Neuroimaging of tissue microstructure as a marker of neurodegeneration in the AT(N) framework: defining abnormal neurodegeneration and improving prediction of clinical status

Table 2 Logistic regression model performance predicting CU and combined MCI/AD clinical status

Logistic regression models	AUC (95% CI)	AIC	PLR
All models including age and sex
(0) No AT(N) predictors	(0) 0.74(0.60–0.88)	(0) − 27.1	(0) N/A
(1) CSF A/T status	(1) 0.93(0.89–0.97)	(1) 108.7	(1) − 102.68
(2) WM NDI	(2) 0.90(0.85–0.96)	(2) − 23.7	(2) 39.65
(3) WM ODI	(3) 0.81(0.65–1.00)	(3) − 5.7	(3) 21.61
(4) GM NDI	(4) 0.83(0.68–0.98)	(4) − 10.6	(4) 26.62
(5) GM ODI	(5) 0.87(0.78–1.00)	(5) 4.9	(5) 11.10
(6) CSF A/T status + WM NDI	(6) 0.96(0.93–0.99)	(6) 15.6	(6) 2.42
(7) CSF A/T status + WM ODI	(7) 0.97 (0.95–0.99)	(7) 25.7	(7) − 7.42
(8) CSF A/T status + GM NDI	(8) 0.97(0.95–0.99)	(8) 22.2	(8) − 4.20
(9) CSF A/T status + GM ODI	(9) 0.98(0.96–1.00)	(9) 27.6	(9) − 9.64

Logistic regression with Firth reduction predicted binomial clinical diagnosis outcomes (CU or MCI/AD). CU status included participants with CU diagnosis (n = 285) and A − /T − , A + /T − , or A + /T + CSF status. MCI/AD status included MCI or AD-diagnosed participants (n = 11) with A + /T − or A + /T + CSF A/T status. All models are controlled for age and sex. Receiver operator analysis (ROC) with area under the curve (AUC) assessed model prediction accuracy. Akaike information criteria (AIC) and penalized likelihood ratio (PLR) assessed model performance. Models with NODDI + CSF A/T status covariates had higher AUC values than the CSF A/T status-only model. The NODDI + CSF A/T status model with the lowest AIC and highest PLR included NODDI-NDI in AD-associated white matter regions

ISSN: 1758-9193