SMART-A extends the open-source mindLAMP app developed by the Division of Digital Psychiatry at Beth Israel Deaconess Medical Center (BIDMC) [30]. MindLAMP is an adaptable platform for biomedical research. The platform is acceptable and feasible for use in older adults and patients with AD [36]. It includes a study portal, a smartphone app (available on both Apple and Android), a database, data analysis, and modeling tools [37]. With special research entitlements, mindLAMP2 can capture Apple HealthKit and Google Health Connect data directly from smartphones, including accelerometer, gyroscope, screen time, geolocation, ambient light, call/text logs, and phone charging time in a safe, ethical, and patient-centric manner (see the “Ethical Considerations” section), and transform this raw data into clinically meaningful constructs such as sleep, mobility, and behavioral patterns through custom pipelines [38,39]. In addition, mindLAMP supports survey or ecological momentary assessment data and currently contains several cognitive tests related to spatial span tasks and visual identification tasks, but also the ability to create new cognitive tasks.

In the Speech Recording task (Figure 5), participants will be given a prompt to respond to open-ended questions such as “describe what you had for breakfast this week.” Participants are asked to press the microphone to record their voice and verbally respond to the prompt for about a minute. The participant’s voice is recorded as an MP4 audio file.

During the Fragmented Letters task (Figure 6), the user is presented with letters of the alphabet written in black ink on a white background that have been slightly fragmented, causing some parts of the letters to blend in with the background [52]. Fragmentation is defined as the number of pixels in the letter that are white, out of the total number of pixels normally in the letter. In each round, the fragmentation percentage increases by 5%, and the game concludes after 10 trials. This game aims to assess both visuospatial and perceptual deficits.

In addition to cognitive games, study participants are asked to complete brief surveys through the mindLAMP app. Orientation questionnaires drawn from the Orientation sections of the MoCA [3] and Clinical Dementia Rating (CDR) [53] are administered during the delay phase of TicTacToe and the Visual Association Memory Test. Drop-down menus are provided on the smartphone for questions such as “What month is it?,” and participants are asked to select the best answer.

Furthermore, participants are asked to complete a brief daily sleep survey and a longer weekly sleep survey. The daily sleep survey consists of 4 questions, inquiring about sleeping and rising time, and how long participants were awake (if at all) during the night. The weekly sleep survey provides several Likert scales for participants to rate how their sleep throughout the week affected their daytime activities.

A unique facet of our approach is the use of smartphones to capture relevant and real-time behaviors via digital phenotyping. We will use the native features of mindLAMP to support digital phenotyping data collection to capture relevant information on behavioral domains pertinent to AD risk, including physical activity, sleep, and social interactions or isolation. While the assessment of these behaviors has traditionally been difficult, with smartphones, it is accessible and acceptable to patients with AD. Throughout the 6-month active data collection period, sensor data collection native to the mindLAMP 2 app [54] will also be activated, including GPS and accelerometer, communication events (anonymized call and text message logs), and phone and screen status. From these raw data streams, we will build relevant behavioral features in a transparent and reproducible manner that is not possible today with wearable data, where companies and the underlying algorithms can predefine features that can change at any time. The Cortex software package [55], developed as open-source software for use with mindLAMP data, will be used for initial behavioral feature creation.

We will recruit 15 patients evaluated by physicians in the Cognitive Neurology Unit of BIDMC with either MCI [56] or mild dementia [15] from AD. All participants with MCI or AD included in the study are diagnosed through neuropsychological testing with confirmatory biomarkers (eg, amyloid-positron emission tomography scan or p-tau 217). Fifteen cognitively normal controls, with no known history or diagnosis of cognitive impairment, will be recruited as well. Baseline evaluation will include the MoCA [3] and the CDR Global [53] score, as a staging measure for the MCI/AD group and a screening tool for the control group. Patients will run the SMART-A app on their smartphones for 6 months. During this period, we will capture passive data continuously and active data on a daily or weekly schedule (Table 1). As part of the Collective, data are entered into the a2 Collective Data Enclave, hosted by the Johns Hopkins Artificial Intelligence and Technology Collaboratory Data Integration and Quality Core. The a2 Collective aims to eventually archive the data in the National Archive of Computerized Data on Aging. All data sent to the Collective and entered into the Data Enclave will be completely deidentified and contain no protected health information. BIDMC and the University of Pennsylvania, the entity under which BIDMC was granted a subaward of the a2 Collective, will be sharing completely deidentified data with the NIH as well, per a previously agreed-upon protocol. Data will be available by request.

