Insights Into the Patient Experience of Hormone Therapy for Early Breast Cancer Treatment Using Patient Forum Discussions and Natural Language Processing

Abstract

PURPOSE

Understanding the real-world experience of patients with early breast cancer (eBC) is imperative for optimizing outcomes and evolving patient care. However, there is a lack of patient-level data, hindering clinical development. This social listening study was performed to understand patient insights into symptoms and impacts of hormone therapy (HT) for eBC using posts from patient forums on breastcancer.org to inform future clinical research.

METHODS

Natural language processing (NLP) and machine learning techniques were used to identify themes related to eBC from a sample of 500,000 posts. After relevant data selection, 362,074 eBC posts were retained for further analysis of symptoms and impacts related to HT, as well as insights into symptom severity, pain locations, and symptom management using exercise and yoga.

RESULTS

Overall, 32 symptoms and nine impacts had significant associations with ≥one HT. Hot flush (relative risk [RR], 6.70 [95% CI, 3.36 to 13.36]), arthralgia (RR, 6.67 [95% CI, 3.53 to 12.59]), weight increased (RR, 4.83 [95% CI, 3.20 to 7.28]), mood swings (RR, 7.36 [95% CI, 5.75 to 9.42]), insomnia (RR, 4.76 [95% CI, 3.14 to 7.22]), and depression (RR, 3.05 [95% CI, 1.71 to 5.44]) demonstrated the strongest associations. Severe headache, dizziness, back pain, and muscle spasms showed significant associations with ≥one HT despite their low overall prevalence in eBC posts.

CONCLUSION

The social listening approach allowed the identification of real-world insights from posts specific to eBC HT from a large-scale online breast cancer forum that captured experiences from a uniquely diverse group of patients. Using NLP has a potential to scale analysis of patient feedback and reveal actionable insights into patient experiences of treatment that can inform the development of future therapies and improve the care of patients with eBC.

Social listening and NLP/ML uncover breast cancer patients’ experiences with hormonal therapies.

INTRODUCTION

Breast cancer (BC) is the most common cancer diagnosis among women in the United States, with an estimated 287,850 new cases in 2022, constituting around 15% of all new cancer diagnoses.¹ Around 65% of female BC cases are diagnosed at the local stage, which is potentially curable with standard locoregional and systemic treatments.^1,2 Standard treatment of early breast cancer (eBC) includes surgery with neoadjuvant and/or adjuvant chemotherapy in combination with human epidermal growth factor receptor 2 (HER2)–targeted therapy and/or hormone therapy (HT) depending on the molecular status of the disease.^3,4 Data show that 70%-80% of all BCs are estrogen receptor- and/or progesterone receptor-positive and, therefore, are likely to respond to HT.^5-7

CONTEXT

• Key Objective

• Can the patient perspective on hormonal therapies and their side effects in treating early breast cancer (eBC) be captured and quantified through social listening and natural language processing (NLP)/machine learning (ML)?

• Knowledge Generated

• Large-scale analysis of patient discussion forum posts identified the symptoms and impact of hormone therapy (HT) in eBC. Hot flush, arthralgia, weight increase, mood swings, insomnia, and depression had the strongest association with HT; 32 symptoms and nine impacts in total had a significant association.

• Relevance (F.P.-Y. Lin)

• Leveraging NLP and ML to analyze patient forums allows for scalable, real-world capture of endocrine therapy experiences in eBC, validating known side effects, and may uncover new, unidentified patient-reported symptoms. This approach provides an additional dimension to inform research into optimizing personalized survivorship care.*

• *Relevance section written by JCO Clinical Cancer Informatics Associate Editor Frank P.-Y. Lin, PhD, MBChB, FRACP, FAIDH.

