Ethical approval was given by the human research ethics committee of the Delft University of Technology (ID 2596). Informed consent was not required since the data used in this study were sourced from publicly available forums, in accordance with institutional guidelines. The data were accessed and analyzed in accordance with the terms of service and privacy policies outlined by the platform hosting the data. As the ethics committee suggested, all data used in this study were anonymized by removing all direct and indirect identifiers (eg, names, location, and user ID) to prevent potential privacy issues within our data set. Confidentiality was strictly maintained during data collection, analysis, and reporting to ensure that no identifiable information was disclosed.

In order to assess patient experiences using machine learning–driven analysis, this study used systematic data collection, advanced topic modeling techniqu and cocreation sessions with domain experts to interpret and validate identified clusters. The Consolidated Reporting Guidelines for prognostic and diagnostic machine learning modeling studies checklist can be found in Multimedia Appendix 1.

To enable research, we used forum posts of patients with CRC scraped from the Cancer Survivors Network USA, an open-source patient community platform. The web-based platform provides support, education, and advocacy for patients affected by CRC that includes current patients, former patients, families, and caregivers. The posts are written with the intention of asking questions to peers and health care providers, as well as forsharing experiences among peers [16]. The initial CRC discussion thread (ie, main posts, comments, and replies) has remained active on the public platform since the year 2000. No distinctions were made between main thread posts and replies to posts.

The data collection involves 2 main steps: first involves using Selenium WebDriver to gather the URLs of discussion pages, and second, using BeautifulSoup to extract data from the HTML elements while keeping sensitive information secure.

Information about the personal (health) status of patients, which could be directly or indirectly identifiable, was excluded from the analysis. Topic modeling, a machine learning algorithm was used to recognize patterns in the platform data. Nonnegative matrix factorization (NMF) was used as a topic modeling technique to analyze the data set and identify topics with weighted keywords [17,18]. The number of topics was determined by evaluating topic coherence scores and model stability [19,20]. Model stability was assessed using Jaccard similarity to reduce the overlap of topics. Human evaluation was used to ensure that the output contains diverse and distinctive topics, without topics unrelated to patient experiences. This evaluation was conducted by 4 researchers to select the number of topics that yield the most diverse and distinctive clusters.

To comprehend each topic, researchers meticulously read and analyzed the top 50 most relevant posts and top 20 related keywords. The most relevant posts and keywords had the highest dominant topic score identified by the algorithm, indicating they were most closely associated with the topic. The algorithm efficiently identified the most relevant posts and keywords within each topic by using Term Frequency–Inverse Document Frequency to calculate the importance of each term in relation to each document [21].

The identified topics were grouped into clusters for topics that revolved around similar concerns. Irrelevant topics (eg, those related to platform-specific issues) were excluded. The frequency of discussions on each topic was recorded by counting the number of posts associated with each topic.

This study uses patient community journey mapping to visually present the identified patient experience topics. The creation of the patient community journey map began with the conduction of 2 interviews with CRC medical doctors to outline the patient journey and its distinct phases such as diagnosis and follow-up of a patient with CRC.

The results of the interpreted topics were deliberated in an in-person cocreation session involving domain experts [22]. The cocreation session took place in the Erasmus Medical Centre in March 2023 and lasted 3 hours. The interdisciplinary team of domain experts reviewed and discussed the outputs comprehensively. The team comprised 2 oncological surgeons, 2 surgical oncology PhD candidates (MD), and an epidemiologist with an interest in oncological research. A total of 3 out of 5 participants have over 20 years of work experience, indicating a high level of expertise. During the cocreation session, clusters from the NMF results were interpreted, and feedback was provided on the generated interpretations. This process ensured a thorough examination and validation of the identified clusters within a clinically relevant context. During this session, the same domain experts also reviewed the patient community journey map. The identified topics were positioned on the journey map, with allocation to specific phases of the patient journey. Distinctions were made regarding whether each topic was taking place at a “hospital” or “home”setting.

A total of 294,166 posts were extracted from the patient forums of the Cancer Survivors Network USA. The posts on the website were created between the year 2000 and 2022. However, 212,107 posts were analyzed, as the remaining 82,059 posts were excluded through topic modeling.

