Recovery is a behavioral health process among individuals affected by substance use and substance use disorders (SUDs) [1]. In 2012, the Substance Abuse and Mental Health Services Administration developed a new working definition of recovery, defining the process as “a…change through which individuals improve their health and wellness, live self-directed lives, and strive to reach their full potential” [2]. This definition shifts away from the traditional emphasis on abstinence along with symptom mitigation and reframes recovery as a holistic, multidimensional, and humanistic process [3]. By emphasizing overall well-being, the Substance Abuse and Mental Health Services Administration broadens the focus to include improvement in multiple life domains for individuals affected by substance use or other similar challenges. This shift underscores the importance of individualized and comprehensive recovery pathways that are accessible and not time-bound and foster supportive environments to promote recovery outcomes [4,5].

Emphasizing recovery is critical given the evolving US substance use crisis, which affects individuals across age groups and demographics [6]. As many as one-third of US adults reported an SUD in the past year, and the number of people facing substance-related challenges is growing substantially [7,8]. Yet, in the United States, most adults with SUD do not receive treatment [6] due to well-documented barriers to care that include cost, geographic inaccessibility, perceived stigma, and the availability of co-occurring disorder treatment [9-11]. Novel strategies that address this crisis with approaches that promote engagement in recovery by meeting people where they are in ways that are both accessible and intuitive are needed.

Growing evidence suggests that social media forums and smartphone apps may offer practical and cost-effective strategies for behavior change, which can significantly decrease substance use [12-14]. Some of the most frequently deployed behavior change strategies used on these forums and apps include self-monitoring, personalization, reminders [14], and online peer coaching [15]. These app features enable users to tailor the app’s functionality to meet their individual needs, resulting in personalized recovery plans, increased app use, prolonged engagement in recovery, and reductions in substance use [13,14,16,17].

Social media platforms, and their respective smartphone apps, also help facilitate social support for substance use and mental health recovery [18-21]. Previous research shows, for example, that online recovery forums on Reddit can provide a sense of belonging and provide practical advice for individuals with SUD through interactions with others who share lived experiences and peer coaches trained to discuss substance use virtually [19,22]. Additionally, hashtags and content related to recovery on platforms such as X and TikTok can raise awareness; reduce stigma; and connect individuals with resources, professional counseling, and other support networks [18,21]. Indeed, these virtual spaces allow people to share their experiences and seek guidance from others who have faced similar challenges, which can potentially support the engagement and connection necessary for promoting behavior change [23].

Other digital smartphone apps, which include closed social media–like platforms designed specifically for people in recovery, provide accessible, highly specific, and flexible opportunities for individuals to engage with recovery networks, often bypassing barriers known to impede care such as stigma [24]. Research likewise suggests that individuals are open to the idea of using social media or similar apps for recovery support, sometimes citing convenience and semianonymity in these spaces [25,26].

Recovery-specific digital platforms or apps—that is, InTheRooms and The Phoenix—have emerged as online resources beyond conventional social media designed to foster a community among people with lived experiences of substance use [23,25,27,28]. Although more research is needed on the long-term impacts of these apps, early indications suggest that they offer recovery-specific tools that provide emotional and practical support, paralleling the benefits of in-person interactions [29].

Individuals wanting to change behaviors may seek online and offline resources to initiate or sustain recovery from substance use and related challenges [30-32]. This study examines The Phoenix: A Sober Community, a nonprofit organization with a nationwide reach that developed a free mobile app (hereafter referred to as The Phoenix app, now known as NewForm). The app was designed to facilitate access to social support across geographical and temporal boundaries. It has since evolved into a robust platform that allows members and volunteers to connect based on shared interests or locations (eg, sober parents and guardians, veterans, arts, and crafts). It includes features such as live-streamed and in-person events, recovery journey tracking, volunteer coordination tools, and peer-to-peer communication.

As a social media app with a closed environment (ie, tailored for a specific purpose), understanding the app’s thematic and emotional digital ecosystem may offer additional insights into the role of digitally delivered peer- and volunteer-mediated interactions in promoting a safe and inclusive recovery environment. This study aims to identify the thematic and emotional climate of The Phoenix app using the in-the-moment interactions from anonymized app member data. Our study is guided by 3 research questions (RQs):

The findings from this study will contribute to the growing body of evidence calling for directed and curated app experiences to promote behavior change in the context of substance use. In particular, the methods, or pipeline, undertaken in this study will also illustrate their effectiveness in highlighting app functionality along several critical domains, including language (ie, words and phrases used on mobile apps), emotions (ie, valence patterns used in mobile apps), and behaviors (ie, identifying observations in which recovery outcomes are salient). Collectively, this pipeline—which uses user-generated content provided by The Phoenix—will further document successful strategies necessary to sustain healthy digital environments for substance use recovery.