While early diagnosis and treatment of BC leads to improved survival outcomes,⁸ 18%-50% of patients discontinue adjuvant HT prematurely, risking poor outcomes.^9-14 Early discontinuation of postsurgery HT is largely attributed to adverse events (AEs) that can significantly affect quality of life and, consequently, treatment adherence.^5,14,15 Therefore, understanding the nature and impact of AEs with HT in eBC from the patients' perspective is crucial to improving treatment compliance, outcomes, and the overall patient experience. Furthermore, as the treatment landscape evolves because of the heterogeneous nature of BC, it is essential to understand the patient experience from existing standard-of-care treatments to inform the next wave of drug development.

Data on the real-world experience of patients with eBC can provide a breadth of insights across a range of biologic risk groups on the severity, impact, tolerability, and identification of and management of AEs from the patients' perspective that have the greatest impact on daily life. Qualitative methods, including patient interviews and focus groups, quantitative methods, including patient surveys and patient-reported outcome measures (PROMs), and mixed methods approaches can provide an understanding of patient experiences and unmet needs to inform drug development and improve patient care.^16,17 Qualitative methods involving patient interviews and/or focus groups are typically the foundation for identifying unmet needs, developing patient surveys and PROMs and informing patient-reported outcome (PRO) measurement strategies for drug development.^18,19 Such methods tend to involve small patient samples and require substantial resources.^20,21 Newer methods and advanced technologies involving artificial intelligence (AI) have begun to emerge to capture patient experiences, such as the analysis of social media posts or unstructured PRO or symptom data, offering an efficient way to gain insights useful for hypothesis generation and to aid in drug development.^17,22-24

To date, research on the impact of treatment in patients with BC, including eBC, has been limited,^25-27 involving traditional qualitative approaches with small numbers of patients with hormone receptor–positive/HER2-negative eBC participating in interviews and focus groups. While these qualitative approaches are effective at gathering first-hand feedback, they are time consuming, expensive to conduct, and are unlikely to represent the heterogeneous BC population.^24,28,29

There are many avenues via which patients, including eBC ones, can share their experiences, voice their concerns, and seek support from other patients who have been through or are navigating a similar path. Many of these support resources exist online and have registered users in the tens or hundreds of thousands, thus providing a valuable resource to gather a breadth of insights into the patients' experience of their disease. breastcancer.org³⁰ is a leading online resource for patients with BC, with over 230,000 registered users and more than 69,000 visits per day in 2021.³¹ Here, we describe the use of natural language processing (NLP) and machine learning (ML) to extract patient insights into the symptoms and impacts of treatment with HT in patients with eBC using a large data set from a patient forum on Breastcancer.³⁰

METHODS

An overview of the study design is provided in Figure 1. We collected a sample of 500,000 entries posted on the Breastcancer³⁰ patient forum between February 1, 2015, and May 31, 2022, using the Sprinklr social listening tool.³² To maintain privacy, patient IDs were not collected, with only the message field retained from each post.

FIG 1.

Study design and data extraction. eBC, early breast cancer; HT, hormone therapy.

Data Cleaning

To exclude posts that discussed literature, studies, or other informational articles, those that contained a URL (identified via regular expression matching) were omitted.

Selection of eBC-Related Posts

A multistage approach was employed to eliminate posts that discussed the metastatic, recurrent, or advanced BC settings. First, posts mentioning an advanced stage (stage IIIb and above) were excluded using a keyword matching approach. Next, two supervised ML models were built to identify posts discussing metastasis or recurrence. The models were built using the pretrained Bidirectional Encoder Representations from Transformers architecture³³ and fine-tuned by using a task-specific classification layer trained on 600 manually labeled posts. The models yielded an area under the curve of 0.78 and 0.81 for metastasis and recurrence, respectively.

Clustering Analysis

To cluster the posts and identify coherent segments, documents were vectorized on the basis of Term Frequency Inverse Document Frequency (TF-IDF), Latent Semantic Analysis (LSA) was used to reduce the dimensionality of the vectors, and KMeans clustering was applied to divide them into clusters. Since posts were short in length (median length: 105 words), and typically focused on a single aspect of patient experience, we favored a computationally efficient method like LSA over a more sophisticated topic-modeling approach like Latent Dirichlet Allocation.^34,35

The posts were then divided into different numbers of clusters between 2 and 100, and the goodness of the split was quantified using the Silhouette coefficient.³⁶ The optimal split was found to be at 60 clusters. For each cluster, words were ranked by their TF-IDF scores to identify the topics discussed and were confirmed by manually inspecting a random sample of 50 posts from each cluster. Selected clusters were visualized by projecting a vector representation of each post onto two dimensions using Uniform Manifold Approximation and Projection³⁷ and colored based on their distinct cluster membership.

