Using patient-reported health data from social media to identify diverse lupus patients and assess their symptom and medication expressions: A feasibility study

Abstract

Objective:

Patient communities use social media for peer support and information seeking. This study assessed the feasibility of using public patient-generated health data (PGHD) from the social network Twitter to identify diverse lupus patients and gather their perspectives about disease symptoms and medications.

Methods:

We extracted public lupus-related Twitter messages (N=47,715 tweets) in English posted by users (N=8,446) in the United States between September 1, 2017, and October 31, 2018. We analyzed the data to describe lupus patients and the expressed themes (symptoms and medications). Two independent coders analyzed the data; Cohen’s Kappa coefficient was used to ensure interrater reliability. Differences in symptom and medication expressions were analyzed using two-tailed Z tests and a combination of one-way ANOVA tests and unpaired t-tests.

Results:

We found that lupus patients on Twitter are diverse in gender and race: about one-third (34.64%, 62/179) were persons of color (POC), and 85.47% were female. The expressed disease symptoms and medications varied significantly by gender and race. Much of our findings correlated with documented clinical observations, e.g., expressions of general pain (8.39%, 709/8,446), flares (6.05%, 511/8,446), and fatigue (4.18%, 353/8,446). However, our data also revealed less well-known patient observations, e.g., possible racial disparities within ocular manifestations of lupus.

Conclusion:

Our results indicate that social media surveillance can provide valuable lupus patient perspective data of clinical relevance. The medical community has the opportunity to harness this information to inform the patient-centered care within underrepresented patient groups such as POC.

Introduction

Lupus is an autoimmune condition characterized by chronic inflammation and intermittent acute flares resulting in progressive organ damage. It is estimated that at least 1.5 million people in the United States (U.S.) are affected by this condition, with over 16,000 new cases reported annually. An increased risk for lupus is related to gender, race, and age, with women of color being the most likely to develop the condition.¹ Lupus symptoms often mimic those of other diseases, making the condition difficult to diagnose. Despite recent advancements in serological and diagnostic testing, the length of time between symptom onset and diagnosis remains quite long.² Given that untreated lupus flares can result in organ damage, the field of lupus could greatly benefit from tools that identify people with lupus, help them manage and predict disease flares, and understand disease symptoms, treatment efficacy, and side effects.^3–5

Lupus is commonly diagnosed among younger individuals between the ages of 15 and 44, making social media a promising medium for connecting with this patient population, e.g., to provide health education. The self-reported data, also referred to as patient-generated health data (PGHD)⁶, generated by patients on public social media sites such Twitter and Reddit provide a new source of analyzable data. While traditional survey data can take years to collect, social media data is readily available and unaffected by recalling bias. It offers real-time insight into patient perspectives and behaviors. Harnessing self-reported patient data from social media to inform patient-centered care is an area of growing interest among medical professionals.

However, there are few social media and health surveillance studies with a focus on lupus patients. Two recent studies reported using PGHD from Twitter to understand the impact of the COVID-19 pandemic within patients diagnosed with rheumatic and/or musculoskeletal diseases.^{7, 8} Another study described how health conditions, including lupus, were represented on Facebook and showed that lupus patient support ranked highest among disease community support pages.⁹ A fourth study evaluated the reliability and quality of videos about SLE on YouTube, finding an abundance of videos with helpful information on SLE that was reliable and high quality.¹⁰

Our study aimed to assess the feasibility of using PGHD from Twitter for two purposes: (1) to identify diverse patients with lupus, and (2) examine their perspectives regarding disease symptoms, and medications. We chose to use data from Twitter for three reasons: (i) Twitter is among the most popular public social networks with a racially and ethnically diverse user base: 25% men vs. 22% women, as well as 23% Hispanic, 29% Black, and 22% White.¹¹ (ii) Research demonstrated the active use of Twitter among patient communities that share their disease experiences, e.g., cancer survivors.¹² These Twitter patient communities are easily identifiable through their use of hashtags: words or phrases preceded by a pound symbol that are used to identify tweets on a particular topic (e.g., #Lupus, #Spoonie #SLE, ). (iii) Health surveillance researchers have demonstrated the usefulness of Twitter data to understand public and patient perspectives on various diseases and health topics, e.g., COVID-19.¹³

Our central hypothesis was that patient-reported data from Twitter could be utilized to identify diverse lupus patients and provide insight into clinically relevant perspectives among this patient population.

