Swiss MS Registry (SMSR)–a longitudinal Citizen Science study

The SMSR is an ongoing longitudinal patient-centered study in Switzerland funded by the Swiss MS Society (https://clinicaltrials.gov/ct2/show/NCT02980640) [11,29]. The study has been approved by the Ethics Committee of the Canton Zurich (PB-2016-00894, BASEC2019-01027). All SMSR participants have provided written informed consent. The SMSR pursues the overarching goal of giving a voice to persons with MS. The SMSR actively collaborates with its participants on various projects as the persons with MS themselves are the best experts on their disease. A detailed description of how the SMSR collaborations with its participants is provided in the Supplement (S1 Text).

The ‘My life with MS’ project

The ‘My life with MS’-project has been developed jointly by scientists and persons with MS. The initial project idea has been sparked by the SMSR’s board consisting of approximately 30 persons with MS that continuously advise on the SMSR’s activities. Part of the board’s role is also to bring up themes that are important for persons with MS and not sufficiently covered in the scientific landscape. Given their relevance and the deficient state of research, the board’s members expressed the wish to research individual-level stories of MS in depth from a lifetime perspective. The SMSR subsequently developed the ‘My life with MS’-project in a multistage, iterative process. The project conceptualization phase included two workshops with persons with MS who contributed to project aims and design, reviewed and edited assessment materials, and provided general feedback. The experiences of the board members and the SMSR’s scientific expertise ultimately developed into the ‘My life with MS’-project which invites participants to tell the story of their life with MS focusing on self-chosen key events. Data collection started in July 2019 and is still ongoing.

In the ‘My Life with MS’ survey, participants were asked to tell the story of their life with MS by identifying key MS-related life events that they considered central to the progression of their MS. Participation in SMSR studies is possible both online and by paper-pencil. Online participants of the SMSR had the option to document up to nine such MS-related events. Paper-pencil participants received sheets for up to three events but could request more sheets if needed. The ‘My life with MS’-survey started with a visual timeline covering the time period between receiving the MS diagnosis and the present day, participants filled in keywords at appropriate points on the timeline, detailing life events/experiences they would like to elaborate on in detail later. For each of these life events, participants completed a brief semi-structured survey. An overview of the survey structure for the individual MS-related events can be found in the Supplement (S1 Fig). Participants first provided a short text description of the event itself, the event’s consequences for them personally, sources of support which they had access to at the time, what they felt that they could have used at the time to deal with the event, and their advice for other persons with MS in similar situations. They then allocated the life events to a range of thematic categories, defined as ‘personal level’ (e.g., mental wellbeing), ‘societal level’ (e.g., social position), and ‘direct environment’ (e.g., family and loved ones). They then indicated on a visual analogue scale and as a numeric value how they felt when the event took place (items adapted from the EQ-5D; range: 0–100, higher values reflecting a better emotional state). [30] In addition, participants further specified their emotional state by indicating which negative or positive emotions (happy, relaxed, sad, depressed, anxious, nervous, tired, other) best reflected their emotional state at that time. Next, they provided a free text description of the life event’s impact on their lives. On a 5-point Likert scale, they further rated the perceived impact of the event in hindsight (very negative, negative, neutral, positive, very positive) as well as the persistence of any event-related consequences (brief, recurrent, permanent [always present], other). They subsequently indicated whether the event had any impact on their broader social network (i.e., family, loved ones) and again, rated the event in terms of impact and persistence. Participants then described in free text who had supported them during the event and, in another text entry, what they felt could have helped them in the given situation. Finally, participants were asked to share their advice for other persons in a similar situation. For most of the closed questions, participants had the opportunity to provide additional comments if they wished. The first life event description was followed by self-report questionnaires covering information on participants’ MS disease, medication, therapies, as well as general physical and mental health over the course of the last half year. Participants then completed the procedure described above for the remaining life events.

Description of the text corpus underlying this study

For the present study, we focused on the event keywords provided in the visual timeline (S1 Fig, survey section A1) and the related event descriptions (S1 Fig, survey section A3). We combined the keywords with the respective event description for all subsequent analyses. That is because not all participants provided a text description complementing their keywords so that these events would otherwise be lost. Again others, regarded the event description as a continuation of the event keywords so that they did not reiterate the keyword in their description.

