Evaluation of Facebook as a Longitudinal Data Source for Parkinson’s Disease Insights

1. Introduction

Social media platforms are used by nearly 60% of the global population [1], with Facebook alone reporting 2.89 billion active users in 2024 [2]. Beyond maintaining social connections [3], older adults use social media to share aspects of their daily lives, including health-related experiences [4]. Prior research has demonstrated that social media data, including explicit health disclosures and behavioral indicators (e.g., changes in posting frequency, sentiment shifts), can be leveraged for medical applications [5,6]. However, the extent to which social media serves as a meaningful data source for specific neurological conditions, including Parkinson’s disease (PD), remains underexplored.

PD, the second-most common neurodegenerative disorder, affects over 6.2 million people worldwide [7]. Its clinical heterogeneity, spanning both motor and non-motor symptoms, complicates diagnosis and progression modeling [8]. PD is diagnosed based on motor symptoms that emerge after extensive dopaminergic neurodegeneration, often estimated to exceed 50% neuronal loss [9]. The prodromal phase, during which neurodegeneration occurs before clinical diagnosis, remains difficult to study prospectively, as most PD cases are idiopathic [10]. Leveraging retrospective data sources that document lived experiences before and after diagnosis may provide insights into the premorbid period and early disease manifestations.

This study evaluates the feasibility of using Facebook as a data source for PD research. We assess whether individuals with PD and related disorders are represented on the platform, the extent of their health-related disclosures, and whether their longitudinal social media activity can provide insights into the prodromal phase. Because PD shares clinical features with and is sometimes misdiagnosed as Essential Tremor (ET) and Atypical Parkinsonism (AP) [11], individuals with these conditions were also included. Caregivers, who frequently seek and share PD-related information online [12], were included as well.

This work establishes social media as a viable source of PD-related information and lays the foundation for future studies leveraging large-scale social media data to identify disease-relevant disclosures and track self-reported experiences over time.

2. Materials and Methods

2.1. Participants and Data Collection

Participants were recruited through the Emory Brain Health Center, research recontact lists, and PD-specific community events. Eligible participants included individuals diagnosed with PD, ET, or AP, as well as caregivers of individuals with those conditions. Participants had to communicate in English. No additional exclusion criteria were applied. Written informed consent was obtained in person or via teleconferencing, following protocols approved by Emory University’s Institutional Review Board (STUDY00005722) and the Declaration of Helsinki.

During structured interviews, demographic and clinical information were collected. Disease severity and symptom burden in PD and AP participants were assessed using the Movement Disorder Society–Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) Parts I (non-motor symptoms), II (motor symptoms), and IV (motor complications) [13], where higher scores indicate greater impairment. ET participants were assessed using MDS-UPDRS Parts I and II, as motor complications (Part IV) are not relevant to ET. MDS-UPDRS Part III (motor examination) was not collected, as physical examinations were outside the scope of data collection.

Participants who opted to share their Facebook data were guided through downloading their Posts, Comments and Reactions, Pages, and Groups in JSON format. They were instructed to download their complete activity history, capturing data from account creation through study participation. All text-based content—posts, comments, media captions, and other written engagement—was authored by the account owner and collectively referred to as "posts." Private messages were neither collected nor analyzed. Study staff provided technical assistance as needed. All data were securely uploaded to a HIPAA-compliant REDCap database [14].

2.2. PD-Relevant Post Identification

2.2.1. Text Extraction from JSON Files

Participant-generated text was extracted from Facebook data exports. Given that individual JSON files were not mutually exclusive, deduplication was necessary. Exact duplicates were removed based on identical timestamps and text, and near-duplicates were filtered by retaining only the earliest instance unless at least 180 seconds had elapsed between occurrences, ensuring that only intentionally reshared content was preserved.

