All the sessions from Transform 2021 are available on-demand now. Watch now.
Mild cognitive impairment — which affects an estimated 15% to 20% of people over the age of 65, according to the Alzheimer’s Association — often progresses quickly and quietly. Those with it are at an increased risk of developing a dementia, making early detection critical.
Researchers at Apple and the University of Tübingen believe the key might lie in iOS app usage habits. To this end, they propose a machine learning approach to reveal differences in patterns between users with and without cognitive impairment. It achieves an area under the receiver operating characteristics of 0.79, they report, indicating that it’s able to correctly spot symptomatic subjects about 80% of the time.
“The ubiquity of smartphone usage in many people’s lives make it a rich source of information about a person’s mental and cognitive state,” wrote the research team in a preprint paper. “Here, we … investigate to what extent [app usage patterns] are informative about a person’s cognitive health.”
The scientists’ unsupervised model — which automatically discovers the representations needed to classify data — identifies different types of interactions from the sequences in which apps are used, which it passes along to a separate cognitive health prediction model. The entire framework is engineered to be interpretable, such that the strength of the relationships between apps (like Messages) and health becomes self-evident, and its diagnoses are strongly informed by the structure of a user’s behaviors.
The system first segments usage over the course of interaction sessions, so that apps opened from the moment a phone’s unlocked to when it’s locked are grouped together. To encode the apps’ similarity, the researchers trained an algorithm to project which app a person might launch given the previous and subsequent three apps used. Apps are represented with embeddings (i.e., numerically), which are averaged together to obtain a single representation of each session. Next, the representations are clustered to identify different session types, and a user’s app usage is represented by a series of data points indexed and summarized by session time and types. This serves as input for the cognitive health predictor.
The team conducted a study involving 113 older adults — 31 with diagnosed cognitive impairment and 82 without — who contributed 12 weeks of phone usage data. They analyzed the four session types with the highest contribution to the model decision, and then they visualized the differences between apps in each session and the overall distribution for the 15 most common apps.
Interestingly, the researchers found that the session types most strongly associated with a high symptomatic score were dominated by pairs of apps: Call and Phone, Messages and Mail, and Mail and Safari, followed by Clock and Calendar. On the other hand, types corresponding to a low score were topped by Messages, Safari, Mail, and Facebook. Overall, the sessions with high Messages and Mail or Mail and Safari usage strongly increased the model’s predicted score for symptomatic, whereas session types with a lot of Messages or Safari sessions or Mail and Facebook sessions decreased it.
The researchers say that for subjects with a high symptomatic score, apps like Phone, Calendar, and Clock contributed measurably. That’s as opposed to apps such as Messages, Instagram, and Camera for those with a low score. But the impact of apps such as Messages or Mail appears to depend on the surrounding apps in the session. For instance, when Messages shares a session with Mail or Safari, it strongly increases the predicted score, while when Messages is alone or in a session with Facebook or Instagram it decreases the predicted score.
The researchers concede that their work has several potential limitations, chief among them the small sample size and the fact that symptomatic subjects were already diagnosed when entering the study. Nevertheless, they say that they’ve proven app usage alone can capture systematic differences between healthy and symptomatic subjects, and they intend to build on the work by incorporating the order of apps in each session, time of day, motion state, and other context.
It’s worth noting that it’s not the first time Apple has leveraged AI to predict the health of a user from their behavior alone. Engineers at the Cupertino company conducted a study involving more than 2,500 people to train an algorithm for detecting falls, which it incorporated into the Apple Watch Series 4 and Series 5.
VentureBeatVentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative technology and transact. Our site delivers essential information on data technologies and strategies to guide you as you lead your organizations. We invite you to become a member of our community, to access:
- up-to-date information on the subjects of interest to you
- our newsletters
- gated thought-leader content and discounted access to our prized events, such as Transform 2021: Learn More
- networking features, and more