Patients and Methods

Ethical approval

This study relied on publicly available Twitter data. The authors adhered to Twitter’s terms and conditions and privacy policy. Study-related data were collected using the secure Research Electronic Data Capture at the University of Southern California (USC). Any examples of Twitter account descriptions or posts included in this report were paraphrased to ensure user privacy. Study approval was obtained from the institutional review board (IRB) at USC (Protocol #HS-19-00048).

Study design and setting

We conducted a retrospective user and content analysis of public Twitter accounts and messages. Twitter data extraction were carried out at the Keck School of Medicine (SoM) at USC. The data analysis was completed at the Keck SoM at USC and Attune Health Research. The study protocol was published in the Journal of Medical Internet Research Protocols.¹⁴

Data collection and sampling strategy

The healthcare social media analytics platform Symplur Signals was used to search public Twitter data. The search terms were selected using an iterative process based on an established conceptual framework for social media data collection and quality assessment,¹⁵ expert input from rheumatologists, patient input, social media experts, and a search of related topics and hashtags in Symplur Signals. Included were Twitter posts in English that mentioned lupus-related keywords, phrases, and hashtags (see Supplementary Table 1) posted by users in the U.S. between September 1, 2017, and October 31, 2018. The location of the users in the U.S. was determined using a mapped location filter provided by Twitter GNIP through the “Profile Geo Enrichment” algorithm (formerly known as GNIP’s Profile Geo 2.0, which was acquired by Twitter).¹⁶

We applied a non-probability convenience sampling strategy using a nomothetic model of generalization¹⁷ by identifying the population (i.e., people with lupus) to which we aimed to generalize the results. The population is the totality of people with defined characteristics (i.e., lupus) to whom the study results are relevant. Although PGHD from Twitter was conveniently available, not every person with lupus (i.e., those not active on Twitter) had a chance of being included in our sample (non-probability sampling). Nonetheless, each person on Twitter that posted content about lupus had an equal chance of being included in the sample (probability sampling), thus, allowing us to reduce selection bias within our Twitter sample.

We attempted to understand lupus patient perspectives and used a hybrid approach of machine learning (ML) and qualitative research methods to confirm the reliability of the user account data included in the analysis. First, we verified the Twitter accounts were from human users (e.g., @JohnSmith) using the ML program Botometer¹⁸ (formerly BotOrNot) established by Indiana University. With a 95% success rate, the program identifies automated Twitter accounts, so-called bots, created by industry and interest groups to influence discussions and promote specific ideas or products.¹⁸ Botometer analyses multiple variables, including the account’s network (diffusion patterns), user (metadata), friends (account’s contacts), temporal amount (tweet rate), and sentiment (content of messages), detecting automated accounts.¹⁸ Second, we manually reviewed each Twitter account (N=1,576) to confirm that the account holder was a person living with lupus (see Supplementary Table 2). We reviewed the Twitter profile description (e.g., “Mom and wife living bravely with #SLE”) and the content from the posts in our dataset. During the manual review, two team members independently coded the demographic characteristics of the Twitter users using their profile description, username, and profile picture. Due to the limited demographic information on Twitter, we focused the analysis on gender and race (Person of Color: POC vs. White) (see Supplementary Table 2).

Data analysis

Coding:

Two research team members independently reviewed the user accounts and posts (thematic analysis) from lupus patients to classify the content based on a priori codes (see Supplementary Tables 2 and 3). To assess the interrater reliability, we used Cohen’s Kappa coefficient.¹⁹ A coefficient greater than 0.8 was considered substantial for this study. In instances where the coders disagreed, the primary investigators aided in making the final decision.

Statistical analysis:

We used descriptive statistics to describe our findings, e.g., symptoms and medications expressed by Twitter users with lupus. Differences in expressed themes were analyzed using two-tailed Z-tests. After dividing lupus medications into categories (see Supplementary Table 3), the number of posts for each treatment group were tallied, and a one-way ANOVA test was run to determine whether the variations were statistically significant. Unpaired t-tests were subsequently conducted between the groups to analyze differences between subgroups. All analyses were performed on Excel, Version 16.38.

