All data in the current study were obtained from participants enrolled in web-LABrainS. The web-LABrainS study is a self-guided, web-based battery of questionnaires (demographics, mobility, medication use, and health history) and assessments [39]. Web-LABrainS has been continually enrolling participants since 2020. Participants who heard about web-LABrainS via word-of-mouth (no promotions or outreach efforts for web-LABrainS have been conducted) contacted the Pennington Biomedical Research Center (PBRC) to be enrolled. Participants were then sent an email containing a hyperlink that took them to the web-LABrainS site to consent to and complete the web-LABrainS battery. Individuals were sent an additional email and hyperlink every 6 months to repeat the web-LABrainS battery.

Because subjects enroll in and complete the web-LABrainS battery of assessments on a rolling basis, for the purposes of this analysis, we pulled participant data from our servers on October 17, 2024. Data from participants who enrolled in the study any time after this date, as well as assessment data from participants enrolled in the study, were not included in this analysis.

Subjects were classified as either dropouts or adherents of web-LABrainS. A participant was classified as a dropout if they completed their baseline assessment and had no subsequent assessments. A minimum of 9 months had to have passed since their baseline assessment for a participant to be considered a dropout. If a participant completed a second assessment—even if it was more than 9 months after their baseline assessment—they were not classified as a dropout. Participants were classified as adherents if they completed assessments at a minimum of 3 time points (baseline, 6-month, and 12-month assessments). Participants were recruited to the study via minimal outreach efforts that largely relied on word-of-mouth and grassroots promotion by the participants in web-LABrainS. Participants were retained in the study by equally minimal retention efforts that were based on 2 email prompts for those individuals who did not open the email link to complete their web-LABrainS assessment at baseline, 6, or 12 months postbaseline. This approach was undertaken to evaluate the feasibility in a low recruitment and retention environment.

The components of the web-LABrainS assessment have been described previously in a study of the feasibility of this self-guided, web-based paradigm [39]. We will briefly summarize the components of the web-LABrainS assessments used in this investigation. First, participants were presented with demographic questions about their date of birth, gender identification, racial and ethnic background, marital status, and education. Participants were then asked about their living situation (including whether they live alone and their type of residence), the types of technology they use, their self-rated comfort with technology, as well as their driving status and frequency. Participants are asked about how they rate their own mobility and are given a short questionnaire adapted from the Life-Space Assessment [40]. Participants proceed to provide their level of concern for their memory and whether they rate it worse than others their age. Participants are asked about the number and type of prescription medications they use, and their ability to read a medication label. They are asked to provide their personal medical history including identifying the presence of current chronic health conditions. To conclude the survey portion of the web-LABrainS battery, participants are asked whether they wish to be contacted for research portions in the future.

In the next section of the web-LABrainS battery, participants take multiple validated cognitive assessments, including an 8-item orientation assessment as well as 4-item immediate and delayed recall assessments. Finally, they take an assessment of acute symptoms of depression and anxiety.

We used 2 types of statistical analysis in this study. We first implemented a between-group comparison of dropouts and adherents. This involved 1-way chi-square tests of independence to assess the significance of the differences in the observed and expected frequencies of all categorical dependent variables that were examined. It also involved Mann-Whitney U tests to assess the significance of the differences in the distributions of all numerical-dependent variables that were examined. Exact P values were reported; however, a significance threshold (α) of .05 was used to report statistically significant results.

Our second form of statistical analysis was a binary logistic regression model, a predictive model trained on our data from which we calculated odds ratios (ORs) and CIs to determine the relative strength of our examined variables as predictors of dropout status. We trained and cross-validated 4 different models. We used different sets of predictors in each of these models and iteratively tuned their hyperparameters for optimal performance. Depending on the model, this involved adjusting the optimization algorithm, regularization method, regularization strength, size trim tolerance for predictor coefficients, convergence tolerance, or number of iterations over multiple runs of the model and selecting that which resulted in the strongest classification performance. We assessed our model’s classification ability using k-fold cross-validation. We first shuffled the dataset and split it into 5 separate folds. Four of the folds were used to train the model, while the last fold was withheld to be used as the test set. This process was repeated with each fold being used as the test set once. The trained model was used to predict the outcomes of the unseen subjects in each set. For each iteration of the model, we generated a classification report that included the model’s precision, recall, and F₁-scores to evaluate model performance for each class (dropout and adherent). The weighted and unweighted averages of these scores across the 2 classes were also provided along with the total model accuracy. The support values for these averages were also included. The 5 classification reports produced for each iteration were then averaged.

