. 2019 Aug 30;7(3):e13139. doi: 10.2196/13139

Table 4.

Comparison of the hazard regression model with repeated measurements and the deep learning model with repeated measurements using validation datasets from the National Health Insurance Service-Health Screening Cohort (40-79 years of age).

Performance index		All-cause dementia	Alzheimer dementia
Performance index		DL-R^a,b versus HR-R^c	DL-R^a versus HR-R
Discrimination
	Difference between AUCs^d	0.034 (0.029-0.039)^e	0.024 (0.018-0.031)^e
	Absolute IDI^f	0.334	0.423
	Relative IDI	3.200	5.351
Reclassification
	Patients move to higher, n (%)	4163 (76.30)	2163 (86.42)
	Patients move to lower, n (%)	0 (0.00)	0 (0.00)
	Controls move to higher, n (%)	17,664 (19.52)	19,231 (21.25)
	Controls move to lower, n (%)	0 (0.00)	0 (0.00)
	NRI^g (%)	56.79^h	65.17^h

^aNRIs were calculated to determine the improvement in the performance of each model to identify individuals whose risk of dementia was more than 50%.

^bDL-R: deep learning model with repeated measurements.

^cHR-R: hazard regression model with repeated measurements.

^dAUC: area under the receiver operating characteristic curve.

^eDifference between AUCs was significant with P<.001.

^fIDI: integrated discrimination improvement.

^gNRI: net reclassification improvement.

^hNRI was significant with P<.001.