Additional measures

For comparison with the emotion classifications derived from the language model, we used the corresponding survey ratings presented in S1 Fig (survey section A7).

Analytic approach

The present research implements a range of different analysis methods. To determine which broader categories of life events persons with MS perceived as central and emotions associated with the descriptions, we implemented NLP techniques including topic modelling. To compare variables across groups, we used descriptive statistical methods. All analysis steps are detailed in the following sections. A flowchart visualizing the individual analysis steps is provided in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Flowchart visualizing the analytical approach and individual steps in the present manuscript.

https://doi.org/10.1371/journal.pdig.0000305.g001

ML-driven natural language processing

Text preprocessing

Text preprocessing prior to the analyses comprised manual spelling correction, automated language detection, and offline-translation of all the text entries to English, conversion of, for example, clinic names and medication brands into generic terms, removal of stop words and transformation of generic terms, lemmatization, and part-of-speech tagging (explained in more detail below).

All text entries were first checked manually for spelling errors and any abbreviations were spelled unless they were widely established (such as ‘MS’ for ‘multiple sclerosis’ or ‘MRI’ for ‘Magnet Resonance Imaging’). Filling in questionnaires can be exhausting for some persons with MS. Apart from typical spelling errors, our text data is thus characterized by the frequent usage of non-standard abbreviations (e.g., ‘Betaf.’ for ‘Betaferon’). Given the specific nature of errors sources and the variety of MS-specific expressions in our data, we initially conducted a manual spelling correction which also spelled out all uncommon abbreviations. On very rare occasions, participants also used Swiss-German words (e.g., ‘lüpfen’ instead of ‘heben’ [‘lift’]). In such cases, Swiss German words were replaced with High German. Since the SMSR conducts its assessments in the three of the Swiss official languages (i.e., German, French, Italian), we then translated all text entries into English. For keywords related to MS symptomatology for which more than one term is available, one term was consistently used, and if available, the one-word term was preferred (e.g., ‘diplopia’ instead of ‘double images’). For text translation, we used the Hugging Face ‘transformers’ library in Python and the Helsinki-NLP Opus language models ‘German-English’, ‘French-English’, and ‘Italian-English’ which can be retrieved from the Hugging Face website (https://huggingface.co/) [15]. All language models were downloaded, and all translations were completed locally in an offline Python environment. We then automatically removed any names (personal, institutions) and places (cities, countries) from the data for anonymization purposes. A subsequent manual check ensured the data’s anonymization. In this step we additionally excluded, e.g., highly specific job descriptions. We subsequently removed any words classified as ‘stopwords’ (that is, common words without specific meaning such as ‘and’) using the ‘stopword’ list of the widely established ‘NLTK’ library [31]. Given the retrospective nature of the study, we expanded the stopword list by a set of nondescriptive terms and time references which repetitively occurred in individuals’ descriptions (e.g., ‘year’, ‘month’, ‘week’, ‘minute’ etc.). We then lemmatized all words using NLTK’s ‘WordNetLemmatizer’. Lemmatization refers to a word’s conversion to its inflective form (e.g., ‘treatments’ to ‘treatment’) so that different derivations of the same word can be analyzed as a unit. As nouns were most informative for our guiding research aim (i.e., which broader categories of life events persons with MS perceived as crucial in terms of the progression of their MS across the life course), we finally extracted all lemmatized nouns from the text entries using the part-of-tagging functionality implemented in the ‘spacy’ library, version 3.2.3 [32]. Finally, we conducted a manual check of the lemmatized nouns to ensure that only anonymous information was included and that only nouns which can found be in English dictionaries were included. Given the novelty of our methods and the specific nature of our data, we manually checked all preprocessing steps and made corrections where needed.

Topic modelling—Latent Dirichlet Allocation (LDA)

To identify different types of latent life-event topics, we implemented topic modelling. Topic modelling is an unsupervised ML method suitable to analyze large-scale unlabeled text data based on word co-occurrences. The overarching idea of the method is that a distinct number of latent topics underlies the text entries and, in parallel, each of the text entries is a composite of different topics in varying proportions. LDA is an established probabilistic modelling technique [33]. The process of topic determination through LDA works such that words are allocated to a topic on the basis of their co-occurrence with other words in the text entries. For this study, we used the LDA algorithm implemented in the ‘gensim’ library, version 3.8.3 [34].