We then selected the best-performing model from the cross-validation phase. Of the 4 models tested, our chosen model was tied for the highest overall accuracy but exhibited more consistent performance across validation folds and better classification of the minority class. This model included the variables that significantly differed between groups from our first analysis (self-rated mobility score, delayed recall performance, total number of prescription medications taken, use of depression medication, and quality-of-life self-rating) along with demographic variables (age, race, ethnicity, gender, and years of education). With that subset of data selected, we preprocessed our data in the same way we did during our cross-validation stage. This involved applying 1-hot encoding to all categorical variables, converting them into a binary format that the model was able to use. It also involved converting all continuous variables to a standard scale with a mean of 0 and SD of 1 to ensure equal contribution of these variables to our model. Finally, we added a constant term to the dataset to be used as a reference by the model.

The model itself was based on a logit function. The coefficients of this function were estimated using a convex optimization algorithm with L1 or Lasso (Least Absolute Shrinkage and Selection Operator) regularization. We applied a regularization strength of 0.01 to the coefficient estimation. We also set a trim tolerance of 0.05 for coefficients with a small absolute value. After the model was fit to the full dataset, we retrieved the coefficients of each predictor and calculated the corresponding ORs. The ORs show how the probability of the outcome—in this case, dropping out— changes with an increase of 1 SD to a numerical variable or the presence of a binary variable.

All statistical analysis was performed in Python [41] using functions from the numpy, scipy, sklearn, and statsmodels packages [42-45]. Outputs from the Web-LABrainS tool were stored in data structures and manipulated using objects and functions from the pandas Python package [46]. Data visualization was accomplished using the matplotlib and seaborn packages [47,48].

All study procedures were approved by the PBRC Institutional Review Board (IRB). Informed consent was provided for all participants prior to the initiation of study procedures. Participants were informed that participation was voluntary and that they could opt out at any time. Informed consent included the ability of PBRC researchers to conduct secondary analysis without additional consent. All data were stored in a deidentified manner to maintain participant confidentiality. No compensation was provided for study participation as outlined in the PBRC IRB-approved study protocol. All data used for this study were obtained from web-LABrainS. IRB approval for the web-LABrainS study was obtained from the PBRC IRB (FWA # 00006218) prior to the initiation of web-LABrainS research efforts. The PBRC IRB approval number is 2020-044-PBRC Web-LABrainS. The study complied with ethical standards outlined in the Belmont Report and Declaration of Helsinki.

Our results in this study support our initial hypothesis that the use of a self-guided, remote, and web-based battery can be used to conduct longitudinal research in community-dwelling OAs. We identified that demographic characteristics such as age, education, gender, marital status, and living alone were not significantly correlated with participant dropout. Similarly, we observed that the reported comfort level with using technology, the amount of technology use, interest in research, and the time of assessment were also not correlated with participant dropout. Conversely, we identified a diverse array of medical-related issues that were significantly correlated with study dropout, including use of depression medication, decreased mobility, and impairments in delayed recall performance, all of which were significantly correlated with study dropout. The importance of each of these findings is discussed below.

While the level of dropouts in this study was high, it is in line with what has been reported by in-clinic longitudinal studies. It is important to point out that recruitment of participants to web-LABrainS in this feasibility study did not involve efforts to target specific populations or efforts to do large-scale recruitment, being that recruitment was largely restricted to word-of-mouth among participants and a byproduct of correspondence with our institute. Additionally, it should be pointed out that participants received 2 reminder emails to conduct their 6- or 12-month assessment batteries and received limited or no additional follow-up (or forms of follow-up) to maintain participant retention in this feasibility study. Given the extremely restricted outreach and retention efforts in this study, we believe that even though our feasibility effort went out to only 12 months, it shows that this approach is a feasible effort to have community-dwelling OAs maintain participation in a longitudinal research effort that involves repeated assessments over time.

Many in-clinic longitudinal studies have investigated which of the variables collected in their demographic surveys, health questionnaires, and physical or cognitive assessments are significant predictors of participant attrition. One of the most commonly identified variables reported in the literature is age, although some studies claim older age makes a subject more prone to attrition [12,14,18,19,21] while others suggest younger subjects have a greater risk of study dropout [7-9,23]. Past research also shows variability in other predictors of study dropout including gender and education. Several studies identify being male as a significant predictor of study dropout [8,12,21], while others have shown that being female increases the risk for study dropout [16,19]. Likewise, past research has found both low levels of education [7,8,10,15,16,20,22,23] and high levels of education [21] to be significant predictors of study dropout. Taken together, data from this study and existing literature suggest that the retention of participants in longitudinal studies may be impacted by some factors (age, gender, and education) that are study-specific and therefore challenging to control for or successfully address.

