Personal Devices May Aid in MCI and Alzheimer Dementia Identification

August 9, 2019

Initial data from a feasibility study conducted with Apple devices and digital apps has shown the potential for these to be able to differentiate people with mild cognitive impairment and mild Alzheimer disease dementia.

Nikki Marinsek, PhD

Early results from a feasibility study have suggested that the use of an iPhone, Apple Watch, iPad, and the Beddit sleep monitoring device, in combination with digital apps, may be able to differentiate people with mild cognitive impairment (MCI) and mild Alzheimer disease dementia.1

The initial findings were presented at the 25th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD). The study is being conducted by a joint effort from Eli Lilly, Evidation Health, and Apple Inc. Using Area Under the Receiver Operating Characteristic curve (AUROC), averaged across splits in a cohort of 84 healthy controls, 28 patients with MCI, and 7 with Alzheimer dementia, the investigators presented a model achieving AUROC = 0.80 using device-derived features and demographic data.

"Over the past few years, we've seen how data and insights derived from wearables and mobile consumer devices have enabled people living with health conditions, along with their clinicians, to better monitor their health," Nikki Marinsek, PhD, first author, and data scientist, Evidation Health, said in a statement.2 "We know that insights from smart devices and digital applications can lead to improved health outcomes, but we don't yet know how those resources can be used to identify and accelerate diagnoses. The results of the trial set the groundwork for future research that may be able to help identify people with neurodegenerative conditions earlier than ever before."

Marinsek and colleagues wrote, that to put their findings in context, a previous study using solely actigraphy data produced an AUROC = 0.62, and while other digital evaluations have been tested in Alzheimer discrimination, none others have been able to assess cognitive status from multiple sensors. They noted that future smart devices might be able to monitor patient symptoms post-MCI or Alzheimer diagnosis, detect MCI vulnerability, and test therapy effectiveness, among other capabilities. “However, additional research and validation are needed before these applications become a reality. Privacy is of particular importance in any clinical application,” they wrote.

The study was conducted over 12 weeks, with the patients using the Apple devices in real-world settings and collected 16 terabytes of data ranging from passively derived sensor data, to questionnaires, to simple assessment activities through the Digital Assessment App, which included psychomotor, reading, and typing tasks.

Ultimately, the devices showed that using the Shapley Additive Explanations (SHAP) approach, 5 of the 20 top features identified were deemed “important” in identifying those who were symptomatic with Alzheimer dementia. They were: slower typing (which has been noted by prior literature), less regularity and later first steps, receipt and sending of fewer text messages, greater reliance on apps deemed “helper apps”, and poorer compliance to study surveys.

"With further study, we may be able to screen people at high risk or detect dementia and Alzheimer's earlier with the devices we use in our everyday lives," said Christine Lemke, co-founder, and president, Evidation Health, in a statement. "These early findings suggest the potential of novel digital measures that are based on data generated and controlled by individuals."

The authors did identify limitations in the work. First, the patterns found in the data were linked to modifiable behaviors, and thus altered behavior not associated with the progression of underlying disease must be accounted for in future work. Second, a passive measuring of cognitive performance may be “self-reinforcing” without knowledge of preventive actions taken. Finally, the group acknowledged the potential risks of creating automated decision-making tools that are trained with data whose distribution may not best represent the target population.

“These results are a starting point, and more accurate classification may be possible with longer longitudinal data, larger cohort sizes, and other advances in passive data collection. Among the next steps in the analysis of this dataset specifically, are more in-depth explorations of accelerometer, audio, and video data,” Marinsek and colleagues concluded.


1. Chen R, Jankovic F, et al. Developing Measures of Cognitive Impairment in the Real World from Consumer-Grade Multimodal Sensor Streams. Presented at: 25th ACM SIGKDD Conference on Knowledge Discovery and Data Mining; August 4—8, 2019; Anchorage, AK. doi: 10.1145/3292500.3330690.

2. Lilly, Evidation Health and Apple Study Shows Personal Digital Devices May Help in the Identification of Mild Cognitive Impairment and Mild Alzheimer's Disease Dementia [press release]. Indianapolis, IN: Eli Lilly and Co; Published August 8, 2019. Accessed August 9, 2019.