Two‐stage cost‐sensitive clinical decision support system for detecting Alzheimer’s disease

Abstract

AbstractBackgroundA two‐stage clinical decision support pipeline was developed to first recommend individuals for further testing based on easily administered screening tests, followed by a more robust identification of impairment based on additional cognitive testing and neuroimaging. The ultimate goal of the pipeline is to recommend further in‐depth diagnostic assessment for those identified as “likely impaired” and reduce the incidence of healthy controls needing to undergo costly diagnostic assessments.MethodThe first stage of the paradigm used easily‐collected demographic and questionnaire data with simple trail‐making and word recall, while the second stage introduced volumetric MRI and full cognitive testing from the MoCA and both the verbal learning and verbal fluency tasks. Both stages of the paradigm utilized gradient‐boosted ensemble classifiers trained on 1181 participants from the ADNI dataset (517 controls and 522 MCI, 142 AD) and evaluated on our cohort of 105 participants (23 controls, 59 MCI, 23 AD).ResultIn the first selection stage, all dementia patients (23/23), 93.2% of MCI patients (55/59), and 43.5% of controls (10/23) were selected for further testing. After the second stage, 73.9% of controls (17/23) were ruled out, and “likely impairment” was identified in all dementia patients (23/23) and 83.1% MCI patients (49/59) with overall 87.8% sensitivity after 50 repeats of a stochastic cross‐validation procedure.ConclusionThe ability of this two‐stage paradigm to rule out controls prior to costly in‐depth diagnostic assessment can be useful to reduce the staffing and time costs associated with appropriately identifying dementia. Overall detection was similar between a dual‐stage and single‐stage paradigm, suggesting that the dual‐stage paradigm is effective at reducing screening costs without compromising efficacy.