Other predictors of participant attrition did not vary in directionality. These variables associated with participant dropout included having lower socioeconomic or job status, lower cognitive function, and lower self-reported or general health [7,10,12,15,19,21,22]. Medical conditions such as chronic stress, cardiovascular disease, coronary artery disease, heart attack, respiratory dysfunction, diabetes, depression, and anxiety were commonly identified predictors of study dropout when surveyed for or directly assessed [9-11,19,22]. Finally, lifestyle factors such as smoking, low physical activity, and low social and community involvement were also commonly associated with study dropout [7,10,12,21]. Living situations played an interesting role in the risk of attrition in previous studies. The presence of familial conflict or young children was identified as a significant predictor of study dropout [9]. Living alone or having fewer potential caretakers also significantly predicted attrition [12,18,19]. Based on the consistency of these findings, future studies should set out to fully address the potential impact of each of these aspects on participant retention. For example, future studies (including web-LABrainS 2.0) should consider incorporating strategies that minimize the negative impacts of lower socioeconomic status (providing participant stipends and maximizing flexibility for when study visits occur), impaired cognition (consider incorporating study partner and simplifying assessment screens), and poor health (allow study staff to assist with evaluation if requested, provide alternative types of assessments for individuals with physical limitations).

Our studies identified several factors associated with dropout in this web-LABrainS feasibility effort and the findings from studies using traditional longitudinal study approaches. For example, we observed that dropout in the web-LABrainS effort was not correlated with age, education, gender, marital status, or living alone. In contrast, dropouts in web-LABrainS had significantly higher levels of depression, decreased self-reported mobility, and decreased memory performance (delayed recall). In addition, we identified 2 novel factors—decreased quality of life and increased number of prescribed medications—that were associated with study dropout. Interestingly, in this web-LABrainS feasibility research effort, no differences in technology use, comfort with technology, or time of assessment were observed between adherent participants and dropouts. Another surprising finding in this study was the finding that there was an extremely high percentage (greater than 95%) of both adherent participants and dropout participants in web-LABrainS who confirmed a willingness to be contacted about future research studies. These data suggest that alternative factors, including the perceived value of participation in the specific study, play a role in participant retention. For example, it is likely that some participants continue in a longitudinal study solely based on personal reasons that have not been discussed above (family member impacted by the disease or dedication to a university, clinician, or researcher). Alternatively, participants may differentially stay in a longitudinal study based on their perceived value of the study results they receive. It should be noted that for web-LABrainS, we routinely receive participant correspondence saying they value the opportunity to participate because of a family member who had dementia or because they highly value the study summaries they receive at the conclusion of each assessment.

There are multiple limitations in this feasibility study. For example, our web-LABrainS participants are not a representative sample of the community, and in future research efforts, it will be critical to optimize outreach such that the recruitment of a representative community sample is reliably obtained. With a more representative sample and larger sample size, we could also expect predictive modeling for risk of dropout to improve in accuracy from the binary logistic regression model shared in this study. Additionally, there was an expected elevated dropout in this web-LABrainS feasibility study due to the extremely limited nature of participant retention efforts (2 email reminders). In future studies, it will be key to identify and implement features of our platform to maximize the retention of study participants. We believe, based on the incidental feedback from web-LABrainS participants over the course of this feasibility study, that the major reason for study participation is the fact that each participant gets a detailed and automated report on their cognitive function at the conclusion of the web-LABrainS assessment. Continuing to optimize feedback and related collaterals to study participants will likely be critical to ensuring the long-term retention of study participants for years or even decades in longitudinal research efforts.

Taken together, the findings in this study provide a framework for the design and implementation of longitudinal research studies that focus on the use of a self-guided, remote, and web-based approach. Additionally, the data in this study provide important findings on how different demographic, health, and technology use aspects correlate with adherence or dropout in the web-LABrainS approach. In addition to being feasible, we believe using this new paradigm for longitudinal research will significantly decrease participant burden, dramatically lower research costs, and allow for significantly greater recruitment of participants across a wide geographic area. In future studies, it will be critical to develop methodologies and approaches that facilitate participation by a more racially and socioeconomically diverse array of study participants. These efforts will likely include outreach on social media platforms and collaborations with stakeholders in the community that align with racially and socioeconomically diverse populations. It will likely be important in these future efforts to make the outreach and onboarding of participants as frictionless as possible for participants by optimizing the user interface and user experience. In the interest of improving the experience of the community stakeholders involved in the future research effort, optimizing the application programming interface of the research platform will be a high priority. The impact of these efforts will have a significantly greater clinical impact if they can begin to be incorporated into clinical care as screening, patient engagement, and caregiver support tools. Our future efforts will attempt to incorporate each of these aspects as well as develop pathways for web-LABrainS participants to be connected quickly with study staff and health care professionals when mild cognitive impairment or dementia is identified as part of their longitudinal assessment.