Over the course of 6 months, patients will complete each cognitive task weekly, a daily survey about sleep time and wake time, and a weekly survey about the quality of sleep. We anticipate collecting a total of 156 cognitive tasks, 180 daily sleep surveys, 26 sleep quality surveys, and 180 days of passive data over the course of the study from each participant. To optimize engagement, patients will be compensated a maximum of US $50 per month for engaging in cognitive tasks and surveys as well as for participating in the passive data collection. At the end of the 6-month study period, the baseline cognitive evaluations are repeated, and participants are given a survey (the mHealth App Usability Questionnaire [MAUQ]) on the usability of the app [57].

Participants with MCI or mild AD are offered optional continuation of the study for a maximum of 12 months. Our goal is for 5 to 10 participants from the AD group to continue this long-term data collection. When participants decide to end their participation, the final assessments completed at the 6-month end of study point (MoCA, CDR, and MAUQ) will be repeated.

The SMART-A study is designed as a longitudinal feasibility and proof-of-concept study of smartphone-based cognitive and behavioral monitoring in individuals with MCI or mild AD and cognitively normal controls. Given the modest sample size and the multimodal nature of the data collected, the analyses are organized into primary, secondary, and exploratory aims. This structure is intended to clearly distinguish feasibility-focused objectives from preliminary validity analyses and from more hypothesis-generating modeling work.

This study is designed as a feasibility and proof-of-concept study rather than a definitively powered validation trial. Because reliable estimates of adherence, within-subject variability, sensor completeness, and effect sizes for smartphone-derived cognitive and behavioral features in MCI and mild AD are not yet established [58], this study is intended to generate the empirical parameters needed for formal powering of a subsequent larger study. The target sample of 15 participants with MCI or mild AD and 15 cognitively normal controls is expected to provide reasonable precision for estimating feasibility metrics and preliminary effect sizes, while supporting exploratory analyses of large between-group differences and longitudinal trajectories. For repeated-measures analyses, power will be evaluated using simulation based on the planned mixed-effects modeling framework, incorporating expected adherence, dropout, and within-subject correlation. All inferential findings from this study will therefore be interpreted as preliminary and hypothesis-generating.

We have received approval from the BIDMC Committee for Clinical Investigations, our institutional review board, for the outlined project (#2024P000931). Participants experiencing cognitive issues will be patients of the principal investigator and coinvestigators as well as patients of physicians at the Cognitive Neurology Unit. No patients will be approached unless they have given assent to be contacted about this study. We will recruit only participants who are able to consent to the testing, such as those with mild dementia (CDR of 1.0 or less), as determined by clinician rating prior to the screening process. All participants will be advised that they can withdraw from participation in this investigation at any time and that their participation in this research is completely voluntary.

For participants with dementia, a clinician who is familiar to the individual will discuss the study with the participant and their family member, review any questions they may have, and ask if they are willing to participate. Participants with dementia with a CDR of 1.0 or lower will be asked to sign the consent form. Although as a very general guideline, participants with dementia or a related disorder with a CDR of 2.0 or better might be considered capable of consent for the main study, our standard practice of “double consent” has been endorsed by the Alzheimer’s Association and is consistent with the suggestions of experts in the field and adheres to the “Points to Consider” published by the NIH Office of Extramural Research. In this context, having the participant and caregiver sign the informed consent for the participant makes perfect sense: it ensures that both the participant and caregiver are comfortable with the participation of the participant with memory impairment. Having both signatures on the consent form has several additional advantages. Allowing the participant with memory impairment to sign for themself along with the caregiver preserves dignity and is consistent with the humane clinical practice of addressing the patient and not just the caregiver in medical discussions.

