Early diagnosis of Alzheimer’s disease using machine learning: a multi-diagnostic, generalizable approach

Table 5 Classifier performance for “HC vs. AD” using morphometric features only using ADNI and OASIS subjects

Experiment	Training set	Testing set	MCC [CI: 95%]	BAC [CI: 95%]	ROC AUC [CI: 95%]	Sens [CI: 95%]	Spec [CI: 95%]	PPV (prevalence) [CI: 95%]	NPV (prevalence) [CI: 95%]	PPV (standard) [CI: 95%]	NPV (standard) [CI: 95%]	TN	FP	FN	TP	MCC p-value (vs. w/ GT)
B1	ADNI MPRAGE (N=194)	ADNI MPRAGE (N=84)	0.814 [0.679;0.927]	90.8% [84.1%;96.2%]	97.3% [93.6%;99.6%]	94.9% [87.2%;100.0%]	86.7% [76.2%;95.5%]	81.7% [69.6%;93.3%]	96.4% [90.5%;100.0%]	87.7% [78.6%;95.7%]	94.4% [85.6%;100.0%]	39	6	2	37	0.974
B2	OASIS (N=365)	OASIS (N=152)	0.564 [0.344;0.771]	75.6% [65.3%;87.2%]	94.1% [89.7%;97.8%]	55.0% [33.3%;77.3%]	96.2% [92.5%;99.2%]	90.1% [73.5%;98.4%]	77.3% [68.9%;87.5%]	93.5% [81.6%;99.0%]	68.1% [58.1%;81.4%]	127	5	9	11	0.998
B3	OASIS (N=517)	ADNI MPRAGE (N=278)	0.722 [0.657;0.793]	84.3% [80.3%;88.4%]	95.5% [93.1%;97.5%]	70.5% [63.0%;78.2%]	98.0% [95.7%;100.0%]	95.7% [90.2%;100.0%]	84.1% [80.5%;88.0%]	97.2% [93.6%;100.0%]	76.9% [72.1%;82.1%]	146	3	38	91	0.623
B4	ADNI MPRAGE (N=278)	OASIS (N=517)	0.641 [0.551;0.733]	87.2% [82.4%;97.0%]	94.3% [91.3%;97.0%]	83.6% [74.6%;91.7%]	90.9% [88.3%;93.5%]	85.2% [80.0%;89.8%]	89.9% [84.7%;94.7%]	90.2% [86.4%;93.4%]	84.7% [77.7%;91.8%]	409	41	11	56	0.265
B5	ADNI MPRAGE and OASIS (N=295)	ADNI MPRAGE and OASIS (N=127)	0.811 [0.701;0.906]	90.6% [85.3%;95.5%]	97.4% [94.9%;99.3%]	91.5% [83.6%;98.3%]	89.7% [81.8%;96.2%]	84.8% [74.2%;94.2%]	94.4% [88.9%;98.9%]	89.9% [82.1%;96.3%]	91.4% [83.3%;98.3%]	61	7	5	54	0.383

MCC p-value refers to the p-value for the MCC metric for the comparison with the equivalent classifier (i.e., same training and test sets) with morphometric and GT features. PPV/NPV “prevalence” are calculated with an AD prevalence of 38.5% (this corresponds to the prevalence of AD relative to HC based on prevalence estimates from the first visit in the clinical setting of 42.0% for HC and 26.3% for AD) [48]. PPV/NPV “standard” are calculated with a prevalence of 50% to allow for comparison with other studies
CI confidence interval, MCC Matthew’s correlation coefficient, ROC AUC area under the receiver operating characteristic curve, BAC balanced accuracy, Sens sensitivity, Spec specificity, PPV positive predict value, NPV negative predictive value, TN true negatives, FP false positives, FN false negatives, TP true positives. The most global classifier is highlighted

ISSN: 1758-9193