Because this study involves the analysis of deidentified and secondary data, it was deemed as nonhuman subjects research and did not require further review by the institutional review board at Indiana University Bloomington (#24091). Given the sensitive nature of the data, preliminary findings from this study were shared with The Phoenix app members through open-access listening sessions. Any excerpts published in this report have been abridged slightly to uphold app member privacy by removing potentially identifiable information, graphic depictions of substance use, and other personal details.

The steps followed in this study are shown in Figure 1. We first preprocessed the data to ensure standard and consistent entries. For our BERTopic analysis, preprocessing steps included removing parts of speech with no clear semantic meaning, such as (1) articles, (2) prepositions, (3) punctuations, (4) abbreviations, (5) any special characters, and (6) numbers. After preprocessing, we ran iterative BERTopic analysis, testing a range of possible topic solutions (k=5…50) to find an optimal topic number. We calculated a coherence score with each iteration as an indicator of model fit, where coherence scores reflect the human interpretability of a fixed set of topics. For each iteration, we plotted each coherence score, with the highest coherence score indicating the optimal solution. Upon running 50 topic iterations, the topic solution with the highest coherence score was 10 (coherence=0.48). Once we identified a topic solution, we applied UMAP and HDBSCAN to reduce dimensions and extract topics with key words (eg, topic 1: Word a, Word b, Word c…Word d). We then used a simple sorting function to match each app entry to a topic based on the overlap of words used in the entry to topic key words. After sorting our topics, the members of the research team met briefly to agree on an interpretation of each topic.

For the VADER analysis, we fed minimally processed text through the VADER lexicon, retaining stop words, punctuations, numbers, and special symbols and characters. Because VADER requires exact-case matching to return an accurate score, that is, VADER can only provide a score if the spelling is correct, we used TextBlob and a loop function to replace apparent typos such that “angryyyy” is automatically corrected to its apparently correct spelling of “angry.” Once each entry received a VADER score and Text2Emotion profile, we created subdatasets consisting of highly negative entries (−0.99 to −0.50) and highly positive entries (0.50 to 0.99), which were reviewed by the research team to ascertain themes.

Broadly, our topic modeling analyses found various topics specific to promoting social connection; our sentiment analysis revealed how people utilize the app to seek or provide social support to those struggling with their recovery journeys. We present the findings from our analysis below.

From the iterative BERTopic analysis, it was determined that 10 topics represented the optimal solution with a coherence score of 0.48. We then extracted top terms per topic and their representation as a whole number and percentage of the larger corpus (Table 1), and based on topic keywords, we created a custom name for each of the 10 topics. Topics extracted from our BERTopic analysis revealed what appeared to be 2 overarching themes: (1) a group of topics pertaining to engaging members through in-person and online interactions (7 topics) and (2) as a forum to discuss more serious topics pertaining to substance use and mental health, including struggles with sobriety, relapse, and strategies to promote substance use reduction or abstinence and improve mental health (3 topics). These overarching themes are again observed in Figure 2, which visualizes an intertopic correlation as an intertopic distance map. Dense and closely knit circles (ie, topics) indicate high topic similarity, while disparate topics represent dissimilarity. The topics in the upper left-hand quadrant of the intertopic distance map largely pertain to app dialog about promoting in-person and online events, which ranged from hikes (eg, “Don’t forget about The Phoenix hike tonight at 8 pm!”), workouts (eg, “See you tonight at...the gym for a great workout session”), online and offline poetry sharing (eg, “I am pasting my poem here; let me know what you think”), and other events deemed part of community building among app members. In contrast, the topics in the bottom-right quadrant, while sharing some overlap with the topics in the upper left quadrant, largely pertained to sharing personal experiences; narratives; or insights into substance use, recovery, and substance relapse.

Upon reviewing topic keywords and intertopic similarities, we then used a sorting function to sort posts into one of the 10 possible topics based on keyword matching. From this sorting, we were able to identify a series of optimal quotations, which added further context and nuance to the topics explained above (Table 2). Table 2 shows the actual conversations unfolding with greater nuance from the dataset.