Data security is one of the most important features of any mobile health study, and we have taken several safeguards to ensure data security in this pilot study. All personally identifiable information will be stored in REDCap (Research Electronic Data Capture) behind the BIDMC firewall, with access restricted to members of the research team. Only members of the research team will have access to the list of participants.

Passive data can potentially contain personally identifiable information, including GPS data and communication logs. The app’s GPS data provide enough detail to identify individual buildings or street addresses with some degree of confidence, although a fair amount of analysis would be required to transform a series of GPS coordinates into a home address for a participant. As our analysis will only look at changes in mobility patterns, no raw GPS data or specific locations will ever be investigated or reported. Call and text log data are data that tell when a phone call was made, or a text message was sent and how long that message was. The app does not collect any content from the phone call, so no data are recorded about what is said or what is typed. The app does record what phone number is called but encrypts this so that no one can determine the actual phone number that was communicated with. The app also records when the participant turns on or off the phone’s screen, when the participant turns off or reboots the phone, and when the participant plugs in or unplugs the phone. These data provide information on a participant’s typical phone usage and do not record any sensitive or identifiable information, only time stamps.

The app never records or stores any personally identifiable information. Every participant is assigned a randomly generated participant ID. For a participant to use the app, a study administrator must create a participant ID, and the user must register it and create a password. A participant ID can only be connected to one phone at a time. To upload data, a phone must have a valid username, password, and phone ID number.

All data on phones, on the server, and in-transit use industry-standard encryption techniques. The phone stores data locally only until they can be automatically sent to the secure server. The phone uses asymmetric encryption, meaning that even the phone cannot read its own data; data recorded on the phone can only be read on the server. Once Wi-Fi is available, the phone deletes any temporarily stored data and then sends them to the server in an encrypted manner. During registration, the device is provided with the public half of a 2048-bit RSA encryption key. With this key, the device can encrypt data, but only the server, which has the private key, can decrypt it. The RSA key is then used to encrypt a symmetric AES key for bulk encryption. These keys are generated as needed by the app, are not stored, and must be decrypted by the server before any data can be recovered. Data received by the BIDMC server are then re-encrypted with a master key provided for that study.

The pilot study described in this protocol aims to affirm that it is feasible for our population of participants with MCI and AD to complete the testing paradigm, and to identify which active and passive biomarkers have the most potential to differentiate between the MCI/AD and control groups. Larger-scale studies may contain some changes to the SMART-A paradigm based on our feasibility findings from the pilot study but will aim to confirm the strongest combination of active and passive biomarkers for AD.

We expect to find differences between the AD group and controls in passive data measures (eg, sleep and movement-related), as patients with AD have been documented to have sleep, socialization, and physical activity deficits [59-61]. While intraindividual variability is expected in all participants, we anticipate those with cognitive impairments to experience more fluctuations in their scores [62], and to exhibit less of a learning effect on the cognitive tasks [63]. Given that memory deficits are considered a hallmark of AD [7], we hypothesize that our cognitive games centered around the memory domain, the Associative Memory Game and TicTacToe, will be the most sensitive in differentiating between our cognitively impaired and cognitively unimpaired groups.

While longitudinal data collection for the pilot study is ongoing, the integration of these novel tests onto the mindLAMP platform highlights how feasible and scalable a digital clinical diagnostic program for AD can be. The feasibility of recruitment and interest in the study, indicated by recruitment numbers being nearly met in 8 months, suggest that this paradigm of smartphone-based multimodal data collection is acceptable and feasible in a cognitively impaired population. While concerns around digital literacy or privacy are valid, our study and others [64-66] indicate that older adults are very capable of using devices and willing to share personal data from their smartphones with the appropriate research safeguards in place. Participants who have completed the 6-month-long trial report that the mindLAMP 2 app was easy and enjoyable to use, and the weekly schedule of tasks was manageable to complete.

