Table 2.

Absolute errors of clinician and model predictions.

Modela	Mean (SD)b	Median (IQR)c	Clinician error differenced	P value
clinicians	1.38 (1.31)	1.10 (0.60-1.80)	N/Ae	N/A
current	1.86 (1.55)	1.50 (0.80-2.58)	–0.49 (–0.68 to –0.29)	<.0001
lm	1.19 (1.03)	0.94 (0.50-1.67)	0.19 (0.04 to 0.34)	.0109
ridge	1.14 (1.01)	0.97 (0.48-1.47)	0.23 (0.10 to 0.36)	.0005
lstm	1.08 (1.01)	0.91 (0.37-1.55)	0.29 (0.14 to 0.44)	.0002
lasso	1.08 (0.95)	0.94 (0.51-1.38)	0.30 (0.17 to 0.43)	<.0001
nn	1.06 (1.02)	0.87 (0.34-1.36)	0.32 (0.18 to 0.45)	<.0001
rf	1.04 (0.91)	0.76 (0.34-1.48)	0.34 (0.20 to 0.48)	<.0001
xgboost	1.01 (0.90)	0.88 (0.37-1.41)	0.37 (0.22 to 0.52)	<.0001

^aModels are as described in Textbox 1, with 210 predictions made by each model. Clinician predictions were from all role groups (attendings, advanced practitioners, residents, and nurses).

^bPrediction mean absolute error and SD (mg/dL).

^cPrediction median absolute error and IQR (mg/dL).

^dMean error differences (mg/dL, clinician absolute error minus model absolute error) with 95% confidence range and comparisons by paired t test. Positive values reflect higher prediction errors by clinicians.

^eN/A: not applicable.