Results

The initial dataset included 47,715 lupus-related tweets posted between September 1, 2017, and October 31, 2018 in English that originated from 1,576 Twitter users located in the U.S. After removing duplicates and posts from commercial and automated Twitter accounts, 40,885 lupus-related user messages remained (see Supplementary Figure 1). Of these Twitter accounts, 179 were determined to belong to people living with lupus (i.e., patients) (see Supplementary Table 4), contributing a total of 8,446 tweets (Table 1). Within this dataset, 7,752 posts (82.17%) corresponded to patients discussing their lupus symptoms and/or medications on Twitter.

Table 1 –

Demographic Characteristics of Twitter Users with Lupus (n=179)

Demographic Variable	Category	User Count (n=179)	User Proportion	Tweet Count (n=8,446)	Tweet Proportion
SEX	Male	18	10.06%	734	8.69%
	Female	153	85.47%	7587	89.83%
	Unclear	8	4.47%	125	1.48%
RACE	White	90	50.28%	5030	59.55%
	Person of Color	62	34.64%	2793	33.07%
	Unclear	27	15.08%	623	7.38%

Table illustrating sex and race breakdown of users determined to be Lupus patients, as well as their corresponding tweets.

Most accounts belonged to female patients (85.47%, 153/179), generating the majority of the posts (89.83%, 7,587/8,446), compared to Twitter accounts by male patients (10.06% of users, 18/179, and 8.69% of posts, 734/8,446). one-third of these users were POC patients (34.64%, 62/179), contributing about one-third of the posts (33.07%, 2,793/8,446). In comparison, roughly half of the users were White (50.28%, 90/179), accounting for more than half of the lupus-related Twitter posts (59.55%, 5,030/8,446) (Table 1).

Symptoms expressed by Twitter users with lupus

The most frequently mentioned symptoms (see Supplementary Table 5) in the tweets written by lupus patients were general pain (8.39%, 709/8,446 posts) and flares (6.05%, 511/8,446), with more than half of the patients mentioning them in a tweet at least once. Additionally, fatigue (4.18%, 353/8,446) and the inability to perform daily activities (3.11%, 263/8,446) were mentioned by more than 40% of users. Tweets with mentions of symptoms related to anxiety (1.21%, 102/8,446) and depression (1.00%, 85/8,446) were also frequent: about one-quarter of Twitter users with lupus mentioned them in a tweet at least once. Arthritis (2.78%, 235/8,446), skin sensitivity (1.78%, 150/8,446), muscle pain/myalgias (1.41%, 119/8,446), joint pain/arthralgias (0.95%, 80/8,446), self-consciousness (0.91%, 77/8,446), and symptoms related to “brain fog” – including confusion (0.91%, 77/8,446) and difficulty concentrating (0.90%, 76/8,446) ²⁰ – were mentioned in a tweet by over one-fifth of lupus patients.

Gender-specific expressions of symptoms among Twitter users with lupus

We found significant differences related to the gender-specific expression of symptoms (Table 2). Compared to posts from male Twitter users, posts from female users contained significantly more mentions of symptoms related to the inability to perform daily activities (p=0.0025), skin sensitivity (p=0.0167), depression (p=0.0163), and arthritis (p=0.0057). The expression of symptoms related to brain fog – specifically, confusion and difficulty concentrating – was also higher in posts written by female users (p=0.0062 and 0.0065, respectively). Twitter posts by male users showed a significantly higher frequency of fatigue expressions (p<0.0001); for all other comparisons, we found no significant differences.