Treatment, Symptom, and Impact Extraction

We applied the TERMite Text Analysis Engine³⁸ to the posts to identify symptoms (regardless of causality) and HTs (anastrozole, exemestane, goserelin, letrozole, leuprolide, and tamoxifen administered in the adjuvant setting) mentioned within each post. TERMite employs Named Entity Recognition and maps terms and phrases to specific categories on the basis of its curated vocabularies and ontologies that are created to identify drugs and symptoms. TERMite-derived symptoms that were considered to impact well-being, rather than direct symptoms of disease or treatment, were analyzed separately. In total, 10 impacts were derived from TERMite and included anger, anxiety, crying, depression, emotional distress, fear, frustration, insomnia, mood swings, and nightmare.

Prevalence and Association Analysis

The overall prevalence of specific symptoms and impacts was defined as the percentage of posts where that symptom or impact was mentioned. Similarly, the prevalence of a symptom or impact associated with an HT was defined as the percentage of posts where the HT and the symptom or impact were mentioned together.

To further understand the association between symptoms and HTs, the top 50 symptoms based on prevalence were selected. Next, a chi-squared test of independence was performed between each HT and each of the 50 symptoms. For each HT, symptoms with a statistically significant association (P ≤ .05) were retained, and the relative risks (RRs) for these symptoms were calculated. RR is a commonly used measure to quantify the strength of co-occurrence.^39-41 In our analysis, the RR indicated the ratio of how often each HT and symptom were observed together compared with how often they would be expected to appear together by chance.

Identification of Symptom Severity

For each eBC post that included a symptom, the sentence that mentioned the symptom was extracted and Parts-of-Speech tagging (Natural Language Toolkit) was applied to identify the 50 most frequent adjective terms. Of these, two subsets were identified that were associated with high and low severity (based on manual inspection; Data Supplement, Table S1). The severity level associated with each symptom was estimated by counting how often one or more of the adjectives in the high or low severity collection appeared in the same sentence as the symptom.

Identification of Pain Locations

To obtain further insight into posts mentioning nonspecific pain, posts were searched for any occurrence of the following loci of pain: arm, hip, ovarian, pelvic, and/or uterine.

Identification of Insights Into the Management of Symptoms Using Exercise and Yoga

Two sets of quotes were extracted from eBC posts that mentioned exercise or yoga, which equated to 2.7% (n = 9,708) and 0.7% (n = 2,372) of the total eBC posts, respectively. Posts in each set were then clustered using TF-IDF to identify topics discussed and were confirmed by manually inspecting a random sample of 50 posts from each cluster (Data Supplement, Fig S1). Sentiments that mentioned symptoms/impacts of interest were extracted from the four clusters with the highest number of posts for further analysis.

RESULTS

Data Cleaning and Selection of eBC-Related Posts

After data cleaning, the number of posts retained for analysis from the original 500,000 was 390,619. Keyword matching and modeling excluded 28,545 posts related to advanced BC, metastasis, and recurrence, leaving 362,074 (72.4%) posts related to eBC that were retained for further analysis (Fig 1).

Clustering Analysis

Of 60 clusters identified among eBC posts, seven prominent clusters (13.5% [48,963/362,074] of original posts) and associated themes were identified that covered aspects of the entire patient journey from diagnosis and biomarker testing to treatment concerns and impacts on everyday life (Fig 2). The Data Supplement (Table S2) shows representative example messages contained within each cluster. The Experience with HT cluster was selected for further analysis.

FIG 2.

