In this research, we have developed a novel data-mining approach for detection of cognitive impairment, SPCIR (Speech Prosody-Based Cognitive Impairment Rating), which can discriminate between mild cognitive impairment and mild Alzheimer's disease from elderly using prosodic sign extracted from elderly speech during questionnaire test. This paper proposes a binary discrimination model of SPCIR using multivariate logistic regression and model selection using receiver operating characteristic (ROC) curve analysis, and reports the sensitivity and specificity of SPCIR for diagnosis (control; mild cognitive impairment/mild Alzheimer's disease).