Table 2 –

Gender-Specific Expressions of Symptoms Among 179 Lupus Patients using Twitter

Symptom	Twitter Post Count from Male Users (n=734)	Twitter Post Count from Female Users (n=7,587)	Z-value	95% Confidence Interval	p-value
General Pain	53 (7.22%)	631 (8.32%)	1	−0.0098 – 0.0318	0.3003
Flares	34 (4.63%)	454 (5.98%)	1.5	−0.0545 – 0.0651	0.1371
Fatigue *	52 (7.08%)	298 (3.93%)	4.1	0.0163 – 0.0467	<0.0001
Inability to perform daily activities *	9 (1.23%)	246 (3.24%)	3	0.007 – 0.332	0.0025
Arthritis *	8 (1.09%)	213 (2.81%)	2.8	0.005 – 0.0294	0.0057
Skin sensitivity *	5 (0.68%)	145 (1.91%)	2.4	0.0022 – 0.0224	0.0167
Muscle pain/myalgia	10 (1.36%)	107 (1.41%)	0.1	−0.0084 – 0.0094	0.9125
Anxiety	7 (0.95%)	95 (1.25%)	0.7	−0.0053 – 0.0113	0.4802
Corticosteroid treatment	4 (0.54%)	88 (1.16%)	1.5	−0.0017 – 0.0141	0.125
Depression *	1 (0.14%)	80 (1.05%)	2.4	0.0017 – 0.0165	0.0163
Joint pain/arthralgia	4 (0.54%)	71 (0.94%)	1.1	−0.0032 – 0.0112	0.2744
Biologic treatment	4 (0.54%)	72 (0.95%)	1.1	−0.0031 – 0.0113	0.265
Self-consciousness	4 (0.54%)	69 (0.91%)	1	−0.0034 – 0.0108	0.3047
Confusion *	0 (0.00%)	77 (1.01%)	2.7	0.0029 – 0.0173	0.0062
Difficulty Concentrating *	0 (0.00%)	76 (1.00%)	2.7	0.0028 – 0.0172	0.0065
Rash	3 (0.41%)	62 (0.82%)	1.2	−0.0026 – 0.0102	0.2291
Inflammation	1 (0.14%)	54 (0.71%)	1.8	−4e-04 – 0.0118	0.0685
Alopecia	6 (0.82%)	47 (0.62%)	0.7	−0.004 – 0.008	0.5157
Antimalarial treatment	7 (0.95%)	41 (0.54%)	1.4	−0.0016 – 0.0098	0.1611
Visual disturbance	2 (0.27%)	42 (0.55%)	1	−0.0027 – 0.0083	0.3163
Immunosuppressant treatment	3 (0.41%)	39 (0.51%)	0.4	−0.0044 – 0.0064	0.7141
Headaches	0 (0.00%)	39 (0.51%)	1.9	−1e04 – 0.0103	0.0525
Swollen joints	5 (0.68%)	25 (0.33%)	1.5	−0.001 – 0.008	0.131
Cannot plan activities or events	2 (0.27%)	28 (0.37%)	0.4	−0.0035 – 0.0055	0.6633
Fever	1 (0.14%)	27 (0.36%)	1	−0.0022 – 0.0066	0.3286
Limited in fulfilling family responsibilities	2 (0.27%)	24 (0.32%)	0.2	−0.0037 – 0.0047	0.8176
NSAID treatment	0 (0.00%)	25 (0.33%)	1.6	−8e-04 – 0.0074	0.1191
Digestive issues	0 (0.00%)	23 (0.30%)	1.5	−0.001 – 0.007	0.1373
Migraines	0 (0.00%)	22 (0.29%)	1.5	−0.001 – 0.0068	0.144
Seizures	0 (0.00%)	25 (0.33%)	1.6	−8e-04 – 0.0074	0.1191

^*Statistical significance at p-value <0.05

Table illustrating the results of two-tailed Z tests to determine whether differences in the frequencies of expressed lupus symptoms between male and female patient-generated tweets were statistically significant.

Symptoms expressed by Twitter users with lupus of different race

We also found significant differences in the expression of symptoms within posts written by White versus POC Twitter users with lupus (Table 3). Symptoms that were discussed more frequently in tweets generated by White users included general pain (p<0.0001), flares (p=0.0223), arthritis (p<0.0001), skin sensitivity (p=0.035), and rash (p=0.0379). In contrast, inflammation (p=0.0339), self-consciousness (p=0.0006), alopecia (p<0.0001), visual disturbances (p=0.0013), confusion (p=0.0149), and difficulty concentrating (p=0.0268), were discussed more frequently by POC. No other significant differences were observed.