The excerpts in Table 2 highlight representative themes, discussions, or comments within our sample. These excerpts were identified using a sorting function that categorized app posts into one of the 10 topics identified by the BERTopic tool. These quotes demonstrate a distinct and recurring theme of community support. This support manifests through encouragement, the promotion of online activities, and in-person events, including hiking, physical activity, mindfulness, meditation, and other similar activities that promote recovery through movement and community engagement.

To further substantiate the topic modeling findings, which were used to identify recurring themes in the social media entries collected from the dataset, we applied VADER sentiment analysis to all collected entries, excluding duplicates and retaining “zero” values. Table 3 provides an overview of our findings, highlighting the overall mean VADER score for the entire sample, with a possible range of −0.99 to 0.99. It also includes a summary of highly positive entries—those scoring above 0.50—and highly negative entries—those scoring below −0.50.

The mean VADER score of our entire sample, after eliminating duplicate entries, was 0.44 with an SD of 0.42. This suggests that, overall, valence, or affect, was highly positive with notable variation across all entries. The median and quartile regions, however, strongly imply a right-skewed distribution, implicating higher levels of positive content. When we further subdivided our sample to extract highly positive and highly negative entries, we observed that over half (52.2%) of our sample contained VADER scores greater than 0.50. Conversely, only 429 entries, or 2.4% of our data, scored below –0.50.

To validate our VADER findings, we ran a Text2Emotion analysis with the overall data. Unlike VADER, which provides a single composite score that ranges from –0.99 to 0.99, Text2Emotion measures 5 emotion categories: (1) Happy, (2) Sad, (3) Surprise, (4) Angry, and (5) Fear. It then scores them as normalized values that sum to 1 for each entry. These scores reflect the intensity of each emotion expressed. Across all entries, “Happy” was the most prominent emotion (mean 0.25), followed by “Fear” (mean 0.19), “Sad” (mean 0.17), and “Surprise” (mean 0.16), while “Angry” was the least prominent emotion (mean 0.04). As with VADER, we noticed some potential misclassification (eg, “That workout killed me today“ scored high for “fear”). However, the findings from VADER and Text2Emotion both implicate the presence of highly positive content.

Table 4 provides 5 abridged excerpts of highly negative and highly positive posts. From these excerpts, we observed clear differences in the content. For highly positive excerpts, we observed people expressing gratitude for their sobriety and mental health; appreciation for The Phoenix, the community, and other support structures; and promotion of in-person and online events available in locations across the United States. As seen in Excerpt: “I am just so thankful to be alive today” and “Just had a great workout, make sure you stop by if you are in the [redacted] area!,” the presence of positive language with words including “thankful,” “alive,” “great,” and phrases such as “just had a great workout” contribute to higher than average VADER scores. Conversely, negative entries seemed to exhibit 1 recurring theme—individual narratives about struggles with sobriety, relapse, mental health, or voicing needs for additional support.

It is also worth noting that some negative entries concluded with words of encouragement, as seen in the excerpt: “If I can get through it, so can you.”

Our study used topic models and sentiment analysis to examine recurring themes, uses, and valences of randomly selected entries from the dataset. Using NLP, it is possible to increase the granularity of analysis to triangulate phrasings counterintuitive to recovery. Using sentiment categories and strategic queries, we attempted to allocate content that may imply highly negative experiences, warranting response or intervention. For example, we queried the data to isolate all entries containing the phrase “I need help,” which also contained a negative VADER sentiment value (ie, values between –0.01 and –0.99). We also searched a few additional phrases that, in the psychology literature, indicate personal rumination, worry, or distorted thinking, including “I am a”—a type of distorted thinking characterized by labeling oneself and “I will never”—a type of distorted thinking characterized by overgeneralizing or fortune-telling. We hypothesized that people using these phrases in a negative context reflected some distress or worry, should they appear in the dataset.

Although these queries returned very few entries, as seen in Table 5, they illustrate how some individuals used the app to seek support from other app members. In this context, seeking support reflected a general struggle with one’s current recovery journey. At the same time, we observed individuals posting poetry, which may not directly reflect personal experiences with substance use, mental health, and recovery.

