Shifting From Active to Passive Monitoring of Alzheimer Disease: The State of the Research

ABSTRACT

Most research using digital technologies builds on existing methods for staff‐administered evaluation, requiring a large investment of time, effort, and resources. Widespread use of personal mobile devices provides opportunities for continuous health monitoring without active participant engagement. Home‐based sensors show promise in evaluating behavioral features in near real time. Digital technologies across these methodologies can detect precise measures of cognition, mood, sleep, gait, speech, motor activity, behavior patterns, and additional features relevant to health. As a neurodegenerative condition with insidious onset, Alzheimer disease and other dementias (AD/D) represent a key target for advances in monitoring disease symptoms. Studies to date evaluating the predictive power of digital measures use inconsistent approaches to characterize these measures. Comparison between different digital collection methods supports the use of passive collection methods in settings in which active participant engagement approaches are not feasible. Additional studies that analyze how digital measures across multiple data streams can together improve prediction of cognitive impairment and early‐stage AD are needed. Given the long timeline of progression from normal to diagnosis, digital monitoring will more easily make extended longitudinal follow‐up possible. Through the American Heart Association–funded Strategically Focused Research Network, the Boston University investigative team deployed a platform involving a wide range of technologies to address these gaps in research practice. Much more research is needed to thoroughly evaluate limitations of passive monitoring. Multidisciplinary collaborations are needed to establish legal and ethical frameworks for ensuring passive monitoring can be conducted at scale while protecting privacy and security, especially in vulnerable populations.

Digital health is a growing field and spans a broad variety of applications using novel technology and methodology for early symptom or disease detection, monitoring, and treatment. The Food and Drug Administration characterizes digital health as “mobile health (mHealth), health information technology (IT), wearable devices, telehealth and telemedicine, and personalized medicines.” ¹ The first mobile medical application with Food and Drug Administration clearance was introduced in 1997. ² Over the past 20 years, there has been an exponential increase in the number of digital health applications in the market, with most Food and Drug Administration–cleared applications geared toward cardiovascular health. ² In 2020, market data showed that $24.93 billion was invested in the digital health space, and this is projected to increase to over $310 billion in 2028. ³ While there has been extensive investment in the development of digital health tools, large‐scale, global research will be needed to determine which tools are most effective for various health domains and disease pathologies.

The potential of digital technologies, however, is constrained by the dominant use of active engagement technologies that allow increases in measurement frequency but still result in relatively sporadic data because they require conscious use of the application/device. To fully realize the potential of digital technologies there must be a shift to predominantly focus on passive monitoring technologies. Passive monitoring entails data capture through embedded or ambient sensors in daily activities that generate continuous data streams that characterize an individual's health. Existing evidence of the value of passive monitoring has come largely in the context of mental health interventions. They have included the unobtrusive capture of accelerometer data, incoming and outgoing call records, and text timestamps to evaluate major depressive disorder, bipolar disorder, and social anxiety severity. ⁴ , ⁵ A 2020 review evaluated the use of passive digital phenotyping, defined as use of data from personal digital devices to measure behavior, in college students. The review evaluated 25 studies with an average sample of 81 participants and identified location through global positioning system, accelerometer, and social information captured through the smartphone as critical mental health measures that could be captured passively. ⁶ While the reviewed studies established the feasibility of passive digital phenotyping and the opportunity for capture of relevant sleep, activity, and social interaction measures, the limited length of follow‐up and the lack of uniformity in data collection methods across studies attenuated the clinical relevance of reported findings.

Digital health realized through passive monitoring demonstrates especially large potential in the evaluation of individuals at risk for dementia or who have a neurodegenerative condition. Alzheimer disease (AD) is the most common neurodegenerative condition, with global prevalence of 50 million individuals. ⁷ Age is closely linked with AD so the global shifts in life span in the coming decades are expected to dramatically increase the prevalence of dementia. ⁸ , ⁹ AD global incidence increased 147% between 1990 and 2019. ¹⁰ Now viewed as a life‐course disease in which the insidious onset process can take up to decades to progress to diagnosed disease, digital assessments in general allow for more frequent data collection with more detailed data metrics, ultimately providing an opportunity to detect cognitive changes and associated behavioral indicators of AD with greater sensitivity and reliability. ¹¹ Passive engagement technologies will make monitoring and detection of AD‐related symptoms much more feasible across decades of times. Strengths of digital assessments and the acceptability of digital technology to older populations have been established in previous reviews. ¹¹ , ¹² , ¹³ , ¹⁴

This review encompasses 2 distinct elements grounded in the experience of the Boston University (BU) American Heart Association (AHA) Strategically Focused Research Network center on Health Technology and Innovation. As an AHA Strategically Focused Research Network center, BU has become a hub for collaboration in the collection, storage, and analysis of digital data, as well as the identification of novel digital tools. The first element of this review is a description of the research literature regarding passive monitoring in AD, as well as a consideration of gaps in current research. The second element of this review is a description of the data collection processes and selected technologies for use at the BU AHA center. Experiences in data collection, data processing considerations, and selection of collection modalities are presented, as well as limitations that require further research to ensure passive monitoring can be conducted at scale.