Table 3 –

Race-Specific Expressions of Symptoms Among 179 Lupus Patients using Twitter

Symptom	Twitter Post Count from White Users (n=5,030)	Twitter Post Count from POC Users (n=2,793)	Z-value	95% Confidence Interval	p-value
General pain *	510 (10.14%)	115 (4.12%)	9.4	0.0477 – 0.0727	<0.0001
Flares *	304 (6.04%)	134 (4.80%)	2.3	0.0018 – 0.023	0.0223
Fatigue	210 (4.17%)	111 (3.97%)	0.4	−0.0072 – 0.0112	0.669
Inability to perform daily activities	156 (3.10%)	74 (2.65%)	1.1	−0.0033 – 0.0123	0.2589
Arthritis *	201 (4.00%)	9 (0.32%)	9.6	0.0293 – 0.0443	<0.0001
Skin sensitivity *	99 (1.97%)	37 (1.32%)	2.1	5e-04 – 0.0125	0.035
Muscle pain/myalgia	81 (1.61%)	30 (1.07%)	1.9	−1e-04 – 0.0109	0.0529
Anxious	62 (1.23%)	34 (1.22%)	0	−0.005 – 0.0052	0.9693
Depression	49 (0.97%)	23 (0.82%)	0.7	−0.0029 – 0.0059	0.5047
Joint pain/arthralgia	41 (0.82%)	28 (1.00%)	0.8	−0.0025 – 0.0061	0.4125
Self-consciousness *	28 (0.56%)	36 (1.29%)	3.4	0.0031 – 0.0115	0.0006
Confusion *	38 (0.76%)	37 (1.32%)	2.4	0.0011 – 0.0101	0.0149
Difficulty concentrating *	39 (0.78%)	36 (1.29%)	2.2	6e-04 – 0.0096	0.0268
Rash *	47 (0.93%)	14 (0.50%)	2.1	2e-04 – 0.0084	0.0379
Inflammation *	25 (0.50%)	25 (0.90%)	2.1	3e-04 – 0.0077	0.0339
Alopecia *	14 (0.28%)	32 (1.15%)	4.8	0.0052 – 0.0122	<0.0001
Visual disturbance *	18 (0.36%)	26 (0.93%)	3.2	0.0022 – 0.0092	0.0013
Headaches	28 (0.56%)	8 (0.29%)	1.7	−4e-04 – 0.0058	0.0921
Swollen joints	18 (0.36%)	9 (0.32%)	0.3	−0.0023 – 0.0031	0.7727
Cannot plan activities or events	15 (0.30%)	10 (0.36%)	0.4	−0.002 – 0.0032	0.6533
Fever	21 (0.42%)	5 (0.18%)	1.8	−3e-04 – 0.0034	0.0781
Limited in fulfilling family responsibilities	15 (0.30%)	5 (0.21%)	0.7	−0.0015 – 0.0033	0.4606
Digestive issues	15 (0.30%)	7 (0.25%)	0.4	−0.002 – 0.003	0.6896
Migraines	16 (0.32%)	4 (0.14%)	1.5	−5e-04 – 0.0041	0.131
Seizures	15 (0.30%)	4 (0.14%)	1.4	−7e-04 – 0.0039	0.1684

POC: Person of Color

^*Statistical significance at p-value <0.05

Table illustrating the results of two-tailed Z tests to determine whether differences in the frequencies of expressed lupus symptoms between White and POC patient-generated tweets were statistically significant.

Medications expressed by Twitter users with lupus by gender and race

We further examined the expressions of medications among Twitter users with lupus. Utilizing expert input from rheumatologists, we compiled a list of the most commonly used medications in lupus patients (Supplementary Table 3). Of the 179 Twitter users in our dataset, 34.6% (62) wrote one or more posts mentioning medications, and tweets containing mentions of medications accounted for less than 1% of the total tweets written by users living with lupus. The most frequently mentioned medications among these lupus patients were corticosteroids (18.99%, 34/179 users), antimalarials (15.08%, 27/179 users), biologics (13.97%, 25/179 users), immunosuppressants (10.61%, 19/179), and NSAIDs (7.82%, 14/179) (see Supplementary Table 6).

