Are Digital Markers of Everyday Driving Comparable to Fluid Biomarkers in Identifying Preclinical Alzheimer Disease?

Abstract

AbstractBackgroundPreclinical Alzheimer’s disease (AD) may have a subtle functional signature reflected in changes in everyday driving behaviour. Driving has the potential to serve as a digital marker for AD when captured continuously and characterized accurately. This study directly compares the diagnostic performance of digital markers of everyday driving, cerebrospinal fluid (CSF), and blood biomarkers in detecting amyloid deposition in AD, using amyloid positron emission tomography (PET) as the “gold standard”.MethodParticipants were enrolled in a longitudinal study on driving and preclinical AD biomarkers at Washington University School of Medicine. The Driving Real‐World In‐Vehicle Evaluation System (DRIVES) Project uses in‐vehicle GPS dataloggers to collect daily driving among cognitive normal drivers. This study included 121 drivers (aged 65+) with complete blood, CSF, and PET biomarker data. We employed three artificial neural network (ANN) to examine the relationship between everyday driving and PET‐amyloid status. The first model solely included driving features. The second ANN model (i.e., “non‐intrusive” model) added age and education level as additional features. The third model included all of the previous features plus APOE e4 status. Plasma Aß42/Aß40<0.1013 and CSF Aß42/Aß40<0.0673 were used to detect PET‐amyloid positivity. To compare the performance of the ANN models and the CSF and plasma biomarkers, area under the receiver operating curve (AUC) from 5‐fold cross‐validation were calculated.ResultIndividuals who were PET‐amyloid positive (n = 46) were more likely to be older (p<.05) and carry an APOE e4 allele (p<.0001). The two groups did not differ in sex or education. In predicting PET‐amyloid positivity, the five most important driving features were the number of left turns per mile, 80th percentile of jerk, minimum vehicle speed, speed variability, and the number of speeding incidents. The CSF biomarker achieved the highest performance (AUC = 0.96±0.04), followed by the model with driving, age, education and APOE e4 status (AUC = 0.88±0.05). The model with only driving features achieved lower performance (AUC = 0.74±0.07), while the non‐intrusive ANN model achieved comparable performance to the plasma biomarker (AUC: 0.82±0.11vs.0.83±0.12).ConclusionDigital markers of everyday driving along with age and education level offer a non‐intrusive and scalable diagnostic tool for preclinical AD that provides ongoing assessments.