Unveiling the sound of the cognitive status: Machine Learning-based speech analysis in the Alzheimer’s disease spectrum

Table 4 Regression metrics obtained by the best models in the prediction of each cognitive domain

Sample	Metric	Attention	Executive function	Language	Memory	Visuospatial function
All	RMAE (%)	18.007	17.421	17.801	18.330	17.880
	MAE	0.474	0.477	0.458	0.530	0.515
	Correlation	0.468	0.563	0.569	0.499	0.574
	EV^a (%)	21.778	31.696	32.345	24.924	32.964
SCD	RMAE (%)	22.063	23.295	16.736	26.588	18.401
	MAE	0.580	0.638	0.431	0.769	0.530
	Correlation	0.062	0.082	0.009	0.076	0.002
	EV^a (%)	–	–	–	–	–
MCI	RMAE (%)	16.282	15.339	15.573	17.023	16.125
	MAE	0.428	0.420	0.401	0.492	0.464
	Correlation	0.370	0.477	0.487	0.383	0.485
	EV^a (%)	12.494	17.759	17.559	10.321	18.114
ADD	RMAE (%)	19.634	19.141	21.483	18.265	20.440
	MAE	0.517	0.524	0.553	0.528	0.589
	Correlation	0.243	0.205	0.330	-0 036	0 347
	EV^a (%)	3.668	–	6.192	–	6.560

For each cognitive domain, the regression metrics obtained by the best models are presented. The random forest (RF) was the best model at predicting the attention score, while the extreme gradient boosting (XGB) performed better on all other scores. The metrics were calculated for the entire sample and stratified by clinical phenotype
Abbreviations: MAE, mean absolute error; RMAE, relative MAE (described in the “Experimental setup” section); EV, explained variance; SCD, subjective cognitive decline; MCI, mild cognitive impairment; ADD, Alzheimer’s disease dementia
^a The EV is not shown when the variance of the true values was lower than the variance of the residuals

ISSN: 1758-9193