Comparing medication expressions by user gender and race showed that White female Twitter users posted most frequently about medications, generating an average of 19.11 posts per treatment category, followed by POC female users who generated an average of 10.44 related posts (Table 4). When analyzing gender and race-related differences in the number of tweets mentioning medications, we found statistically significant differences (p=0.00412) between all groups examined (i.e., White males, White females, POC males, and POC females) across all nine treatment categories (Table 5).

Table 4 –

Gender & Race Differences in Medication Expression, Counts of Twitter Posts Written by Users with Lupus

Medication	Male White	Male POC	Female White	Female POC	Total
Immunosuppressant	3	0	21	16	40
Corticosteroid	3	1	55	28	87
Biologics	4	0	45	23	72
NSAID	0	0	18	6	24
Antimalarial	5	2	26	11	44
Anticoagulant	0	0	0	2	2
JAK Inhibitor	0	0	0	1	1
TNF Blocker	0	0	0	1	1
Other Treatment	0	0	7	6	13
Total	15	3	172	94	284
Average	1.67	0.33	19.11	10.44	31.55

Table illustrating the number of tweets written by users with lupus that mentioned medication categories, stratified by gender and race. The average number of tweets per medication category was also included for each race and gender group.

This analysis only included tweets corresponding to users whose gender and race were able to be identified.

Table 5 –

ANOVA Analysis of Differences in Medication Expression in Tweets Written by Users with Lupus by Gender & Race

Source of Variation	SS	df	MS	F	p-value
Between Groups	2051.44	3	684.81	5.3749	0.00412 *
Within Groups	4077.11	32	127.41
Total	6131.55	35

Table illustrating the results of the one-way ANOVA test of variance used to determine whether differences between users of differing gender and race were statistically significant with respect to average number of tweets per medication type.

^*Statistical significance at p-value <0.05

When keeping the user race constant, we found significant differences in the frequency of medication mentions among both White and POC users (p=0.0201 and p=0.0076, respectively) in favor of females tweeting more often (Table 6). Similarly, we found statistical significance when comparing White male and POC female users (p=0.0193), as well as White female and POC male users (p=0.0131).

Table 6 –

Unpaired t-test of Differences in Medication Expression in Twitter Posts by Users with Lupus by Race and Gender

	Standard Error of Difference	95% Confidence Interval	t	df	p-value
Female White & Female POC	7.490	−7.21 to 24.54	1.1571	16	0.2642
Female White & Male White	6.757	3.12 to 31.77	2.5818	16	0.0201 *
Female White & Male POC	6.726	4.52 to 33.04	2.7920	16	0.0131 *
Male White & Male POC	0.726	−0.21 to 2.87	1.8358	16	0.0851
Male White & Female POC	3.375	−15.93 to −1.62	2.6007	16	0.0193 *
Female POC & Male POC	3.313	−17.13 to −3.09	3.0520	16	0.0076 *

^*Statistical significance at p-value <0.05

Discussion

The primary findings of this study demonstrated that PGHD from social media could be used to identify people with lupus and study their perspectives regarding disease symptoms and medications. Our study further showed that the identified lupus patients are diverse in gender and race: most Twitter users with lupus were identified as female patients, and they contributed a vast majority of the tweets analyzed. Additionally, about one-third of the patients were POC, introducing new opportunities for using social media surveillance to gain insight into patient perspectives of hard-to-reach and underserved patient groups.

Social media can also be leveraged as a tool for educating and connecting with patients. For example, a recent study successfully used social media to deliver lupus-related educational resources to Latin American patients, generating high patient engagement and satisfaction.²¹ Further, Twitter has also been employed to inform patients of emerging research and medical developments.²² Future research should characterize the communication patterns of diverse lupus patients on different social media sites (e.g., Reddit) and continue to examine the potential for engaging this patient population through these media.

