Strambi, Luigi-Ferini

Abstract Introduction This study investigated the impact of visual editing of sleep recordings for use in predicting neurodegenerative disorder (NDD) risk. Methods All participants underwent in-home Sleep Profiler studies. Forty-four were diagnosed with mild cognitive impartment (MCI = 40) or Alzheimer’s disease (AD = 4)(MCI/AD: 55% women, 72.6+/-8.6 years), 26 with REM sleep behavior disorder (RBD: women = 27%, age = 67+/-9.8), and 29 were controls (NC: women = 69%, age = 63+/-11.5). A 4-class machine learning algorithm used age plus eight sleep biomarkers to assign NDD risk probabilities. Records with CG-assigned probabilities >70% were labeled Probably-Normal, and from 45-70% Likely-Normal. Similar thresholds were used to assign Likely- or Probably-Abnormal. NDD risk characterizations were generated based on auto-scoring (“auto”), and again after the recordings were visually edited for accuracy (“edited”). Results Comparing the auto- vs. edited-NDD classifications in the MCI/AD group, the same NDD risk classifications were observed in 91% (39/43), with 2 cases shifting from Likely-Normal to Probably-Normal, one shifted from Likely-AD to Likely-Normal with AD Indications, and one changed from Probably-AD to Likely-AD. The same NDD risk classifications were observed in 81% (21/26) of the RBD cases, 3 shifted from Probably-Abnormal to Likely-Abnormal, one shifted from Likely-Normal to Probably-Normal, and one changed from Likely-Abnormal to Likely-Normal. The same NDD risk classifications were observed in 86% of the NC group, with 4 cases shifting from Likely-Normal to Probably-Normal. Conclusions These findings suggest that the likelihood of a gross misclassification of NDD risk resulting from auto-detected sleep biomarkers is relatively low with inconsistencies directed toward slightly increased NDD risk classifications.