Table 2 Comparison of discrimination and calibration performance for prediction of 3-year risk of stroke using Cox and Gradient Boosted Tree (GBT) modeling approaches.

Modeling Approach and Included Data	Men		Women
Modeling Approach and Included Data	Discrimination AUC [95%CI]	Calibration χ² (p-value)	Discrimination AUC [95%CI]	Calibration χ² (p-value)
Single measurement of risk factor inputs
Cox: Single measurement at most recent visit	0.779 [0.709–0.845]	16.8 (p = 0.05)	0.756 [0.692–0.814]	17.3 (p = 0.04)
GBT: Single measurement at most recent visit	0.811 [0.753–0.867]	5.6 (p = 0.78)	0.743 [0.681–0.798]	7.3 (p = 0.61)
Sequential measurements of risk factor inputs
GBT: Sequential measurements at three visits	0.795 [0.721–0.858]	1.9 (p = 0.99)	0.741 [0.677–0.796]	14.3 (p = 0.11)
GBT: Longitudinal summary ^a of sequential measurements at three visits	0.789 [0.721–0.851]	5.0 (p = 0.83)	0.724 [0.660–0.782]	9.2 (p = 0.42)
GBT: Longitudinal summary of stroke risk estimates ^b at three visits + Single measurement at most recent visit	0.786 [0.719–0.851]	29.5 (p < 0.01)	0.750 [0.683–0.811]	20.2 (p = 0.02)

^a Longitudinal summary of risk factor inputs included (i) mean, standard deviation, minimum, and maximum values recorded for continuous risk factor inputs, and (ii) mean and standard deviation for binary risk factor inputs.
^b Longitudinal summary of risk estimates included individual-level random slopes and random intercepts from linear mixed effects models; variance of Cox-generated risk estimates across three visits; and absolute changes in Cox-generated risk estimates between consecutive visits.

Quick links

Search