Using NMF topic modeling, the topic coherence score, and model stability did not provide clear insights as the coherence score consistently declined. Therefore, human evaluation was required. Initially, 40 and 50 topics were considered the best amount of topics. Subsequently, the cut-off of 45 topics was also reviewed. Ultimately, 50 topics were chosen as the number that provided the most diverse and distinctive topics. This was confirmed by examining keywords and the most relevant posts for each cluster. Additionally, one topic bin identified 82,059 posts as unclassifiable," which could not be categorized into any of the 50 topics.

A total of 50 topics were identified with the use of NMF topic modeling. Thirteen topics were excluded as they were unrelated to patient experiences, such as platform use and expressions of gratitude. The data export of the relevant topics is shown in Table 1. These are the key patient experiences found in our study.

A total of 10 clusters were derived from 37 topics. The patient community journey map, shown in Figure 1, serves as a visual guide to navigate through the nuanced dynamics of patient experiences.

The topics in the home context have a more emotional content, as emotionally charged keywords such as “confused,” “bad,”and “worried” are more often discussed. Conversely, the hospital context is marked by the clinical terminologies “drug” and “node.” This reflects the distinct atmospheres characterizing discussions in different contexts.

The clusters commanding the highest share of posts are “Daily activities while living with CRC” (38,782, 18.3%) and, “Understanding treatment including alternatives and adjuvant therapy” (31,577, 14.9%) underscoring their important role in shaping the experiences of CRC survivors. Patients expressed significant concerns about test results, as indicated by the 14.1% of posts within cluster 4. As shown in Figure 1, “Understanding treatment including alternatives and adjuvant therapy” is a cluster that goes through almost the entire journey, while the “Daily activities while living with colorectal cancer,” is more in the home context and starts later in the follow-up.

This study delves into the daily experiences of patients using data from a web-based platform, applying topic modeling as a tool to efficiently analyze and interpret large volumes of patient-driven daily life experience data. We analyze patient experiences (eg, struggles, tips, and coping strategies) reported on a platform to better understand patient needs. This understanding of patients’ needs has positive potential for future care path development such as information provision or psychosocial support in remote patient monitoring. The use of machine learning techniques in analyzing qualitative data holds significant promise to improve various aspects of health care. This includes optimizing patient care, by using advanced methods such as using symptom checkers to guide patients to the most suitable care journeys and support disease diagnosis. This approach will ultimately lead to patient-centered care to support and empower patients.

The identified clusters in this study align with findings from previous qualitative and survey research, emphasizing the potential to optimize the use of machine learning in qualitative studies of health care in the years to come [4,23]. Based on our analysis, it becomes evident that patients exhibit a notable interest in actively participating in discussions with other CRC survivors to exchange experiences regarding the daily challenges of living with CRC. This underscores the crucial role of social support and the dissemination of pragmatic, day-to-day coping strategies within the broader patient community. This theme aligns with observations from a systematic review detailing the experiences of CRC survivors, where the significance of coping and addressing functional limitations emerged as a central theme [23].

The cluster with the highest share of posts (cluster 10; 38,782, 18.3%) revolves around the challenges associated with resuming daily activities while living with CRC. This substantial volume of posts highlights the significant impact of this particular aspect on the experiences of CRC survivors, as well as the needed change in patients’ mindset (cluster 8; 26,379/212,1087, 12.4%). The systematic review of survivorship experiences further supports these findings [23]. The alignment between these dominant themes in the posts and the insights from the qualitative literature underscores the robustness of our findings [24]. These findings emphasize the importance of integrating comprehensive support mechanisms to facilitate the successful reintegration of patients into their daily lives.

On the platform, patients frequently engage in discussions centered around understanding treatment options (cluster 2; 31,577, 14.9%), highlighting its significance as a key topic. This observation aligns with the qualitative study conducted by van Deursen et al [4], where participants similarly emphasized the need for more information regarding their treatment. It is noteworthy that patients on the platform also frequently express concerns regarding test results (cluster 4; 29,984, 14.1%). This anxiety and fear of recurrence are recognized as common concerns among cancer survivors [25], highlighting the potential to offer valuable support in addressing this. Additionally, the positive outcomes associated with the interaction with family, as reported in the systematic review [23], further underscore the importance of familial support within the context of patient experiences in our findings [26].

