9%) 356/421	83.9% (79.3%, 87.8%) 255/304	86.3% (78.7%, 92.0%) 101/117
Seoul	0.948 (0.933, 0.964)	0.11^a	89.9% (87.6%, 91.9%) 722/803	93.1% (90.1%, 95.3%) 375/403	86.8% (83.0%, 89.9%) 347/400
Seoul	0.948 (0.933, 0.964)	0.18^b	90.4% (88.2%, 92.4%) 726/803	91.6% (88.4%, 94.1%) 369/403	89.2% (85.8%, 92.1%) 357/400

95% confidence intervals are shown in parentheses.

^aOptimal AI-ECG probability threshold as defined in the algorithm derivation cohort.

^bArtificial intelligence electrocardiogram probability threshold as optimized for each cohort using the Youden index method.