Posts flagged as shared memories were retained, as they represented intentional reposts, but associated in-text timestamps (e.g., “X years ago”) were removed to prevent redundancy. Text encoding was standardized, and the final dataset was stored in a HIPAA-compliant REDCap database [14].

2.2.2. PD-Related Term Dictionary and Search Strategy

A PD-specific term dictionary with over 1,000 terms was developed from academic literature [8], WebMD, and our prior work on fall-related disclosures in PD [15]. To enhance coverage, ChatGPT-4 generated medical and colloquial synonyms, and the dictionary was refined through manual review. The dictionary spanned motor and non-motor symptoms, treatments, assistive devices, caregiver roles, community engagement, genetics, advocacy, and research participation. See Appendix A for the full dictionary.

The text was lowercased, punctuation was removed, stopwords were filtered, and words were stemmed [16]. A keyword search flagged posts containing either stemmed or unstemmed terms.

2.2.3. Ground-Truth Dataset

A subset of 6,750 posts from 14 individuals with PD and 5 caregivers was manually reviewed. Each post contained at least one keyword from the PD-related term dictionary and was independently assessed by two reviewers for relevance to PD using a high-sensitivity approach.

Posts were included if they explicitly mentioned PD, related disorders, treatments, or lacked sufficient context to rule out a PD connection (e.g., "I had a follow-up with my doctor today"). Posts were excluded only if sufficient context indicated they were unrelated (e.g., "I am so fatigued from COVID").

Given that both caregivers and individuals with PD may discuss others with PD, posts were evaluated for general relevance to PD, not just the poster’s personal experience. Conflicts were resolved through discussion.

2.2.4. Classifier Development

The dataset was randomly split into 80% training and 20% testing. All feature transformations, including vectorization, scaling, and encoding, were derived from the training set and applied to the test set. Posts were cleaned by removing URLs, tags, hashtags, special characters, punctuation, extra spaces, and stopwords. The processed text was then represented in two ways. First, lemmatization was performed using the NLTK WordNet lemmatizer [17] without explicit part-of-speech tagging, meaning words were lemmatized assuming noun forms by default. Lemmatized text was vectorized using TF-IDF, capturing single words, bigrams, and trigrams while excluding terms appearing in fewer than two documents to reduce noise. Second, tokens were mapped to cluster-based embeddings using a word-clustering approach derived from a Twitter-based corpus [18]. Clusters were vectorized using TF-IDF, capturing only single words and applying the same frequency threshold.

Model features included vectorized lemma and cluster representations, normalized age at time of posting, and one-hot encoded categorical variables (gender, PD diagnosis status). Multiple classifiers, including K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Random Forest, AdaBoost, Naïve Bayes, Decision Trees, and XGBoost, were trained using five-fold cross-validation. GridSearchCV [19] was used to optimize macro-averaged recall, ensuring balanced sensitivity across PD-relevant and non-PD content. We also experimented with a soft-voting ensemble that combined all classifiers above.

Performance was evaluated using macro-averaged recall, precision, and F₁-score to account for class imbalance. 95% confidence intervals (CIs) were estimated via bootstrapping, and model discrimination was assessed using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) scores. The classifier with the highest recall was applied to the full dataset.

2.3. Analyses

Demographic and clinical characteristics were summarized using descriptive statistics. Measures of central tendency are reported as Mean ± SD unless otherwise noted. Group differences between Facebook users and non-users were assessed using t-tests for continuous variables and Fisher’s exact tests for categorical comparisons.

Facebook-derived text data were analyzed from 45 participants, including a PD-specific subset (N=29) analyzed separately to examine pre- and post-diagnosis content. Facebook engagement was measured by calculating the average duration of use (time between the oldest and newest post), total posts, and time on Facebook before PD diagnosis.

PD-related posts were identified using the best-performing classifier, applied to posts containing at least one term from the custom PD dictionary. The volume of PD-related content was summarized descriptively. Additional analyses examined posting frequency before and after diagnosis and excluded exercise-related content based on predefined terms. A rule-based analysis identified participants who explicitly mentioned ‘Parkinson’ or ‘\bPD\b’ in any post.