Moving From Active Engagement to Passive Monitoring

Passive monitoring relies on different sensors embedded in devices including smartphones, wearables, and in‐home instruments. The digital data streams produced from various sensors have been derived into measures of AD‐related behaviors, such as sleep, cognitive function, gait/balance, and so on. Thus, use of the embedded sensors and software can be used to monitor and detect relevant AD behaviors, from ambient sensors located both within and outside of the home. The distinction between passive and active monitoring falls along a spectrum. Smartphone applications require downloading, installation, and often a set of initiation steps. Wearable devices are frequently characterized as a passive instrument, even though they require a sustained connection to Bluetooth, the participant to upload and sync the app to the device, and to charge the device periodically. While these devices make possible high‐powered, high‐fidelity continuous data capture; the requirements for initiating use of even passive engagement technologies may be overly complicated for some older adults, especially individuals with cognitive impairment. ¹⁵ Figure 1 describes a framework for understanding the range of active to passive engagement activities. The data collection methods included in the bottom of the figure describe a series of tasks that span the array of digital health monitoring tools, which might be used in research or clinical practice. The terms are described in Table 1.

Figure 1. Characteristics of active and passive engagement technologies and related data collection methods.

Table 1.

Definition of Data Collection Approaches Spanning Active to Passive Digital Monitoring

Method	Definition	Examples
Staff‐administered in‐clinic testing	Evaluations that require a participant or patient to be physically present and engaged with a test administrator	Neuropsychological testing Timed up‐and‐go gait measurement
Staff‐administered remote testing	Evaluations that do not require a participant or patient to be physically present but require a dedicated time to meet with a test administrator	Neuropsychological testing over a remote platform Phone‐administered questionnaires/cognitive tests
Self‐administered remote testing	Evaluations conducted at home or other locations but require a participant to actively take time to complete a set of tests	Computer‐administered questionnaires Smartphone‐based cognitive and voice‐based tasks
Wearable data collection	Evaluations that entail some additional hardware in use to capture data and can be completed without the participant being physically present	Accelerometer worn during daily activities Pulse oximeter worn during sleep
In‐home sensor collection	Data collection requiring no participant interaction, with unobtrusive hardware placed in the home capturing data without any physical staff present for observation	Wall‐based infrared sensor system
Personal device sensor collection	Data collection requiring no participant interaction, with existing hardware	Digital phenotyping capture of smartphone accelerometer, gyroscope, call logging

The top of the table represents the most active approaches, while the bottom represents the most passive approaches.

A literature search was conducted to assess how active and passive engagement technologies have been used to detect and monitor behaviors indicative of AD. Previous reviews describe the potential of passive monitoring for of AD and other dementias (AD/D) symptoms. Kourtis et al ¹² published one of the first papers to provide a comprehensive overview of how various AD characteristics could be measured with personal computing sensors. The following section provides a summary description of what existing digital monitoring work has been done in AD/D research. This review does not represent a comprehensive review of digital technologies to evaluate AD/D‐related behaviors. The primary goal of this review was to identify and highlight studies characterizing how different health measures relevant to AD/D have been measured by digital technologies.

Active and Passive AD Monitoring in Research Practice

Table 2 references various studies that collectively represent the breadth of use of active to passive engagement specific to the context of AD/D. This table is not meant to comprehensively characterize all such studies, but rather provides exemplars within health domains of interest across the active–passive engagement technology spectrum. Health measures were selected on the basis of the potential for measurement by a passive monitoring modality (in‐home sensor or personal device sensor collection). The measures listed are not inclusive of all measures relevant to AD/D but instead represent areas that have evidence of demonstrated promise for fully passive monitoring. Wearable monitoring of vascular changes and eye scanning are examples of further digitally captured measures relevant to AD, ⁵⁴ , ⁵⁵ but the use of passive monitoring to collect these measures in relation to AD/D‐related outcomes has been limited. Rapid development of passive in‐home sensors, however, is expected to generate the potential for robust heart rate and respiration collection over time with minimal engagement. ⁵⁴ , ⁵⁶ , ⁵⁷ The set of digital measures accessible through passive monitoring is expected to dramatically increase over time as this research field grows and technology continues to advance.

