Clinical validation of a voice-based AI tool for screening cognitive impairment: a prospective multicenter study

Early, scalable screening for cognitive impairment is needed beyond traditional paper-and-pencil tools.

To evaluate the diagnostic accuracy and usability of SPICK, a Korean-language, voice-based cognitive screener.

In a prospective multicenter validation study (ITT N = 399), the reference diagnosis cognitively impaired (mild cognitive impairment [MCI] or Alzheimer's dementia [AD]) or cognitively unimpaired (CU) was determined independently by board-certified neurologists; SPICK provided the index test result. The primary endpoint was SPICK sensitivity and specificity versus prespecified thresholds (≥ 80% and ≥ 59%). Secondary endpoints included subgroup performance (MCI vs CU; AD vs CU), accuracy, area under the receiver operating characteristic (ROC) curve [AUC], and System Usability Scale (SUS).

For cognitive impairment detection, SPICK achieved sensitivity 85.71% (95% confidence interval [CI] 81.46-89.31%) and specificity 74.29% (62.44-83.99%), meeting thresholds; AUC 0.800 and accuracy 83.71%. Subgroup analyses showed MCI sensitivity 79.64% (72.73-85.47%) and specificity 74.29% (62.44-83.99%), and AD sensitivity 91.98% (86.67-95.66%) and specificity 74.29% (62.44-83.99%). Among 395 participants with usability data, mean SUS was 67.32 (SD 19.10; median 70.0).

SPICK demonstrated clinically meaningful accuracy with acceptable usability, supporting its potential as an automated, voice-based screening tool for diverse clinical settings.