3. Results

3.1. Dataset Characteristics and Participant Engagement

3.1.1. Recruitment and Demographics

We attempted to contact 369 individuals and successfully reached 139. During phone screenings, participants were asked whether they had a Facebook account. Of those contacted, 59% reported having a Facebook account, while 20% did not specify their Facebook status. Among individuals who confirmed having a Facebook account, 85% agreed to participate (see Figure 1A).

Of the 139 individuals reached, 62 enrolled, and 60 completed the interview. The final dataset included 38 individuals with PD, 4 with ET, 3 with AP, and 15 caregivers. The sample was 57% male, with an average age of 66.53 ± 12.99 years. Most participants were white (80%), non-Hispanic or Latino (93%), and college-educated (78%). Among participants with PD (N = 38), the average age was 69.21 ± 10.11 years. Disease characteristics included an average disease duration of 8.62 ± 4.73 years, an MDS-UPDRS Part I score of 14.03 ± 7.48, an MDS-UPDRS Part II score of 13.81 ± 9.68, and an MDS-UPDRS Part IV score of 6.25 ± 4.

Among the 60 participants, 46 shared their Facebook data. Eleven participants did not have Facebook accounts and three had Facebook accounts but did not complete the data-sharing process. Of those who shared Facebook data, 30 had PD, 3 had ET, 1 had AP, and 12 were caregivers.

3.1.2. Comparison of Facebook Users and Non-Users

We compared Facebook users and non-users based on demographics and clinical status. Participants with Facebook accounts were significantly younger than those without accounts (64.46 ± 13.24 vs. 75.72 ± 6.41 years; t(32.14) = 4.16, p < .001, d = -0.91). However, there were no significant differences between Facebook users and non-users in gender (p = .09), race (p = .49), ethnicity (p = 1.00), or education level (p = .34).

Among participants with PD (N = 30), those with Facebook accounts were also significantly younger than those without (67.24 ± 9.89 vs. 77.96 ± 5.69 years; t(15.52) = 3.84, p = .002, d = -0.46). Facebook users and non-users with PD did not differ in MDS-UPDRS Part I score (p = .86), Part II score (p = .32), Part IV score (p = .39), or disease duration (p = .81).

3.1.3. Facebook Account History

Among the 46 participants who shared Facebook data, one had an account but never generated or shared any content, resulting in empty JSON files. This participant was included in overall Facebook user counts but excluded from subsequent analyses. All reported results are based on the remaining 45 participants.

On average, participants had used Facebook for 13 ± 3 years and posted 4,491 ± 5,637 times. Among those with PD (N=29), the average length of Facebook use was 14 ± 3 years, with an average of 4,018 ± 5,570 posts. On average, participants with PD had used Facebook for 5 ± 6 years before their diagnosis, with 90% creating their account before diagnosis (see Figure 1B).

3.2. Identification of PD-Relevant Posts

3.2.1. Ground-Truth Dataset

Reviewers achieved substantial interrater reliability (Cohen’s Kappa = 0.79) [20]. Among the 6,750 reviewed posts, 2,400 (35.6%) were classified as PD-relevant, while 4,350 (64.4%) were deemed irrelevant, reflecting the broad range of topics discussed. This imbalance was partly due to keyword ambiguity, where certain terms appeared in non-PD contexts.

3.2.2. Classifier Performance

Table 1 presents the macro-averaged classification metrics for each model. The Naïve Bayes classifier achieved the highest recall (0.86 95% CI: [0.84-0.88]). Although the soft-voting ensemble was implemented, the Naïve Bayes classifier was selected for final PD-related post identification due to its superior recall, ensuring sensitivity to PD disclosures. Please see Appendix B for model hyperparameters.