Table 2.

Reference Studies for Measurement of Health Domains for Neurodegenerative Disease Measurement Methodology

	Methodology
	Staff‐administered in‐clinic testing	Home‐based active assessments	Wearable monitoring	In‐home sensor collection	Personal device sensor collection
Gait	Zhou et al, 2022 16 Huang et al, 2022 17	Serra‐Añó et al, 2019 18	Gietzelt et al, 2013 19	Dodge et al, 2012 20 Kaye et al, 2012 21 Hagler et al, 2010 22
Speech	Yamada et al, 2022 23	Atreya et al 2019 24 Fristed et al, 2022 25			Yamada et al, 2020 26 Al Hameed et al, 2019 27
Sleep	Weihs et al, 2022 28		Thomas et al 2021 29 Khosroazad et al, 2023 30 Mantua et al, 2016 31	Suzuki et al 2007 32 Khosroazad et al, 2023 30 Hsu et al, 2017 33 Au‐Yeung et al, 2022 34
Physical activity	Liang et al, 2010 35		VandeBunte et al, 2022 36 Bangen et al, 2023 37		Suzuki and Murase, 2010 38
Social engagement	Holtzman et al 2004 39 Bennet et al 2006 40			Suzuki et al 2007 32 Hsu et al, 2019 41 Peterson et al, 2012 42
Behavior patterns			Fleming et al, 2021 43	Reynolds et al, 2022 44 Suzuki et al, 2007 32 Wu et al, 2021 45 Hayes et al, 2004 46	Bernstein et al 2021 47 Eby et al, 2012 48 Seelye et al, 2018 49 Kaye et al, 2014 50 Babulal et al, 2021 51
Fine motor activity	Holmes et al, 2022 52				Austin et al 2012 53

A descriptive overview and comparison of studies is provided below.

Gait

Gait speed, dual‐task speed decline, and gait variability have been associated with changes in cognition that may represent early changes in dementia, particularly in non‐AD dementias where a strong vascular contribution to cognitive impairment, vascular dementia, or dementia with Lewy bodies is suspected. ⁵⁸ These measures have largely depended upon devices that require active engagement collected by staff‐administered walking assessments in research environments. ¹⁶ In‐clinic walking assessments can generate measures, including gait variability, stride length, and gait speed, which have been shown in multiple studies to be associated with risk of cognitive decline and dementia. ¹⁶ , ¹⁷ , ⁵⁹ However, real‐world walking data using research‐grade wearable devices allow for possibility of gait and functional mobility assessments in home settings and show clear associations with mild cognitive impairment (MCI) and dementia. ¹⁸ , ¹⁹ , ⁶⁰ Use of personal consumer devices, such as smartphones for gait measurement, has been explored, although research on gait measures to evaluate or distinguish cognitive status or predict AD has been limited. Smartphone‐based evaluation has been validated as accurate in comparison to wearable sensor systems and instrumented walkways in laboratory settings in people with Parkinson disease (PD) ⁶¹ and healthy individuals. ⁶² Stand‐alone smartphone evaluations have also been conducted in PD ⁶³ and multiple sclerosis. ⁶⁴ , ⁶⁵ There has been limited real‐world application of personal devices for gait evaluation, especially for comparisons of smartphone‐derived metrics to distinguish between healthy subjects and those with AD/D. However, comparisons between different neurodegenerative diseases could be further complicated by the fact that copathologies may occur.

Passive monitoring of gait is feasible through either in‐home monitoring or continuous capture of embedded smartphone sensor data. Home‐based infrared sensors are the primary method for passive monitoring and have been used to develop distinct trajectories of weekly walking speed and walking speed variability for early stages of AD. ²⁰ Home‐based passive sensing, such as the pioneering work of the Oregon Center for Aging and Technology group has established the presence of discriminative measures of gait in home settings. ¹⁵ Use of embedded smartphone sensors for evaluation of AD/D‐related changes in gait is an area for further exploration.