Rationale for topic determination. A challenge and controversial issue in topic modelling is the determination of the optimal number of topics to be modeled. This choice is key as it has to be decided on prior to the LDA analysis and thereby impacts on the nature of the topics to be generated [35,36,37]. Following recent recommendations and subsequent applications, we combined a data-driven and empirical approach to determine the optimal number of topics [38,39]. Specifically, we generated a series of models for comparison purposes. We started with a two-topics model and subsequently increased number of topics by one until 40 (that is, a 40-topics model). We checked different topic models manually for coherence and interpretability by exploring how the topic representations changed in terms of their keywords and the keywords’ importance as we successively increased the number of topics. To ensure that our interpretation of the respective topic representations was correct, we continuously verified our understanding with the underlying text data that was representative of the respective topics (i.e., which had a high probability score in the LDA model). Increasing the number of topics generally results in splitting topics that were previously one into smaller subtopics. A key challenge in this process is to strike a balance between capturing the larger underlying patterns of the corpus without neglecting important details. Finally, we compared the models’ goodness in terms of their topic’s coherence (i.e., the ‘c_v score’), which provides a statistical proxy in this respect. While the ‘c_v score’ can be considered as a general reference point, domain expert knowledge is key for fine-tuning topic models to ensure that topics are coherent and well-interpretable at the content level. In line with Occam’s razor, we opted for an LDA model optimizing topic coherence while being as parsimonious as possible.

Clear allocation to a distinct category. We considered an event to be clearly allocated to a certain category (1) if it was the dominant category (that is, the highest percentage of the topic probability score), and (2) if this probability score was larger than 33%. We decided on the one-third allocation upon prior exploration of the probability score distributions in the present sample and found this cut-off to be appropriate for noise reduction purposes.

Visualization. To visualize the distinct topics in an information-dense fashion, we computed ‘word clouds’ using the ‘wordcloud’ library, version 1.8.1 [40]. The ‘word clouds’ consist of the most frequent key words underlying a specific topic whereas the word size corresponds to their relative frequency of its occurrence across all text entries.

Fine-grained emotion analysis

We performed the emotion analysis based on the text entries. While study participants documented a high number of event descriptions (n = 4297), they completed the full event questionnaire, including self-rated emotion categorizations, for only 58.72% of all events (n = 2523). This was most common when an individual documented many events. We conducted the emotion analysis using the full, translated data including in English language in which translation errors had been corrected, all abbreviations were written out in full (e.g., ‘multiple sclerosis’ instead of ‘ms’), specific terms such as medication names had been converted to generic expressions (e.g., ‘medication’ instead of ‘Interferon’) and which still included all stopwords, verbs, adjectives, nouns, and punctuation.

Rationale for model selection. We determined the event descriptions’ emotional content using a pretrained language model from the ‘Hugging Face’ open-source platform and community [15]. To decide on a pretrained emotion classification model, which is suitable for our type of text data, we conducted an upfront evaluation of the available models by assessing their classification performance on our data. Our prior evaluation found the ‘Emotion English DistilRoBERTa-base’ model to be most appropriate for the ‘My life with MS’-data [41]. This pretrained language model outputs a probability score for each of the six emotions theorized by Paul Ekman as ‘basic human emotions’ (‘fear’, ‘joy’, ‘sadness’, ‘surprise’, ‘anger’, ‘disgust’) together with a neutral category. The ‘Emotion English DistilRoBERTa-base’ model is based on the distilled version of the original ‘RoBERTa-base model’ which is smaller in size but has similar performance compared to the original. The model is suitable text data in English language. For the emotion classification task, the ‘distilled RoBERTa-base model’ has been fine-tuned on six publicly available datasets including data labelled with the respective emotions [42].

Analysis. For emotion determination in Python, we used ‘Transformers’ library (https://huggingface.co/transformers) which has been developed by ‘Hugging Face’ to make the website’s models conveniently accessible to the community [15]. The Transformers library provides application programming interfaces and tools for download and training of open-source pre-trained models that are available on the website. For emotion analysis, we first loaded the model into Python. We then converted the text data into a numeric format for further processing of the text data using the ’RobertaTokenizerFast’ tokenizer. This tokenizer is suitable for any language model based on the ’RoBERTa model’ [43]. Each unique participant text entry was assigned a score for each of the six represented emotions. We then normalized the model output for respective emotions using the ‘softmax’ function from the ‘scipy’ Python library [44]. This transformation allows the results to be interpreted as probabilities which makes it more intuitively understandable [45]. To compare the emotion scores derived from this analysis with the corresponding, binary ratings from the survey, we calculated point-biserial correlations. We then visualized the resulting correlation using the R package ‘corrplot’, version 0.92 [46].

