AbstractBackgroundPresence of significant subjective complaints about cognition (SCD) is considered the first behavioral manifestation of Alzheimer disease (AD). However, SCD has not yet overcome the challenge of becoming a reliable preclinical AD marker. Severity indices were proposed to improve the accuracy of complaints when predicting the risk of AD (Jessen et al., 2010).Our aim was to compare the predictive accuracy of ML algorithms using a more (95%ile) or less (5%ile) restrictive cut‐off point in severity of complaints for classification in Low (LSC) and High (HSC) subjective complaints groups.MethodOne hundred and ninety‐nine participants from the Compostela Aging Study (ComPAS) were identified at baseline as SCDs and completed three follow‐up measurements (54‐72 months). Considering their total scores in the Subjective Memory Questionnaire, participants were classified as LSC and HSC following using two distinct cut‐off points: complaint scores above or below than 5%ile vs 95%ile. Participants were labeled as ‘worsening’ or ‘stable’ based on their progression or not to MCI or AD. ML classifier algorithms (Random Forest, Support Vector Machine, Extra Tree) were applied to forty‐one measures (socio‐demographic, time, health, cognitive, behavioral, cognitive reserve) collected at baseline.ResultsThe best performing model was the Random Forest (95%ile: PPV =.87; Sensibility =.92; Specificity =.38; 5%ile: PPV =.66; Sensibility=.61; Specificity =.51). The confusion matrix (Figure 1) showed that: a) both, the more (95%ile) and the less (5%ile) restrictive criteria mostly classified the HSC‐stable participants as LSC‐stable; and b) criterion 5%ile, but not 95%ile, was able to differentially identify progressors based on their complaints. Episodic memory, executive functions, depression, cognitive reserve and progression timing measures assume the highest levels of importance in the ML algorithm that differentially predicts the progression of SCD to MCI and AD according to the level of complaints.ConclusionsFor both the criteria, algorithms failed to successfully identify stable SCD participants. The less restrictive criterion resulted in a better classification than the more lenient one to differentially identify LSC and HSC participants who get worse. Cognitive, affective, cognitive reserve and progression timing were the more prominent variables in conforming the predictive algorithm.
Read full abstract