Speech and Voice

Recorded verbal responses can be used to conduct analysis of both linguistic features (number of words, length of sentences) of speech, and acoustic features (tone, pitch) of voice. A 2021 review of automatic speech and voice analyses for AD and MCI identified 24 diagnostic studies, with most studies establishing a diagnostic accuracy >88% for AD, and >80% for MCI. ⁶⁶ , ⁶⁷ Among the studies reviewed, almost all involved some active engagement speech elicitation paradigm such as reading activities, story recall, picture description, or sentence repetition tasks. In‐clinic administration of speech collection has been fully automated through tablet‐based sensing to reduce the burden on research staff ²³ ; however, this approach continues to place travel and time burden on study participants. The use of smartphone‐administered voice collection reduces these burdens. ²⁴ Structured smartphone assessments have been used to yield diagnostic accuracy of AD, as well as prediction of AD‐relevant amyloid beta biomarker status within clinical subsamples, ²⁵ but the active engagement required to open and use a structured assessment at some cadence may make longitudinal follow‐up more difficult.

Speech and voice measures represent a clear target for passive data collection given the universality of speech production and the accessibility of microphones through hand‐held voice recorders, smartphones, computers, and home personal assistant devices. Two studies with substantially different approaches to the capture of voice data through existing activities demonstrate the opportunities for passive speech collection. One approach captured conversational voice data generated through a regular monitoring phone call service and found atypical word and topic repetitions to be detectable as an indicator of AD. ²⁶ Another approach captured voice data during conversations with a neurologist at a patient visit and used pure acoustic features to differentiate between patients with progressive neurodegenerative disease and patients with subjective memory complaint with 96.2% accuracy. ²⁷ Both approaches involve data collection through unstructured conversations that would have occurred as part of individuals' existing activities. There is strong support for the potential of AD detection using voice and speech data, and further passive engagement collection could expand opportunities for analysis of unique characteristics of unstructured speech while also improving scalability.

Sleep

Sleep has been studied as a biomarker for AD for many years. Reduced rapid eye movement sleep, increasing prevalence of sleep disorders, and more have all been linked to early stages of MCI. ⁶⁸ Research has previously focused largely on in‐clinic sleep studies (polysomnography, actigraphy) and questionnaire‐based self‐report measures. ⁶⁹ Wearable sensors have enabled home‐based monitoring of sleep, which has made much more feasible measurements from geriatric populations. Studies show that wearable devices show a significant correlation between total sleep time, sleep efficiency, and rapid eye movement sleep detection when compared with a more traditional in‐clinic polysomnography. ³¹ Wearable‐derived sleep measures of sleep quality oxygenation have been associated with white matter, a pathological measure of AD/D. ²⁹

The Maine Sleep and Aging Study explored how multiple sleep measures could be used together through collection of actigraphy, a wireless mattress sensor, and self‐report questionnaires in participants with normal cognition and MCI. Latency between movement arousals and coupled respiratory upregulation derived from the sleep monitoring produced >85% sensitivity and specificity. ³⁰ As with other measures, collecting sleep data at home provides opportunities for additional evaluation of variability and change over time at a minimal cost and engagement. At‐home passive sleep monitoring via infrared sensors has suggested that tracking sleep through movement and sleep time could be an effective method for detecting dementia. ³² Ambient home sensors have been able to provide evidence of the opportunity for passive monitoring of sleep‐related metrics. ³³ , ³⁴

Physical Activity

Self‐report questionnaires are typically used to evaluate physical activity. Questionnaire‐derived exercise measures have been associated with AD biomarkers, with elevated Pittsburgh compound‐B, tau, and phosphorylated tau 181 being linked to less exercise. ³⁵ Moving further along the active to passive continuum, wearables provide objective and more accurate measures of daily activity and the opportunity for evaluating new measures of activity. Comparisons of self‐report and wearable measurement of activity and older adults demonstrate that wearables have greater construct validity and are less subject to biases in self‐report measures such as a greater likelihood of overreporting among those with worse memory performance. ³⁶ Waist accelerometer–derived moderate to vigorous physical activity levels have been associated with better memory and executive functioning.

Infrared sensing–based activity monitoring demonstrated that individuals characterized as having cognitive decline had reduced indoor movement over a 1‐year study period, whereas individuals without decline maintained similar baseline movement levels. ³⁸ As with other modifiable risk factors, temporality for the relationship between reduced activity and cognitive decline is complex; however, more passive continuous monitoring provides opportunities for novel approaches.

Social Engagement

Research exploring the size and complexity of social networks and AD/D has typically relied on survey measures for evaluating the frequency and strength of social contacts. ³⁹ Social network characteristics evaluated by structured interview have been linked with an attenuation of the relationship between brain pathology and cognitive effects, ⁴⁰ , ⁷⁰ further strengthening the evidence of the influence of social networks on clinical neurodegenerative disease manifestation. As is the case with several of the other domains explored here (eg, activity, sleep), the temporality of the connection with disease is a complex area to evaluate. This is further influenced by reliance on self‐report measures for characterizing social engagement that can be expected to shift substantially throughout the life course.