Visualizations. Data visualizations were computed using the ‘gglot2’ library, version 3.3.5, in R. We used box plots for a standardized visualization of the eight topics’ fine-grained emotion characteristics with the raw data displayed as dots.

Descriptive analyses

All SMSR participants were invited to participate in the present study (n = 2602, status March 2022) and 1039 (39.93%) of them participated in the ‘My life with MS’ survey, either online (n = 873; 84.02%) or as paper-pencil version (n = 166; 15.98%). The average age of participants was 48.55 years (SD = 12.58 years; range: 19–83 years) and 75.65% were female. Participants provided a total of 4309 event descriptions. Sixteen event descriptions had to be excluded because they did not contain meaningful information (e.g., due to the use of unusual abbreviations). This resulted in a final sample of 4293 event descriptions with an average 4.13 events per person (SD: 2.21; range: 1–11; S2 Fig). The translated text descriptions (still including all stopwords) were on average of 18.27 words in length (SD: 29.57; range: 1–253). Of the 1039 participants, 79.30% provided their text entries in German (n = 824), 17.61% in French (n = 183), 2.79% in Italian (n = 29), and 0.28% in English (n = 3).

Central life events for the progression of the MS

First, we examined which broader categories of life events (‘topics’) persons with MS perceived as central in terms of the progression of their MS. Modelling coherence values for an increasing number of topics showed highest coherence value for an LDA models with 8 topics (and with 30 or more topics; S3 Fig). In line with Occam’s razor, we opted for an LDA model optimizing topic coherence while being as parsimonious as possible. Among those options, manual examination confirmed that modelling eight topics led to the most distinct topics and thereby best facilitated their interpretation. The eight topics were (1) ‘diagnosis’, (2) ‘medication, treatment’, (3) ‘relapse, child’, (4) ‘work’, (5) ‘birth, health’, (6) ‘partnership, MS’, (7) ‘rehabilitation, wheelchair’, and (8) ‘injection, symptoms’. The 10 most frequent keywords per topic are displayed in Fig 2 with font size reflecting a words’ relative frequency. Exemplary, anonymized event descriptions which are most representative for each of the topics are provided in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Word cloud visualizing the ten most frequent key words of the 8 overarching topics of MS-related life events, displayed for each of the eight topics separately.

Word size corresponds to a specific word’s relative frequency compared to the total number of analyzed words. The words’ color is distinct for the eight topics. AE = adverse effects; IV = abbreviation for the disability insurance in Switzerland; MSG = Swiss MS Society.

https://doi.org/10.1371/journal.pdig.0000305.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Characterization of the eight overarching topics based on the most frequent keywords and exemplary, anonymized event descriptions for the most frequent emotions.

https://doi.org/10.1371/journal.pdig.0000305.t001

Emotions

We subsequently examined the emotions associated with different types of life events (Fig 3). We found the key MS-related life events to be predominantly linked to emotions of negative valence, most notably to sadness with means per topic category ranging between 0.41 and 0.15. Fear was also a frequent emotion with means per topic category ranging between 0.21 and 0.04. The emotions ‘anger’, ‘surprise’, and ‘disgust’ were rare among all topic categories. Importantly, feelings of joy were linked to the topics ‘birth / health’ (M = 0.19, SD = 0.34) and ‘partnership & MS’ (M = 0.14, SD = 0.27). It is important to note that the text entries exhibited a high amount of heterogeneity as indicated by the large standard deviations and as visualized in Fig 3. A plot which also includes the rare emotion disgust is provided in the supporting information (S4 Fig). For 2523 of the 4294 text entries, participants also provided survey ratings (58.76%). We validated the emotion classification based on the pre-trained language model by correlating the respective probability scores of the emotions for each entry with the binary survey ratings (emotion was present vs. emotion was not present at the time) provided by the participants. A correlation plot displaying the point-biserial correlation coefficients for all comparisons is provided in S5 Fig. Our results show that the emotions identified by the pre-trained language model are positively correlated with the binary survey ratings (i.e., fear, sadness, joy).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Emotion probability scores for ‘sadness’, ‘fear’, ‘anger’, ‘joy’, ‘surprise’, and the overall frequency of the emotions across all events are displayed separately for the eight topic categories.