It is noteworthy that our findings corresponded with previous clinical observations but also revealed less well-documented patient perspectives. The male-to-female ratios of posts and lupus patients on Twitter were consistent with what is observed clinically. Assessments of overall health found that female lupus patients are more likely to report having poor or fair health,²³ which is reflected in our findings: female Twitter users with lupus posted more frequently regarding flares and the inability to perform daily tasks due to their condition. We also observed more frequent posts about skin sensitivity and arthritis among female users. This is clinically notable, given that female SLE patients present more frequently with malar rash and arthritic symptoms.²⁴ With regard to SLE-associated neuropsychiatric symptoms, we found that female users more frequently posted about depression symptoms. At the same time, however, male patients are more reluctant to report symptoms related to psychological issues.²⁵

Social media surveillance on Twitter could potentially aid in improving the early detection of lupus through an improved understanding of the gender and racial dynamics of symptoms and medication use. We observed significant variations in the expression frequency of visual disturbances in favor of POC. While ocular manifestations of lupus have been well-documented²⁶, there is little clinical data on potential racial disparities regarding this symptom. Moreover, these symptoms are associated with system-wide inflammation that can be fatal. Therefore, immediate recognition is vital for improved disease outcomes.²⁶ Further research should determine whether our findings are consistent with clinical observations.

Finally, nonadherence to medications, which can manifest as poor execution of dosing schedules or discontinuation of medications, is a pervasive issue within SLE patients, particularly for immunosuppressive and antimalarial treatments.²⁷ In lupus patients, this practice is associated with a higher risk of flares, morbidity, hospitalizations, and poor renal outcomes, which requires the development of targeted interventions to improve medication adherence.²⁸ Social media such as Twitter could aid in the dissemination of such interventions. Among adolescents and young adults, social media-based interventions have been associated with increased medication adherence and improved health outcomes.²⁹ Another potential application concerns the association between depression and patient noncompliance³⁰. Twitter could be used for patient surveillance and possible engagement. The feasibility of identifying individuals with depression on Twitter has already been demonstrated.³¹ Our data demonstrated that lupus patients discuss symptoms of depression on Twitter, indicating its potential as an intervention.

Limitations of this study are inherent to observational investigations, which mandate caution in interpreting the findings. We acknowledge the lack of lupus patient perspectives from other social networks and the wider, offline lupus patient communities. The external validity of this study was further limited because we only included a sample of posts in English by individuals that we could identify as lupus patients based on their profile descriptions and tweet content. The data provided here should be considered a foundation for generating hypotheses for further evaluation within the context of clinical expertise and patient preferences.

Additionally, user the demographic classification was based on self-reported information from these Twitter users in their profile description and tweets. Although we excluded data from users who did not clearly state their lupus status, there is a risk of misclassification, particularly if users provide false information. However, we are not aware of any research that suggests patient imposters on social media: that is, individuals falsely posing as patients.

In summary, PGHD from people with lupus on Twitter can complement other efforts to collect patient-reported health outcomes and generate a richer picture of how symptoms, medications, and other aspects of the disease experience are being perceived by diverse patient groups. Patient-reported data from Twitter revealed both clinical aspects that are well-documented as well as less well-known patient experiences, which support hypothesis generation and warrant further examination. Twitter as a platform for examining patient perspectives and health outcomes should be studied further as a unique and complementary source of information. Integrating the patient perspective from social media may offer additional dimensionality to the routine monitoring of health outcomes and drug safety, as well as more broadly capture symptoms or experiences relevant to patients that may otherwise remain under-recorded.

To date, the scientific examination of PGHD from social media is underrepresented in rheumatology research.³² The medical community has the opportunity to harness this information to inform the patient-centered care within underrepresented patient groups such as POC. As demonstrated in other disease contexts, surveillance research on Twitter and other public social media can reveal clinically relevant emerging patient needs, adverse drug reactions, and patient safety issues. ^{33, 34}

Key points.

• Researchers increasingly use patient-generated health data (PGHD) from social networks such as Twitter in health research studies. However, research using PGHD is currently underrepresented in the field of rheumatology.

• This study demonstrates that Twitter provides clinically relevant data from diverse lupus patients, including their views on disease symptoms and medications.

• Much of the findings were consistent with documented clinical observations, but the data also revealed racial disparities in the expression frequency of ocular lupus symptoms, of which there is little available clinical data.

• Our work suggests that social media surveillance could provide a real-time and readily available data source for collecting self-reported patient data from diverse lupus patients to improve patient-centered care.