This study tested a computational NLP pipeline to analyze randomly allocated posts collected from The Phoenix app. Broadly, our findings suggest that app members use the app to promote social connection and community. These findings align with a growing body of literature indicating that mobile apps are effective systems for delivering behavior change interventions. We provide additional context below.

Numerous studies use VADER as a proxy for public mood or subjective well-being and as a primary measure to understand social attitudes expressed on social media during major public health events [38,40-42]. These studies typically involve the composite analysis of very large collections of posts from various social media platforms such as X, Reddit, and YouTube, often around a singular fixed topic ranging from substance use and recovery [38], reproductive rights and access [43], collective mental health [44], and others.

Although our study employs a similar use of VADER (ie, as a tool to measure the mood of text data), we observed key differences in our findings that diverge from previous studies. Indeed, prior studies using VADER sentiment analysis in both health and social media contexts sometimes report a mean VADER score near 0 [43]. Distributions of these VADER scores are commonly bimodal [45], indicating a near-equal presence of highly negative and highly positive content. When calculating a mean VADER value, this bimodal distribution can sometimes yield a mean VADER value near 0 with a high standard deviation, making it challenging to draw reasonable conclusions about collective mood, well-being, or valence regardless of the topic—though more research on this topic is needed that addresses VADER limitations is needed.

However, in our analysis of over 17,000 recovery-related entries, we observed a mean VADER score of 0.44, reflecting a generally high presence of positive content. A more detailed review of the descriptive statistics revealed that the VADER score distribution was not bimodal but positively skewed. This finding is further supported by the relatively low number of highly negative posts (ie, those scoring lower than –0.50), which comprised just over 2% (n=429) of the corpus. In contrast, highly positive posts (ie, those scoring higher than 0.50) comprised more than 50% (n=9363) of our sample.

The general positive tone uncovered by VADER is further supported by the Text2Emotion analysis—where “happy” was the most dominant emotion—and themes uncovered in our BERTopic analysis. Some themes broadly referenced various activities and interests, while others appeared to originate from volunteers marketing specific online and in-person events to the membership. These themes may implicate online social connection, which can drive people to continue using the app as part of their recovery process.

Critically, we observed 1 topic that referenced relapse and sobriety struggles. This topic illustrates how, even if in very small quantities, people are willing to discuss their struggles openly, which we interpreted as an effort to seek support. Collectively, these topics offer evidence that, as a recovery smartphone app, curated app experiences may offer more directed discourse needed to support and sustain recovery. Future research should compare these findings with data from other similar recovery apps to provide further evidence that digital tools offer cost-effective approaches to promoting recovery regardless of geographic access.

In recent years, a growing body of literature has explored social media’s role in promoting social connection, particularly in relation to health and well-being [32,46-48]. The findings from that work highlight that the degree and quality of social connection tend to vary and often depend on the specific platform, medium of the platform, and context.

There are numerous examples of how social media can facilitate social connections. Studies on Facebook groups have found that the degree of connection offered was particularly beneficial for older adults [47] and non-native English speakers [49]. These groups are also the most dominant Facebook user base [49]. On X, before the 2022 acquisition by Elon Musk, people could search for communities by grouping related content through hashtags, which created pockets of conversation around specific topics. Reddit remains widely regarded as one of the strongest conventional social media platforms for fostering social connections through curated subreddits with trained moderators [50]. However, recurring issues across these platforms remain, such as low-quality health information, trolling, and exposure to harmful content counterintuitive to recovery [51].

Prior studies have explored the strengths and weaknesses of social media for recovery using conventional social media data, such as X or Twitter, including how communities form around topics such as alcohol abstinence during Dry January [52] and how others communicate about personal and familial drug use online [53]. While platforms such as Facebook, X, and Reddit have been leveraged to connect people with shared experiences, their general nature may not always yield positive experiences. This is potentially due to the overwhelming number of unique posts across a wide range of topics, a majority of which do not pertain to health and well-being, or inconsistent moderation that filters negative content.

To offset negative social media interactions, some argue that the approaches to safeguard spaces via social media–like mobile apps can create a curated social media experience tailored to specific needs, thereby enhancing both the quality of perceived social connection and perceived benefits of support communities [54,55]. The Phoenix app exhibits the characteristics of effective safeguarding through advanced monitoring and interactions between members and volunteers. The more curated nature of the app—designed with the express purpose of connecting people in recovery—allows for a more directed experience that reaches its intended audience.