Digital tools are capable of accurate capture of many characteristics of social engagement, including the timing and duration of telephone contacts, the frequency of activities, or simply expanded self‐report collection through ecological momentary assessment. In‐home sensors can be used to identify the extent to which individuals move about their homes and leave their homes. Early work using infrared sensors in home demonstrated that individuals with impairment made fewer outings from the home. ³²

Fine Motor Activity

Digital measures of fine motor control have frequently focused on typing patterns. Several studies have evaluated typing in structured clinical assessments through specialized keyboards, resulting in the identification of correlations between typing‐derived measures and traditional cognitive testing results. ⁵² Analysis of typing data on personal computers has been directly compared with the finger‐tapping test typically used to assess AD‐related motor symptoms in clinical neuropsychology settings and demonstrated significant correlations. ⁵³

Behavior Patterns

Friends and family members of those with AD often can recall AD‐related behavioral changes that occurred years before symptoms reach clinical thresholds for diagnosis. AD‐related behavioral changes are particularly difficult to detect in the earliest stages of the disease because of moment‐to‐moment fluctuations in their manifestation. Normal cognitive performance and function levels are constantly changing in response to a myriad of factors; thus, accurate detection of a change due to an insidiously neurodegenerative process divergent from the normal heterogeneous spectrum is challenging. This detection is substantially more difficult given the common practice of sporadic assessments, which further lends itself to inconsistent measurements. Passive digital monitoring offers solutions for capturing daily behavior and changes in behavior that are inaccessible via other methodologies and mapped over time and can then differentiate between normal fluctuations versus ones that are pathologically driven. Capturing and characterizing such behavior has been a major focus of studies using in‐home remote sensor systems. The Oregon Center for Aging and Technology laboratory has established the relationship of measures of mobility stability and frequency with gait speed and cognitive impairment. ⁴⁵ MCI status has also been associated with an attenuation of seasonal variability of time spent in the bedroom overnight as measured by in‐home sensors. ⁴⁴ MCI diagnosis has also been linked with distinct patterns of kitchen and bathroom usage in the Oregon Center for Aging and Technology study. ⁷¹ Data collected from existing personal devices with installed software have also revealed significant differences in the duration, timing, and number of computer sessions between older adults with and without cognitive impairment. ⁴⁷

Another area of research has focused on using passive driving monitoring via global positioning system and driving sensors to assess behavioral patterns in older adults. These technologies offer the benefit of having the ability to collect accurate, real‐time data on a large cohort of individuals. Comparisons between early‐stage dementia groups and control groups have shown that early‐stage dementia groups have a significantly smaller driving space when compared with the control group and are much more likely to get lost sleep, physical activity, and respiration. ⁴⁸ Machine learning approaches have also been applied to detect unique signatures of preclinical AD and global positioning system driving data. ⁵¹

Summary

Digital technologies offer opportunities to detect behavior signatures and changes across a range of AD‐related symptoms, as demonstrated by the studies highlighted here. Both active and passive engagement monitoring strategies demonstrate promise in advancing abilities to detect AD earlier in disease progression without relying on highly costly and invasive approaches. The measures listed above such as gait, speech, sleep, and so on are not unique to AD/D. Indeed, both active and passive digital monitoring approaches provide opportunities for advanced detection and monitoring in the evaluation of non‐AD/D forms of neurodegeneration and those associated with vascular dementias as well. This review's focus is primarily driven by the spotlight on AD/D in existing digital research studies and the use of these technologies in research cohorts that primarily focus on AD/D‐related end points. Digital capture of motor activity and gait measures for earlier detection of PD has been an area of substantial research progress. ⁷² Speech analysis has been used to identify distinctive speech patterns in those at risk for vascular dementia. ⁷³ The measures listed in Table 2 have been implicated as clinical features differentiating healthy controls from individuals experiencing a range of prominent neurodegenerative diseases. ⁷⁴ Further research is needed outside of AD/D‐specific settings to identify how these measures could be used to differentiate between AD/D and other common neurodegenerative diseases, as passive collection of granular digital data over extended follow‐up is expected to allow for analytics that could improve the sensitivity of diagnoses. Evaluating a specific disease may be complex due to the contribution of multiple and varied disease causes in many cases. Digital markers may thus provide the most promise in identifying individualized disease trajectories. ⁷⁵ Addressing the shortcomings of existing digital monitoring research practice will be essential to moving forward research toward its full potential.

