Identifying digital phenotypes of risk for Alzheimer's disease and related dementia among Hispanic/Latino persons living in the United States: A protocol for a prospective quantitative study

Culturally relevant considerations

Study procedures follow a culturally informed approach. Our study team includes several bilingual/bicultural members, including one Multiple Principal Investigator (MJM) and all study staff who had participant contact via recruitment efforts or data collection. Study visits are conducted in Spanish or English based on participants’ primary language, determined via methods used in prior studies. ⁴⁷ These methods include a validated algorithm based on a brief measure assessing participant's language preference, fluency, and use in daily life ⁴⁸ and confirmed via performance-based measures of fluency in English ⁴⁹ and Spanish. ⁵⁰ We have implemented several strategies to reduce practical barriers to research participation including: a testing site in a community with high Latino representation; home visits when needed; study visits during weekends and after regular working hours; transportation to/from study visits; meals during visits; childcare reimbursement; compensation for time and effort; and return of study results. Based on a comprehensive literature review, we identified the most appropriate assessment instruments for middle-age and older Latino persons. We chose those with available English and Spanish versions and strong psychometric properties for Latino persons.

Recruitment and retention

Recruitment activities occur in San Diego County (California). Recruitment efforts leverage an NIH-funded longitudinal observational cohort study of Latino persons (R01MD013502), an existing participant pool at the San Diego State University SBLRC, and other participant registries at UCSD. Bilingual study staff reach out to participants in these cohort studies who had previously provided consent to be contacted for future studies via telephone and provide an overview of the study using a low-literacy recruitment script. Building on years of experience fostering trust in the local Latino community, we also recruit new participants, with a focus on persons residing in Latino neighborhoods in southern San Diego County, CA. Activities for recruitment of new participants include presentations to community organizations that serve the older adult Latino community, participation in health events, flyer distribution, establishing a formal partnership with a promotora de salud, referrals from current participants, and mailings to Latino neighborhoods. Recruitment materials were codeveloped in English and Spanish, utilizing plain language and considering cultural appropriateness.

To ensure evidence-based retention efforts for the duration of the study we will: 1) form a direct relationship with each study participant, 2) call and check-in with participants during the first three days and every two weeks of the digital data collection period, and 3) encourage participants to let us know when their contact information changes and do periodic checks on the accuracy of the records. We will make reminder calls to participants one week, then again one day, prior to their follow-up visit. In addition, our team utilizes participant retention methods based on previously published principles, including: 1) informing individuals that they will be followed up for the duration of the study; and 2) obtaining permission at the time of entry to allow follow-up through other contacts (e.g., family members, trusted confidants).

Laboratory assessments

Neurobehavioral assessments

Neurocognitive function is assessed via a comprehensive neuropsychological assessment battery, which measures seven ability domains (Table 1). All tests are available in English and Spanish and have demographically adjusted normative data for Latino populations. All tests are completed at the baseline and 36-month visits, and a subset at the 18-month visits (noted in Table 1). The neurobehavioral assessment includes also self-report measures of current problems with instrumental activities of daily living ⁸ and self-report of cognitive symptoms. ⁵¹

OPEN IN VIEWER

Table 1.

Comprehensive neuropsychological test battery.

Neurocognitive domain	Cognitive test (score)
Learning	Hopkins Verbal Learning Test-Revised (Total Recall)*(53) Brief Visuospatial Memory Test-Revised (Total Recall)(53) UDSv3 Craft Story (Immediate Verbatim)(52)
Delayed Recall	Hopkins Verbal Learning Test-Revised (Delayed Recall)*(53) Brief Visuospatial Memory Test-Revised (Delayed Recall)(53) UDSv3 Craft Story (Delay Verbatim)(52)
Language/Verbal Fluency	UDSv3 Animal Naming (Total Score in 60 s)*(52) UDSv3 Phonemic Fluency (P Spanish and F English Total Score)*(52) UDSv3 Multilingual Naming Test (MINT Total Score)(52)
Attention/Working Memory	UDSv3 Digit Span Forward (Total Correct)*(52) UDSv3 Digit Span Backwards (Total Correct)*(52)
Processing Speed	WAIS-III Digit Symbol (Total Score)(55) WAIS-III Symbol Search (Total Score)(55) UDSv3 Trail Making Test A (Time in s)*(52)
Executive Functioning	Wisconsin Card Sorting Test-64 (Perseverative Responses and Total Errors)(54) UDSv3 Trail Making Test B (Time in s)*(52)
Visual-Spatial Skills	UDSv3 Benson Complex Figure Copy (Total Score)(52)

Note. All tests are administered at baseline and 36 months. *Tests administered at 18 months.

UDSv3: Uniform Data Set-Version 3; WAIS: Wechsler Adult Intelligence Scale.

