Identifying mild cognitive impairment and mild Alzheimer’s disease based on spontaneous speech using ASR and linguistic features

Out of 51 reviewed papers, only seven did three-way or four-way classification.

“Out of 51 reviewed papers, only seven did attempt three-way [50, 62–64] or four-way [12, 65, 66] classification.”

Seven reviewed papers are concerned with three or four disease stages.

“Outside of those, seven papers are concerned with three or four disease stages [12, 50, 62–66], two explore longitudinal cognitive changes (although still through binary classification) [67, 68], and one describes discourse patterns [69].”

Egas López et al. reported 56% and Gosztolya et al. 66.7% for 3-way classification in balanced datasets.

“Other studies attempting 3-way classification experiments in balanced datasets are Egas López et al. [62], 56% and Gosztolya et al. [63] with 66.7%.”

Off-the-shelf ASR tools are not readily prepared for dementia research, and therefore some reviewed studies developed their own custom ASR systems.

“It is becoming increasingly obvious that off-the-shelf ASR tools are not readily prepared for dementia research, and therefore some reviewed studies developed their own custom ASR systems [44, 63].”

The main strategies used to address class imbalance are subsampling, stratified CV, and careful choice of evaluation methods including the UAR metric and ROC curve analysis.

“In the studies that seek to address class imbalance in their datasets, the main strategies used are subsampling [11, 60], use of statistical methods such as stratified CV [74], and careful choice of evaluation methods including use of the UAR metric [63] and ROC curve analysis [97].”