Gaps in Current Research Practice

The use of digital technology for AD/D research is still at a nascent stage. As a result, there are notable research gaps. Most digital health studies examine 1 methodology or measure at a given time. Studies typically focus on a specific metric, such as gait or cognition, and are administered via 1 approach (eg, staff‐administered in‐clinic evaluation, self‐administered, smart home device). Isolated methodologies result in single modality data that are not sufficient for accurate monitoring of complex, fluctuating AD‐related behaviors, and result in a lack of sufficient specificity and sensitivity needed for wide adoption in research and clinical settings. The current bias in digital data collection is leading to biased results that are placing further constraints in unleashing the real potential of digital. For example, in the context of gait evaluation, initial findings suggest that differences between disease states may be more easily detectable in clinical settings than through real‐world data collection. ⁷⁶ Given the complex presentation of AD/D, further evaluation of how measures compare across modalities is needed. Isolated investigation of a single domain prevents examination of how multiple AD/D‐related health measures may be evaluated together to develop more precise disease detection and monitoring. Simultaneous evaluation of drawing, speech, and gait within a clinical population found classification accuracy consistently and substantially improved with the addition of features from the multiple modalities. ⁷⁷

Further research including longitudinal follow‐up of several years is needed. The RADAR‐AD (Remote Assessment of Disease and Relapse–Alzheimer's Disease) study has expanded the modalities across a slate of remote monitoring technologies including walking tests, active engagement decision‐making tasks, activity tracking, and a wearable camera. Passive monitoring applications in the RADAR‐AD study include capture of global positioning system coordinates, phone usage, and additional sensor data, but data collection for the protocol spanned an 8‐week period. ⁷⁸ A 2019 review of real‐life, home‐based digital monitoring studies found that study periods ranged from 3.5 days to 3.6 years. ⁷⁹ Opportunities for low‐touch, long‐term engagement are a key to highlighting the value proposition for passive monitoring technologies, especially software that runs on existing participant‐owned technologies. The Oregon Center for Aging and Technology Platform developed by Kaye et al demonstrates how sensor‐based technologies can be deployed longitudinally in real‐world, uncontrolled settings. ⁸⁰ The Life Laboratory deployed by Kaye et al had been used to collect naturalistic data capable of measuring each of the AD‐related behaviors described above for as long as 11 years as of 2018. ⁸⁰ Longitudinal follow‐up will be valuable for building out additional digitally derived measures of variability and change, as well as for an examination of the predictive power of digital data in years before clinical diagnosis.

Additional studies that explore multiple health measures and collection methodologies should also be explored. Distinct methodologies and measures can synergistically improve classification accuracy. One of the most notable benefits of digital measurement is the opportunity it provides for more inclusive global research practice given appropriate consideration of what is the maximally inclusive deployment. For example, in‐home sensors and wearable devices are infeasible for most global clinical and research contexts due to bandwidth concerns, hardware costs, scalability concerns, and the like. The availability of a multitude of well‐established methodologies that can be applied in specific contexts would provide opportunities for applications use cases in both high‐ and low‐resource settings.

Development of technology‐agnostic digital health measurement approaches can occur through multiple pathways. One approach would be to engage in multisensor, multimodal measurement within a specific population and to develop pipelines for combining the different data outputs together for unified analysis. This approach involves a higher degree of participant burden, high levels of staff involvement, and the availability of sufficient funding to provide multiple technologies to participants. A second approach could involve sharing of research data and research findings more broadly and transparently to the research community. The recent changes in the National Institutes of Health's data‐sharing policy aim to advance this goal of widespread data sharing. Through sharing data from heterogeneous populations using distinct sensors and methodologies, there may be concerns about opportunities for any sensible data analysis. Expanded digital analyses could be enabled through further research using raw digital signals generated from sensor‐based collection or other digital recordings. Recording and sharing raw digital signals increases the opportunities for comparisons across collection methodologies with disparate hardware specifications. In the case of passive monitoring in particular, the focus on tying stagnant derived measures to traditional cognitive tests substantially reduces the data resource. Comparison of raw signal, as opposed to downscaled derivations, might lead to improved frameworks for implementing technology‐agnostic measurement that is necessary for continued progress.