Neurocognitive data will be utilized to compute three types of outcomes: 1) Cross-sectional continuous neurocognitive performance: Raw test scores for each test are converted into demographically adjusted T-scores (age, education, gender).^52–55 Adjusted T-scores are averaged to compute global and domain T-scores. ¹⁸ 2) Diagnosis of MCI. Objective presence of MCI is determined via Jak/Bondi diagnostic criteria. ⁵⁶ Per these criteria, MCI is defined by impairment in at least one cognitive domain. Domain impairment is quantified by scores of >1SD below the normative mean on at least two tests within the domain. 3) Longitudinal neurocognitive change will be analyzed utilizing cognitive T-scores at baseline, 18 months and 36 months. ⁵⁷

Digital phenotyping data

Descriptions of the digital data collected are provided below and summarized in Table 2 and include keystroke metadata, sleep, physical activity, geolocation, and phone use activity. Participants use their own smartphones for the study, and Fitbit Versa 2 or 4 watches are provided to collect sleep and physical activity data over 30 days post baseline laboratory visit.

OPEN IN VIEWER

Table 2.

Summary of digital data collected.

Data source	Data processing	Off the shelf features and established metrics
GPS and GIS data	PALMS processing	Time spent at home Distance from home Total walking distance Neighborhood characteristics
Keyboard Typing Patterns	KeyWise processing	Typing speed variability Average interkey delay Backspace usage Circadian rhythms accelerometer during typing
Wrist-worn accelerometer	TLBC processing	Step count Distance Overall physical activity Total sleep time Sleep efficiency
Social Activity	No preprocessing	Average screen time Variability in screen time Most used apps Categories of apps used

Note. GPS: Global Positioning System; GIS: Geographic Information System; PALMS: Personal Activity and Location Measurement System; TLBC: Two-Level Behavior Classification.

Feature engineering overview

We will use a combination of existing features provided by industry partners (e.g., robust statistics metrics previously used to predict age and smartphone usage), ⁸⁴ and we will engineer additional features using nonlinear dynamical systems methods developed to study intrasubject variability dynamics over time (e.g., recurrence quantification ⁸⁵ and multiscalar entropy analysis ⁸⁶ ) as well as traditional time-series methods. Intrasubject variability will provide interpretable features that may be indicative of subtle differences due to MCI and/or CVD risk.

Using the keystroke data as an example, an outline of the specific features we aim to extract to demonstrate the potential richness and predictive power of digital phenotypes is as follows. We plan to extract a comprehensive suite of keystroke and session-level features from each individual's daily typing data. Specifically, we will quantify interkey timing metrics—such as the median, harmonic mean, 95th percentile, root mean square of successive differences, and median absolute deviation of alphanumeric-to-alphanumeric intervals, as well as session characteristics such as median and mean session length and total typing duration per day. We will also capture error-related behaviors (e.g., backspace and autocorrect event rates) and derive variability and complexity indices (e.g., multiscale entropy of interkey intervals).

Prior research indicates that these measures predict both demographic and cognitive signal: typing speed can predict chronological age, ⁸⁷ while entropy-based variability in keypress timing correlates with planning performance on the Tower of London task. ⁸⁸ Furthermore, within-subject fluctuations in daily typing speed have been shown to track trail-making task performance and depressive symptom severity.^89,90 In work currently in review on younger adults (ages 21–40), we derived 39 keyboard-dynamics features and used principal component analysis to reveal that the first component (overall typing speed) and second component (usage patterns such as pause behavior) each significantly predict fluid cognition (NIH Toolbox composite) and TMT-B completion time. ⁹¹ For the current sample, we will extend this approach by incorporating pause-based features—capturing intrasession idle durations—to enhance sensitivity to fine-grained changes in processing speed and attentional control.

Lastly, we will use artificial neural networks, specifically a deep learning method that combines convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) called CNN-LSTMs. It will extract features from the individual data streams to examine hidden markers that are predictive of the dependent variables. Such models can automatically “featurize” the raw data, including temporal patterns, during the end-to-end training process. ⁹² To reduce the risk of overfitting, we will regularize the model aggressively and explore unsupervised pretraining to learn the lower layers of the CNN features.

Feature engineering analyses

For feature extraction from the passive metrics, the goal is to develop stable parameter estimation of the intrasubject variability using both standard and ML approaches. Standard approaches to capture intrasubject variability involve tools from dynamical systems (complex systems tools), such as entropy-based methods for physiological time series methods, and have been used successfully in bipolar disorder to predict changes in mood self-report, ⁹³ motor activity, ⁹⁴ motor activity in depression, ⁹⁵ and a range of other physiological disorders such as heart disease. ⁹⁶ Multiscalar entropy analysis ⁹⁷ will be applied to measure intrasubject variance, and our own simulations show we need to at least 3000 data points to achieve stable estimates. For keystroke metadata, we can safely assume 300 keystrokes/day × 30 days = 9000 keystrokes/person. For GPS/GIS data, the polling rate is 12 times an hour × 24 h × 30 days = 8640 data points. For the other metrics of intrasubject variance where the time-series is shorter (e.g., gait speed/day, sleep/night), we will use recurrence quantification analysis, another complex systems tool for nonlinear time-series data. Recurrence quantification analysis can be understood as the generalized autocorrelation function and allows the detection of deterministic elements of a signal what might look stochastic in nature ⁹⁸ and has been successfully applied to EEG, ⁹⁹ heart rate variability, ¹⁰⁰ and postural sway. ¹⁰¹ Recurrence quantification analysis produces several metrics (including entropy) from nonlinear time series, such as where in time the person is repeating patterns and how stable patterns are in time.