Only events for which (1) the respective category is the "main category" and (2) the assignment to this main category is at least one third (33%) are included in the graph. Emotion probability scores range from 0 to 1. Higher scores indicate a higher probability for a text description being corresponding to the respective emotion category. Emotion probability scores are displayed using boxplots. The boxplot’s middle line refers to a topic category’s median, the box itself displays the interquartile range (IQR; the range from the 25^th to the 75^th percentile), and the whiskers mark the minimum (25^th-1.5*IQR) and maximum (75^th-1.5*IQR). The thick dots represent outliers beyond the minimum/maximum. The small dots display the raw data.

https://doi.org/10.1371/journal.pdig.0000305.g003

Topic characterization

The eight topic categories consisted of five directly MS-related topics (‘diagnosis’, ‘medication, treatment’, ‘relapse, child’, ‘injections, symptoms’, ‘rehabilitation, wheelchair’) and three topic categories that address broader life goals (‘partnership, MS’, ‘work, ‘birth, health’). Diagnosis was the most frequent topic category (34.75%), followed by ‘medication, treatment’ (23.64%) and ‘work’ (11.06%). Each of these topic categories was defined by its most frequent terms. In contrast, the topic categories ‘relapse’ (12.30%), ‘birth’ (5.26%), ‘rehabilitation, wheelchair’ (5.33%), or ‘injections, tingling’ (2.84%) were rather heterogeneous in nature. The eight distinct topic categories are detailed below.

MS-related topics

Medication, treatment (topic 2)

The most frequent references in this topic category were related to immunomodulatory treatment, which is usually prescribed to prevent the worsening of a disease. Emotions most frequently associated with this topic category were sadness and fear, possibly related to uncertain treatment outcomes (e.g., continued worsening despite treatment), negative treatment outcomes, or adverse effects.

Relapses, child (topic 3)

This topic category was mainly defined by the word relapse. However, the additional keywords indicated a rather heterogeneous context of relapse. As the keywords ‘child’ and ‘IV’ (referring to disability insurance) suggest, relapse affects many aspects of the lives of persons with MS. Others may indicate permanent impairments leading to disability (and the need for support from the disability insurance, which in Switzerland is abbreviated to ‘IV’). As a concrete example, some participants expressed fear of relapse during the postpartum period. On the emotional level, text entries are again characterized by sadness and, to a lesser extent, fear.

Injections, symptoms (topic 8)

This topic category is of a relatively heterogeneous nature. The common feature of the text entries in this category was the constant interference of symptoms (often tingling) with daily life, and the price of controlling them is often high. Reference is made to injections, the usual mode of administration of older disease-modifying drugs. Such injections are painful, and participants report increasing painful indurations under the skin, some of which require surgical removal. Both symptoms and therapy affect all aspects of life.

‘Rehabilitation, wheelchair’ (topic 7)

Progressive disability, as indicated by the keyword "wheelchair," for example, was the focus of this topic category. Progressive disability was also associated with the increasing need to rely on walking aids or to cope with more severe impairments (e.g., through rehabilitation measures). Membership of the Swiss MS Society for information and networking purposes was also mentioned in this topic category. Negative emotions were less common in this category, possibly related to the heterogeneous nature of this category. Indeed, rehabilitation was most often perceived as positive, while the need for a wheelchair or the prospect of needing one was perceived as negative.

Topics concerning broader life goals

Work (topic 4)

The work-related consequences of MS are at the core of topic category 4. As terms such as resignation (in the sense of quitting), unemployment, or reduction reflect, this theme category likely reflected the personal struggle that persons with MS experience in remaining employed and managing their daily workload despite multiple impairments. The text entries in this theme category were often characterized by feelings of sadness.

Birth, health (topic 5)

This topic category was primarily characterized by the term ‘birth’, which mostly refers to the birth of one’s own child or a grandchild. These events were mostly associated with joyful, positive feelings. However, several other prominent keywords indicate considerable heterogeneity in the statements (examples provided in Table 1). Other text entries in this topic category tended to focus on education (e.g., continuing education) or symptom management (physical therapy).

