Background: The detection of subtle cognitive impairment in a clinical setting is difficult, and because time is a key factor in small clinics and research sites, the brief cognitive assessments that are relied upon often misclassify patients with very mild impairment as normal. In this study, we seek to identify a parsimonious screening tool in one stage, followed by additional assessments in an optional second stage if additional specificity is desired, tested using a machine learning algorithm capable of being integrated into a clinical decision support system.
Methods: The best primary stage incorporated measures of short-term memory, executive and visuospatial functioning, and self-reported memory and daily living questions, with a total time of 5 minutes. The best secondary stage incorporated a measure of neurobiology as well as additional cognitive assessment and brief informant report questionnaires, totaling 30 minutes including delayed recall. Combined performance was evaluated using 25 sets of models, trained on 1181 ADNI participants and tested on 127 patients from a memory clinic.
Results: The 5-minute primary stage was highly sensitive (96.5%) but lacked specificity (34.1%), with an AUC of 87.5% and DOR of 14.3. The optional secondary stage increased specificity to 58.6%, resulting in an overall AUC of 89.7% using the best model combination of logistic regression for stage 1 and gradient-boosted machine for stage 2.
Conclusions: The primary stage is brief and effective at screening, with the optional two-stage technique further increasing specificity. The hierarchical two-stage technique exhibited similar accuracy but with reduced costs compared to the more common single-stage paradigm.