Bridging the Gap Between Active and Passive

The BU AHA Strategically Focused Research Network Center is engaged in multiple projects with novel digital data collection, curation, and analysis. The data collection protocols for the feasibility and validation projects were approved by the BU Medical Campus Institutional Review Board, and all participants provided written informed consent. Detailed descriptions of all technologies and collection protocols are described elsewhere. ⁸¹ The multimodal, multiengagement data collection at the BU Alzheimer's Disease Research Center (ADRC) are conducted in tandem with data analysis and sharing initiatives. Analysis goals will include developing measures that encompass multiple technologies, multiple participant use schedules, and varying levels of task adherence and selection. The Center's data collection activities aim to develop a multimodal digital monitoring data set from participants with clinical data collected at the BU ADRC. ⁸² , ⁸³ These data will be shared widely with the global research community through the AHA Precision Medicine Platform and Alzheimer's Disease Data Initiative. Key lessons learned from the project include further understanding of the burden of data collection and management associated with active and passive monitoring digital technologies as currently constructed. Critically, those who engaged with passive typing–based technologies had the greatest adherence longitudinally, suggesting that additional passive‐sensing platforms are of great interest. ⁸¹

Figure 2 provides an overview of the range of collection approaches and AD/D‐related health measures administered through these projects. Within the figure, icons represent that a technology and activity are included in the platform that encompass a given modality and measure combination. Each icon does not represent an individual technology. With the platform including a wrist‐worn wearable capturing telemetry data, there will be opportunities to analyze gait, sleep, social engagement (mobility patterns), behavior patterns (rest–activity cycles), and motor activity (through wrist movement). The study offers passive engagement technologies alongside active engagement tools including smartphone applications, wearable devices, and in‐clinic sensors. Platform components are subject to change over time to maintain pace with the evolving digital collection market, so additional passive engagement tools may be integrated in the future.

Figure 2. Ongoing and planned* data collection measures for BU AHA SFRN studies.

*All sensors except for passive in‐home sensor are being actively used to collect research data. The in‐home sensor is a planned element of the protocol to be included upon institutional review and approval. BU AHA SFRN indicates Boston University American Heart Association Strategically Focused Research Network.

Limitations and Considerations for Scalable Monitoring

There are several concerns in digital data collection, multimodal data collection, and passive monitoring specifically, that warrant further research and attention. First and foremost are concerns about privacy and security. Security risks are expected to differ on the basis of characteristics such as socioeconomic status, education, and various societal factors. ⁸⁴ Smartphones provide the unique opportunity to collect data in even the lowest‐resource settings; however, not all smartphones are built the same. Differing hardware and software will lead to uneven distribution of security risks. Susceptibility to targeted attacks may also lead to individuals with less education and less technology literacy, including older adults, having greater vulnerability to security breaches.

Passive monitoring entails data collection in all moments of a person's activities. Collection of data such as geoposition can invoke concerns of intrusive monitoring for many participants, as may capturing keystroke data that include sensitive personal content. If commercial applications are used in research settings, these sensitive identifiable data elements may be stored on third‐party servers, and no server is unequivocally protected from breach. The low‐touch experience of contributing data through passive monitoring is taken to be a benefit for the reduced burden on participants or patients; however, the generation of data without conscious effort also generates further concern for privacy and autonomy. Without appropriate checks in place, a participant or patient could forget that they are involved in data collection, resulting in a potential infringement on autonomy through the collection of data without active consent. Furthermore, the insidious nature of AD/D may make individuals more vulnerable to forgetting that they are part of continuous data collection longitudinally. These potential concerns require further legal and ethical review before large‐scale passive monitoring could be applied in clinical settings. Feasibility studies have found that for a substantial majority of participants, privacy is not a notable concern. ⁸⁵ Further findings from qualitative research will reveal more information about user perspectives on continuous passive monitoring. Attention to perspectives from diverse participant bases are especially necessary, particularly given the history of structural mistreatment in human subject research that have been experienced by minoritized communities in the United States that might feel especially concerned about privacy and security.

Outside of recruitment for research trials, further concerns for autonomy arise from digital data collection by commercial digital technology companies. Features of voice, movement, social engagement, and a host of other behaviors are constantly captured by commercial applications. Technology users may endorse the collection of these measures through Terms of Service when first using a technology, although the ongoing discovery of how these measures could relate to health raises concerns about data security and user consent. Several issues are raised by the wealth of health‐related behavior metrics by large technology companies. How can the ethical concerns of using data that may be technically authorized but not actively consented for use by research participants be addressed? Could researchers request that these companies share the wealth of collected data for research analytic purposes? These questions cannot be easily answered and require interdisciplinary collaboration between legal, regulatory, and scientific research organizations at the local, national, and international levels. In current practice, obtaining active consent and implementing systems to periodically affirm participant consent for passive data collection and use represent the best approach to ensuring ethical research conduct.