UMAP projection of text contained in the seven clusters identified within eBC posts. eBC, early breast cancer; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; HT, hormone therapy; IDC, invasive ductal carcinoma; PR, progesterone receptor; UMAP, Uniform Manifold Approximation and Projection.

Prevalence of Symptoms, Impacts, and HTs

We compared the prevalence of symptoms and impacts in posts related to eBC and those related to eBC and HT to demonstrate strong associations of symptoms and impacts with disease stage and therapy. Analysis revealed 147,788 (40.8%) posts that mentioned a symptom across eBC-related posts. The 10 most prevalent symptoms were pain (unspecified), fatigue, ulcer, nausea, edema, infection, lymphedema, pruritus, cyst, and hot flush (Fig 3A). A total of 33,916 posts mentioned at least one of the TERMite-derived impacts across all eBC posts, with the top three identified as anxiety, fear, and frustration (Fig 3B).

FIG 3.

Top 10 most prevalent (A) symptoms and (B) impacts mentioned in eBC posts. Most prevalent (C) symptoms and (D) impacts mentioned in eBC posts that included at least one HT. eBC, early breast cancer; HT, hormone therapy.

A total of 24,764 eBC posts were identified that mentioned at least one HT and there were 1,125 distinct symptoms mentioned in these posts. The 10 most prevalent symptoms were pain (unspecified), hot flush, arthralgia, fatigue, lymphedema, ulcer, nausea, severe headache, edema, and weight increase (Fig 3C). All 10 TERMite-derived impacts appeared in these posts, with the top three identified as anxiety, depression, and fear (Fig 3D).

Associations Between HTs and Symptoms

Thirty-two symptoms had significant (P < .05) associations with at least one HT. Overall, hot flush (RR, 6.70 [95% CI, 3.36 to 13.36]), arthralgia (RR, 6.67 [95% CI, 3.53 to 12.59]), and weight increased (RR, 4.83 [95% CI, 3.20 to 7.28]) had the strongest associations with HT. Symptoms that had a low overall prevalence in eBC posts (<1%) that had a significant association with HT included severe headache (RR, 2.09 [95% CI, 1.09 to 4.00]), dizziness (RR, 2.17 [95% CI, 1.27 to 3.73]), back pain (RR, 1.96 [95% CI, 1.27 to 3.01]), and muscle spasms (RR, 3.26 [95% CI, 2.06 to 5.16]). Association by individual HT is shown in Figure 4A. The Data Supplement (Table S3) shows representative example messages that mention hot flush, arthralgia, and weight increased. Many posts reported “severe,” “horrible,” or “terrible” hot flushes with HT, which sometimes affected the quality of sleep or the ability to exercise. Arthralgia was reported as “serious” or “terrible” in some posts and affected ease of movement. Some posts suggested exercise/stretching and dietary supplementation to alleviate arthralgia. Posts often mentioned that weight gain was their worst symptom from HT.

FIG 4.

RRs for statistically significant (P ≤ .05) associations between HTs and (A) symptoms and (B) impacts mentioned in eBC posts. All symptoms with at least one statistically significant association with an HT were identified and ranked by their total RR across all HTs. The top 10 symptoms and nine impacts with significant associations with HTs are shown here. eBC, early breast; HT, hormone therapy; RR, relative risk.

Associations Between HTs and Impacts

A total of nine of the 10 impacts included in the analysis were identified as having a significant association with at least one HT. Overall, mood swings (RR, 7.36 [95% CI, 5.75 to 9.42]), insomnia (RR, 4.76 [95% CI, 3.14 to 7.22]), and depression (RR, 3.05 [95% CI, 1.71 to 5.44]) showed the strongest associations with HT. Association by individual HT is shown in Figure 4B. Data Supplement (Table S3) shows representative example messages that mentioned mood swings, insomnia, and depression. Posts reported “terrible” or “intolerable” mood swings accompanied by “feelings of doom” and “dark episodes,” which often affected personal and professional relationships. Insomnia was described as “bad,” “horrible,” and “severe” and affected quality of life. Posts mentioning depression reported “bleak” and “very dark” feelings that sometimes led to persistent crying and suicidal thoughts.