Further evident in our data, which included topics related to social interaction and recovery, and sentiment analyses indicating high levels of app positivity, the app displays early signs of a supportive and balanced social media ecosystem where safeguarding is effective. By safeguarding the app, app members are protected from incongruent content that may be encountered on conventional social media platforms, creating a safer experience that builds trust among app members and volunteers.

Our findings suggest that digital recovery apps can promote a sense of community and belonging among people in recovery from substance use. To substantiate these observations, we applied an unsupervised NLP framework, which has been used in previous studies to contextualize how people express themselves about certain health topics or leverage online communities for social support or information seeking [32]. Collectively, these findings contribute to prior work calling for directed app experiences for behavior change, though additional research could add further context to our findings to address potential challenges associated with social apps designed with specific populations and objectives in mind. Still, additional work with and beyond the Phoenix is necessary.

Future research should strongly consider the following. First, The Phoenix app is distinct from other social media platforms in its emphasis on safeguarding and advanced monitoring to filter harmful content. While our sample did not identify any content deemed counterintuitive to recovery, further analysis of larger Phoenix datasets could provide deeper insights into the effectiveness of these processes—particularly for app members early in recovery or other vulnerable groups who are likely at a greater risk of negative effects from engaging with triggering or harmful content (ie, veterans and LGBTQIA-identifying members). From a cross-disciplinary perspective, additional work with other computer science fields, including machine learning and artificial intelligence, can help develop additional tools to directly intervene when someone posting on The Phoenix app writes in distressing tones. Qualitative interviews and codesign sessions with app members can also help explore personal experiences with the app’s safety and support features, especially when compared with other mobile apps with a similar focus. Critically, any insights generated from that work would accomplish 2 tasks: (1) establish software-specific best practices for recovery-related mobile apps and (2) identify nuances in recovery communication across online versus conventional, residential programs.

Second, while our analysis points to the success of the app in fostering a supportive recovery environment, our findings have only been compared to prior studies using conventional social media data on substance recovery. Future studies should consider conducting comparative analyses of recovery forums across curated recovery apps, specifically using both qualitative and computational methods. An across-app comparison could add additional validity to our findings, particularly regarding the relative strengths of resources and tools offered by the Phoenix relative to tools and resources offered on apps from other recovery groups and organizations. These additional research directions will strongly contribute to our understanding of smartphone apps’ roles and functions as promoters of positive behavior change and lend further evidence to optimal smartphone app designs to promote sustained engagement in recovery.

This study is subject to limitations we hope to address in future work. First, our study involved a cross-sectional analysis of randomly allocated posts derived from The Phoenix app. While our findings offer preliminary insights into the use and purpose of the app, it is likely that our data are not entirely reflective of in-the-moment discourse that may demonstrate other app utilities and uses. For example, the dataset contained specific channels or communities directed at particular topics and experiences, including pets, art, and others. Although it is possible our data may be overrepresented by one or a few channels, it is not currently possible to assess this with the specific sample. We strongly encourage future research to use our pipeline with additional Phoenix data that removes, or at least compares, volunteers to app members. Second, we acknowledge that our NLP approach is entirely unsupervised, meaning no training data were leveraged. While this may limit the novelty of our approach, we contend that the analyses used provide a strong scaffold that is necessary to design, program, and ideally, implement more sophisticated tools for recovery, including knowledge graphs, retrieval-augmented generation systems, and fine-tuned large language models. Third, we also acknowledge that, at the time of writing, The Phoenix app has been rebranded as NewForm. Although The Phoenix app and NewForm are similar in functionality, it is likely that newer features may result in marginal variations in our findings.

This study tested a computational pipeline aimed at contextualizing the digital environment of a recovery-focused app developed by The Phoenix, a national nonprofit with over 65,000 monthly active members. Notably, our pipeline aimed to extract language, emotions, and behaviors expressed by app members to ascertain app use and its impact in fostering recovery support in a digital environment. These findings—which illustrate the broadly positive uses of an app as evidenced by the themes and emotions expressed by app members—add further evidence to the growing body of research calling for online recovery resources to address the chronic shortage of costly residential treatment options. Further, these findings likewise implicate key recovery dynamics in digital spaces—including how people in active recovery discuss their journeys, support one another, or share resources. Extracting and modeling these insights can further aid in the design of more robust tools aimed at keeping people in recovery by meeting them where they are in their journeys.