While passive monitoring provides a strong opportunity for continuous collection of data with minimal associated staff and participant burden, nonetheless, there is still an effort load, and it may appear in unforeseen ways. Data collection oversight remains a necessary component of practice to ensure data are collected accurately and consistently across the sample. Current applications for home‐based sensors require a high degree of staff burden given the need to install devices in the home and monitor data streams. ⁸⁰ Even in the case of software installed on existing personal devices, turnover of personal computers and smartphones is common, and in many cases, participants may need to be recontacted to redownload applications, or to conduct periodic software updates. Staff must also engage with each participant at the outset of studies to assist in downloading relevant software and linking these data to some secure storage location. These factors much be considered when conducting large‐scale monitoring and infrastructure support to address these issues is still needed, but the extent of engagement remains significantly less burdensome than traditional in‐clinic instrumented measurement that requires travel and may involve multiple hours of participation from the research participant.

An additional perceived benefit of passive monitoring has been to shift from stimuli‐specific cognitive tests, which have been found to generate bias by cultural, socioeconomic status, and educational attainment. Through collection in activities of daily living, passive monitoring enables what should be culturally agnostic measurements of relevant AD‐related behaviors. This assumption has not been well tested. The collection of data during activities of daily living could also lead to measurement bias. Different occupations, living situations, and many other factors influence metrics accessible through passive monitoring such as rest–activity patterns, social engagement, unstructured conversation, and so on. ⁸⁶ Additional biases may be introduced through the difficulty in establishing “ground truth” diagnostic outcomes for model prediction with accepted diagnostic tools largely developed in high‐resource settings with research participants largely from European White ancestry. ⁸⁶ , ⁸⁷

Data processing and advance analytic expertise is another barrier to widespread use of passive sensing. The data captured by unobtrusive sensors may be streamed over 24‐hour periods, producing high volumes of data. Storage of these data can be expensive, and management and analysis of multimodal digital data streams is complex and currently largely dependent on highly specialized training. The constantly shifting digital technology landscape further complicates analyses given that a longitudinal study with years of follow‐up is expected to see software and firmware upgrades that yield substantial differences in data output. Prioritizing raw data capture alleviates some of these concerns, but changes in firmware could be expected to modify the comparability of data before derivation of composite measures. Maintaining documentation through the capture of metadata will be needed to track changes, but advanced analytic techniques capable of technology‐agnostic processing will be the most scalable solution for the heterogeneity in digital tools. Use of data in their native format provides the strongest support for further scientific discovery, but the derived data from digital technologies offer a clear benefit to care teams and patients. Composite measures of cognition and function may offer tangible indicators of decline or improvement, allowing for more targeted intervention points for lifestyle or pharmaceutical interventions. Any given raw data file can be used to generate an overabundance of composite features, and the selection of features most relevant for clinical use will need to be driven by collaboration between clinical and analytic experts. Clinical relevance and reliability for these end points will be driven by the robustness of raw data processing to heterogeneous and dynamic collection technologies. Given that passive monitoring methods are still emerging, there are limited tools that can be broadly used by the broader research community to analyze these data and an insufficient number of experts able to create these tools.

Conclusions

Research is accumulating on the association between various measures attainable via passive monitoring tools and symptoms of neurodegenerative disease. While the evidence of associations between passively captured behavioral measures and AD is encouraging, shifting from active engagement to passive monitoring remains an aspiration. Inertia in this movement toward passive engagement is not attributable to the available tools. Hardware, software, and advanced analytics continue to evolve and provide the foundational tools to enable a significant shift toward measurement that is less burdensome and more sensitive to changes in health. What remains is the dedicated time, energy, and funding support to realize the potential of tools that are currently applied well below their potential. As these resources are put into place, a more substantial shift can be expected. With this shift toward passive monitoring, subsequent shifts in our understanding and approach to AD will come. Further shifts are expected to follow across all chronic diseases, with more accurate and realistic monitoring providing opportunities to transition from lifelong treatment to sustainable prevention.

Sources of Funding

Funding support for this project comes from the American Heart Association (20SFRN35360180), Gates Ventures, and Alzheimer's Research UK. Additional funding support was provided by the National Institute on Aging (AG016495, AG062109, AG068221, AG072589, U01AG068221); the National Cancer Institute (R21‐CA253498); the National Heart, Lung, and Blood Institute (R01‐HL159620); the National Institute of Diabetes and Digestive and Kidney Diseases (R43‐DK134273); the Alzheimer's Drug Discovery Foundation (201902–2 017 835); and the Alzheimer's Association (AARG‐NTF‐20‐643 020).

Disclosures

Dr Au is a scientific advisor to Signant Health and NovoNordisk, and a consultant to Biogen and the Davos Alzheimer's Collaborative; she also serves as Director of the Global Cohort Development program for the Davos Alzheimer's Collaborative. The remaining authors have no disclosures to report.