While other studies have investigated the use of digital measures for screening and monitoring of AD [23-26], our SMART-A paradigm is unique in combining multiple data streams from a real-world setting to assess several areas of functional and cognitive deficits. Passive data streams have not shown promising results as digital biomarkers for AD on their own [26], but we anticipate that, in combination with cognitive testing and survey reports, we will obtain a stronger signal than from any one data stream alone. As noted from other studies using the mindLAMP app, paradigms combining passive and active data streams in other clinical populations [67-69] are feasible and can provide at least preliminary results. The SMART-A paradigm proposed in this protocol has the potential to expand the applications of the mindLAMP app and to strengthen the currently available collection of digital biomarkers for AD.

It is important to consider that this initial study is only a pilot, with a relatively low sample size of 15 MCI or AD and 15 controls, and a relatively homogeneous population. Most participants are highly educated (having earned a bachelor’s degree or higher), are generally comfortable with using technology, and are English speakers from the United States. Further work testing the feasibility and preliminary usefulness of the SMART-A paradigm in other groups will be needed to ensure its application to all groups. In addition, the sample size will need to be much larger to determine whether each digital biomarker can detect or predict cognitive impairment, instead of just a preliminary association. However, the results from the pilot study will inform future iterations of the SMART-A paradigm in larger-scale studies. Recruitment for this pilot study was limited to those with mild cognitive difficulties; while this was done to ensure participants had the capacity to understand the instructions of the cognitive games, further research will be needed to validate these tests as cognitive biomarkers in more advanced dementia populations. Many older adults use Android phones, and sometimes older phone models, as their primary device; however, the mindLAMP app is most compatible with iOS devices. Glitches or bugs are more common in those with Android devices, which can lead to more frustration and difficulty completing tasks than those with iOS devices. Due to the ever-evolving nature of technology, the mindLAMP platform is unable to support any software lower than Android 8, potentially posing a barrier to accessing the SMART-A paradigm.

Our pilot study will establish preliminary data on the acceptance of the SMART-A paradigm in an MCI or AD population and on how the MCI/AD group differs on each measure from the control population. Further analysis of passive data, cognitive tests, and self-reported symptoms on a larger sample size than the pilot study presented here will allow us to optimize the SMART-A protocol to include only the measures that most accurately predict overall cognitive status, both for screening and monitoring purposes.

Our long-term goals are to develop a validated machine learning model to automatically detect cognitive and/or behavioral impairments and to incorporate this into an AI-grounded smartphone/tablet app, which can be used to identify and track patients with very early-stage AD while at home. This app will be used to assess different stages of cognitive impairment and detect clinically meaningful changes such as the transition from independence to supervision. Such an app will offer a range of uses as a diagnostic tool, for home monitoring, and to measure the efficacy of intervention trials. The use of ML models in analyzing the combination of passive and active data collected under this protocol is promising for the development of automatic cognitive monitoring over time, by using longitudinal patterns to make predictions. This allows us the ability to link these together and develop an adaptive version of the app that will actively pop-up cognitive tasks as soon as it detects clinically significant ecological data decline—often referred to as just-in-time adaptive interventions.

In a world where older adults are more reliant on technology than ever before, we can capitalize on a patient’s own devices to glean information about their cognitive health. The innovation of the SMART-A paradigm involves leveraging the capability of smartphones to assess passive data to capture real-time ecological momentary assessments, behavioral data, and surveys, in addition to active data collected via cognitive assessments. Smartphone-based cognitive assessments, when combined with digital phenotyping, may offer a scalable and ecologically valid approach to detecting and monitoring AD in real-world settings. If the pilot study presented provides preliminary evidence of clinical usefulness in an MCI or AD population, future iterations of this framework have the potential to enhance early detection, enable continuous monitoring, and support future machine learning–based automated identification of cognitive impairment, ultimately facilitating earlier and more personalized care.