A novel finding from our analysis was the open expression of patients’ sentiments regarding the opinions of medical professionals (cluster 1; 9756, 4.6%). Not completely surprising as this is possibly due to the reluctance of patients to express negative feelings in a setting where researchers are present, such as focus groups or interviews. Furthermore, qualitative studies primarily concentrate on the follow-up phase, excluding discussions on the experiences of surgery and the subsequent recovery phase. These aspects might provide valuable insights into the comprehensive journey of patients.

To reflect on our approach, topic modeling allows extraction from a vast pool of large-scale patient data. The alignment of our study with existing literature is indeed promising. Additionally, topic modeling offers a novel approach by sparing patients the burden of active participation in qualitative studies. Rather than focusing on a single aspect, as is often the case in conventional studies, this method presents a broader spectrum of patient experiences beyond specific questions from researchers. Notably, topic modeling relies on patient-driven data, in contrast to a physician-driven approach. This method’s replicability allows seamless comparisons between various medical conditions. This not only contributes to a more extensive comprehension of diverse experiences but also facilitates the identification of commonalities and distinctions of different medical conditions. For example, we explored the present method with patients with pulmonary fibrosis and found comparable experiences to those with CRC, their increased focus on treatment options highlights the distinctive challenges inherent to their specific condition (unpublished data). Given the large data sets on forums, we can quantify patients’ experiences with the implementation of topic modeling. This complements the challenges of focus groups and interviews, where the ability to quantify and prioritize findings is limited. As patients’ experiences change throughout their journey, the simplicity of using topic modeling remains a flexible and efficient approach for understanding patients’ experiences and priorities in the era of remote monitoring [27]. It offers a convenient and adaptable method, without burdening patients [28].

Using stories from patients who prefer to share their experiences on a digital platform introduces a risk of selection bias. It is essential to address the potential overexpression of emotion in our findings as the platform primarily captures narratives from individuals who are willing to share their experiences on the web. However, it should be recognized that these individuals represent the intended user group for new care path developments such as remote monitoring support approaches. Patients who are not feeling well, patients who do not have access to the internet, or reserved patients may be less inclined to engage in web-based discussions and might be underrepresented [29]. To complement this, in previous research, we conducted interviews with vulnerable patient populations to ensure their perspectives were included in our understanding of patient experiences [9]. Similar themes such as the need for information provision, concerns about test results, and challenges in reintegrating into daily life were found in this study. Additionally, our approach demands a significant investment of human labor and time to interpret the meaning of topics after generating groups of topics. Thus, exploring another artificial intelligence technique to interpret patient experiences is a promising direction. As an example, a large language model, a pretrained model, can be used to interpret and summarize large quantity textual data [30]. The use of machine learning, even in the part where we used human experts’ knowledge, is a promising area to labor-effectively understand the patients’ experience while having the least safeguarding support of human validation. Another limitation is the lack of direct involvement of patients with CRC in the co-design or review of the map. We plan to address this in the future as this can provide valuable insights. However, the primary purpose of this method is to enhance the methodology while minimizing the burden on patients. To overcome this limitation, we sought input from highly experienced medical professionals’ perspectives. Three of these experts possessed more than 20 years of medical and research experience, specifically with patients with CRC.

In future research, efforts should be directed towards further refining machine learning techniques with the ultimate aim of minimizing or eliminating the need for human involvement, as this will enable more frequent monitoring of patient experiences and thereby facilitate responsiveness to those experiences. Addressing the nuanced interpretation of emotions in patients remains a challenge for artificial intelligence systems. A promising direction is the automatic interpretation of patient-reported outcome measures using machine learning and NLP. Patient-reported outcome measures offer structured data that can improve AI’s ability to understand emotional nuances. This approach aims to improve remote patient monitoring, reduce the burden on health care professionals, and identify those who face psychosocial challenges [31]. This approach not only empowers patients in managing their health but also leads to a proactive and personalized approach.

In conclusion, topic modeling and the use of patient forum data offer a robust and efficient approach to understanding patients’ experiences in their daily lives. This approach reveals the challenges patients encounter in their daily life such as getting back to daily activities and the need for understanding their treatment. This study not only provides a comprehensive overview of patient experiences through web-based platforms but also highlights the potential for improving patient monitoring systems. This machine learning technique of identifying patient experiences contributes to a more efficient way of building value-based health care. By integrating these insights into the development of remote monitoring solutions, a patient-centered approach can be created that not only addresses medical concerns but also caters to the broader spectrum of challenges individuals face in their daily lives.