Machine learning and statistical analyses

Alzheimer's Disease and Related Dementias risk groups (MCI+/-, CVD+/-) will be predicted from data obtained via passive digital measures collected over 30 days. Digital features from the passive metrics (both standard and machine-learned) will be entered into ML methods designed for high-dimensional multivariate predictors in small sample sizes for cognitive T-scores and CVD risk. ¹⁰² Interactions with sex and with language use and other sociocultural factors will also be examined. LASSO will select the most impactful passive digital features while limiting model complexity (help prevent overfitting), maintaining type I error levels in a small sample relative to the number of parameters. ¹⁰³ XGBoost, ¹⁰⁴ a method that incorporates pruning and regularization methods (like LASSO), will help assess the importance of individual features. For the classification of ADRD risk group, modern ML classifiers will be used with standard metrics to assess model fit (i.e., raw accuracy, area under the ROC, specificity sensitivity, and precision). Predictors will be tested stepwise to improve interpretability. The specific algorithms selected will depend on a balance of power relative to underlying assumptions^105,106 that are met with an emphasis on model interpretability (e.g., QDA, random forest, XGBoost). To guard against overfitting in all our machine-learning models, we will first set aside 20% of subjects (stratified by group) for a final hold-out validation. Within the remaining 80%, we will perform k-fold cross-validation (e.g., k = 10), monitoring variability in balanced accuracy and F1-score across folds. To reduce feature-space complexity, we will preselect predictors informed by prior literature and, where necessary, apply dimensionality reduction (e.g., PCA, T-SNE). For our CNN-LSTM pipelines, we will additionally withhold 50% of each individual's time-series streams (Fitbit, GPS, KeyWise, screen time) during training, and employ dropout and regularization (e.g., L₂ weight-decay) in the convolutional and recurrent layers. Finally, we will evaluate generalization performance on the 20% subject hold-out set mentioned above.

The specific cross-validation technique will depend on algorithm selection (e.g., leave-one-out, K-fold, or bootstrapping). ¹⁰⁷ To assess individual feature importance SHAP (SHapley Additive exPlantations) values will be extracted. To balance type I and II error rate and control for multiple testing, we will use an optimal alpha strategy (a = .05 to .005) depending on whether the question at hand is a main hypothesis or exploratory and relative analysis being conducted (e.g., ML analysis, classical or robust parametric and nonparametric statistics), and how the p-values are derived for the specific algorithm tested (e.g., classical, Bayesian, or bootstrapped). We will use Holm–Bonferroni adjustments for ≤5 tests ¹⁰⁸ and false discovery rate methods for >6 tests. ¹⁰⁹ When dependent outcome measures are correlated, corrections will be calculated based on the effective number of independent tests. ¹¹⁰ For the longitudinal exploratory aim, generalized linear mixed-effect models ¹¹¹ will model neurocognitive change and its association with changes in digital data features over time. ¹¹² The variable selection methods above will be used to explore which features are the most impactful. ¹¹³ Finally, we will explore the standard feature importance analysis method of drop-feature evaluation, an ablation analysis where we withhold a subset of features at a time and rebuild the models.

Sample size and power analyses

We conducted a sensitivity analysis to determine the minimum effect sizes detectable in our study, using a multiple regression sensitivity analysis. Drawing on previous studies with passive metrics, we anticipate identifying 18–20 passive features relevant to the primary study aim. With 20 predictors, we estimated small-to-medium effect sizes between individual predictors and cognition T-scores, achieving a power of 0.80 with Cohen's f² = 0.026 (Cohen's d = 0.32; R² = 0.025) and a power of 0.95 with Cohen's f² = 0.044 (Cohen's d = 0.42; R² = 0.042). For the examination of sex, language use, and other sociocultural factors, we will also explore the smallest effect size necessary for accurate participant classification into predefined groups using logistic regression, the oldest and most conservative method, assuming moderate intercorrelation among normally distributed predictors (R² = 0.30): a power of 0.80 analysis resulted in an odds ratio of 1.49 or Cohen's d = 0.22, and a power of 0.95 resulted in an odds ratio of 1.69 or Cohen's d = 0.29. These results indicate adequate power to detect small-to-medium effect sizes.