Table 3.

HealthImpact Accuracy in the Training, Internal Validation, and External Validation Populations. HealthImpact Thresholds Were Chosen to Signify Low (50), Intermediate (75), and High (90) Risk of Incident Diabetes at Three Years

HealthImpact Threshold	>50	>75	>90
Training dataset (n = 473,049)
Sample size	132,819	41,782	15,213
Sensitivity	69.23	38.10	17.86
Specificity	76.49	94.42	98.42
Internal validation dataset (n = 303,025)
Sample size	69,945	19,813	7,042
Sensitivity	69.18	37.74	18.01
Specificity	80.11	95.68	98.80
External validation datasets
Year 1 (n = 1,383,743)
Sample size	424,294	167,124	71,309
Sensitivity	80.54	53.70	31.07
Specificity	69.91	88.40	95.15
Positive predictive value	2.99	5.05	6.85
Negative predictive value	99.68	99.40	99.17
Year 2 (n = 1,054,142)
Sample size	342,925	139,977	61,203
Sensitivity	80.90	54.45	32.13
Specificity	68.73	87.79	94.88
Positive predictive value	6.30	10.40	14.03
Negative predictive value	99.28	98.67	98.17
Year 3 (n = 827,969)
Sample size	278,043	116,235	51,758
Sensitivity	80.88	54.52	32.35
Specificity	68.54	87.78	94.92
Positive predictive value	10.35	16.70	22.25
Negative predictive value	98.76	97.73	96.90