Overall model discrimination was strong, with an AUC of 0.94 (Figure 2), indicating high separability between PD-relevant and irrelevant posts.

3.3. PD-Relevant Posts

The Naïve Bayes classifier was applied to detect PD-related content. On average, participants made 104 ± 176 PD-related posts. 96% of participants had at least one post flagged as PD-relevant. Additionally, 67% of all participants explicitly referenced PD at least once.

Among individuals with PD (N = 29), each participant made an average of 99 ± 172 PD-related posts. Of these, an average of 24 ± 58 posts occurred before diagnosis, and 75 ± 156 occurred after diagnosis. 69% of individuals with PD explicitly mentioned PD at least once, and 93% had at least one post flagged as PD-relevant (see Figure 1C).

Exercise is an important activity for maintaining quality of life in PD [8], though it may be less relevant to detect prodromal symptoms. Since exercise-related terms were included in the term dictionary, we conducted a sensitivity analysis excluding exercise-related posts. Excluding exercise-related posts, the average number of PD-related posts per person before diagnosis decreased from 24 ± 58 to 15 ± 34, suggesting that some pre-diagnosis PD-related content was driven by exercise but not entirely. After diagnosis, the decrease was smaller (75 to 49), suggesting that exercise-related content accounted for a substantial portion of PD-related discourse but was not solely responsible for the observed differences.

4. Discussion

4.1. Principal Findings

This study establishes Facebook as a feasible and informative data source for PD research, demonstrating that individuals with PD maintain long-term social media histories spanning both pre- and post-diagnosis periods. PD-related content surfaced before formal diagnosis, suggesting that early symptoms and concerns may emerge online. While some pre-diagnosis posts referenced exercise, a non-specific aspect of PD management [8], excluding these did not eliminate PD-related content, reinforcing that prodromal disclosures extend beyond fitness discussions.

Beyond individuals with PD, caregivers and individuals with other movement disorders also contributed to PD-related discussions, reinforcing that insights can be gleaned from a broader community. The high prevalence of caregiver engagement aligns with findings from Chu and Jang [12], who reported that caregivers play a central role in online PD discussions. This finding suggests that social media reflects not only patient experiences but also the perspectives of those supporting them, which may be valuable for understanding disease impact beyond clinical reports.

While keyword-matching and traditional machine learning approaches likely underestimated true PD-related discourse, the study demonstrates the potential of social media data for identifying early disease signals. Future work should integrate context-aware pipelines, including large language models (LLMs), to enhance detection and interpretation.

4.2. Ethical Considerations

This study highlights the ethical advantages of our Facebook-based approach, where participants actively consented to data sharing rather than having their posts passively scraped via an API. Social media research on platforms like Reddit typically involves extracting user-generated content without direct user awareness, even if permitted by the platform’s terms and conditions. Following the Cambridge Analytica scandal, Facebook restricted API access [21], leading us to collect data directly from account owners instead. While this method enhances ethical integrity, it limits scalability and restricts datasets to unidirectional conversations, as only account-owner content is retrievable. Future work should explore ways to balance participant agency with efficient data collection.

4.3. Limitations

Digital literacy and internet access vary widely by demographic [22], influencing the generalizability of social media-based health research. Our sample was predominantly white and college-educated, with Facebook users significantly younger than non-users. While PD severity and duration did not differ between Facebook users and non-users, this finding may reflect selection bias, as individuals with milder disease may be more likely to participate in research. Additionally, the use of keyword-matching and traditional machine learning methods likely led to under detection of PD-related content. Future studies should investigate how digital behaviors shape participant engagement and consider approaches to enhance sample diversity and detection accuracy.

5. Conclusions

This study establishes Facebook as a feasible and ethically sound data source for PD research, demonstrating that individuals share PD-related information both before and after diagnosis. These findings highlight social media’s potential for disease monitoring and early detection. Further refinement of computational methods and integration with clinical data could enhance the utility of social media-derived insights.