Descriptive characterization of individuals per topic category

A descriptive characterization of study participants based on the topic categories is provided in Table 2. In the table, participants are divided into those who made at least one entry on a particular topic category and those who made no such entry. Since the topic categories ‘diagnosis’ and ‘medication / treatment’ were reported by most participants (see S6 Fig), we examined only the topics ‘relapse / child’, ‘work’, ‘birth, health’, ‘partnership & MS’, ‘rehabilitation / wheelchair’, and ‘injection, symptoms’. The descriptive analysis revealed, however, that individuals-level characteristics (i.e., MS type, living situation, parentship, etc.) were distributed relatively homogeneous.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Relationship between six of eight overarching event topics and individual-level measures.

https://doi.org/10.1371/journal.pdig.0000305.t002

Limitations and Future Research

The present research has several limitations which merit consideration. First, the event descriptions differ substantially in their level of detail. While a number of participants provided mainly keywords, some provided comprehensive descriptions detailing thoughts and emotional experiences. Very short text entries lack context information and are therefore prone to misinterpretation. In terms of the overarching topics, we considered both short and detailed texts as informative sources of information as our analysis was based on nouns only, which were almost always present in both short and long text descriptions. Given the novelty of our methods and the specific nature of the data (MS context), we incorporated several manual steps to ensure the integrity of the methodology–a somewhat time-consuming process. Future research would therefore benefit from studies that evaluate the performance of such open-source language models on various health-related datasets (e.g., those that contain numerous medical terms and those that use everyday language). Second, there may have been self-selection in recruitment in the sense that mainly persons with MS who had a desire to tell their story participated in the study. In general, persons with a recent MS diagnosis, a milder MS disease course, or a very advanced disease stage (e.g. those who are entirely dependent on external help) are somewhat underrepresented in the SMSR. [47] Finally, our analytic approach is suitable for revealing general patterns in the text. A more fine-grained investigation of individual-level patterns and their contextual meaning would still require a more detailed review of the text entries in combination with additional individual-level context data (such as information on MS type and progression, wellbeing measures).

A current controversial debate concerns the limitations and risks of using artificial intelligence, which NLP is a subdiscipline of, in the exploration of health data. At its core, this discussion revolves around the risk of potential biases that the model has learned through the training data. In the present research, we have implemented pre-trained language models for automatic text translation and emotion classification. Instead, the LDA topic modelling technique is not subject to a priori bias as it does not rely on a pre-trained language model. With respect to automated text translations, a potential bias in our study is unlikely because all translations were manually reviewed by native-level speakers of the respective languages who also had expertise in MS. The pre-trained language model we implemented for emotion classification was fine-tuned on several public datasets which mainly include social media posts which that have been labelled for emotion categories by manual annotators [41]. To ensure that emotions would correctly classify the data in our study, we pre-tested several models as described in the methods section. The emotion classifications were also manually reviewed by CH, which verified that the overall quality of the classifications was high and appropriate. However, future research would benefit from emotion classification models trained on diary or patient narrative data to ensure high model performance.

Our findings have several implications for future research into the life stories of persons with MS. Future research on complex health conditions may benefit from a citizen science approach. Leveraging the mostly untapped expertise of ‘citizen scientists’ for research purposes may help uncover blind spots in health research. Their knowledge may also help pave the way for tailoring better disease management strategies (including medical, symptomatic, and lifestyle aspects) to individual needs, contexts, and everyday challenges. Language represents a key medium for harnessing the expertise of citizen scientists. Future research would also benefit from further development of language models that have been trained on data from individuals with a particular health condition and are therefore more accurate when applied to new data in a similar context. However, persons with severe chronic conditions may struggle to write longer text and default to using abbreviation and providing keywords instead of a coherent text, ultimately limiting the automatic text analytics. A promising avenue for future research would thus be daily-life speech sampling. The emergence of open-source speech-to-text models might circumvent the need to rely on commercial providers and pave the way for revolutionizing daily-life research. Finally, the high participation of registry participants in project development and participation indicates that many individuals with MS have experiences that they are willing to share. Our experience is indeed that the SMSR participants’ motivation for participating in research is often rooted in their wish to contribute to advancement which will benefit others in the long term. Future research may benefit from intensifying citizen-science collaborations to deepen our understanding of how MS impacts on individuals’ everyday experience and what is most needed to improve the quality of life for persons with MS.