Evaluation of Pain Locations

For eBC posts that mentioned nonspecific pain in which at least one HT was mentioned (n = 4,000), the most commonly reported locations were the hips and arms, followed by the pelvis, uterus, and ovaries (Fig 5).

FIG 5.

Prevalence of nonspecific pain by location included in eBC posts mentioning at least one HT. eBC, early breast cancer; HT, hormone therapy.

Evaluation of Hot Flush Severity

For eBC posts that mentioned hot flush in which at least one HT was mentioned (n = 2,299), severity levels were mild in 12.5% and severe in 14.8% of posts; severity levels were not specified in 72.7% of posts.

Insights Into the Management of Symptoms/Impacts Using Exercise and Yoga

Nine cluster themes were identified among eBC posts that mentioned exercise and eight cluster themes mentioned yoga; four each were selected for further analysis (Data Supplement, Fig S1). eBC posts within each cluster that mentioned symptoms/impacts of hot flush, arthralgia, fatigue, anxiety, or depression, and specifically the sentiments describing these symptoms/impacts, were examined (Fig 6). Across all symptoms and clusters for exercise-related eBC posts, 53.3%-82.5% reported positive or neutral sentiments; this was 22.2%-100.0% for yoga-related eBC posts. In the cluster Managing side effects of HT between 50.0% and 63.5% of posts had a positive sentiment associated with exercise with respect to each of the symptoms/impacts; this was somewhat lower (between 11.1% and 43.0%) for posts relating to yoga. Among exercise posts, cluster 6 (Pain management and exercise) included a higher proportion of negative versus positive sentiments for depression. Similarly, among yoga posts, cluster 2 (Managing side effects of HT) had a higher prevalence of negative versus positive sentiments for anxiety and depression.

FIG 6.

Sentiments related to (A) exercise and (B) yoga within select clusters of eBC posts. eBC, early breast cancer.

DISCUSSION

In our study, NLP and ML were used to identify the overall real-world prevalence of symptoms and impacts in patients with eBC, determine associations between symptoms/impacts and HT, and extract patient insights from over 360,000 online forum posts on Breastcancer.³⁰ This social listening study found that patients with eBC in the real-world setting experienced a wide range of symptoms and impacts that were significantly associated with HT, including hot flush, arthralgia, weight increased, mood swings, insomnia, and depression, with several side effects (including back pain and muscle spasms) having a significant association with HT despite their low overall prevalence (<1%) in eBC-related posts. When comparing the overall prevalence of symptoms in eBC posts with the prevalence of symptoms among those that mentioned HT, three of the five highest ranking symptoms differed, with hot flush, arthralgia, and lymphedema becoming more prominent in posts mentioning HT. Similar results were observed for the impact analysis, albeit less pronounced, with depression becoming the second most prevalent impact when an HT was mentioned. These findings are generally consistent with the known safety profiles of HTs based on data from global clinical trials.^42-47

Analysis of pain location among eBC posts that mentioned an HT showed that pain in the hips and arms was most common. In the analysis of hot flush severity, the majority (73%) of eBC posts that mentioned an HT did not specify the severity level of a hot flush, with 13% and 15% of posts reporting mild and high severity, respectively. Furthermore, eBC posts describing the use of exercise and yoga to manage symptoms/impacts were predominantly positive. These results demonstrate that the NLP and ML methodology can be used to identify population-level insights into eBC disease experience and further examine subpopulations of interest.

Several actionable insights were identified in our study that could support future clinical research. These include the identification of symptoms and impacts experienced by patients on HT in the real world, as opposed to those observed in the controlled setting of a clinical trial. Insights could also shape educational initiatives for patients with eBC on the basis of methods used by patients to manage their symptoms such as exercise and yoga. Additional insights not captured in our study could include symptoms that lead to discontinuation of HT in the real-world setting, thus helping patients manage the symptoms to improve adherence to treatment. Future work addressing the more exhaustive analysis of interventions will help to better quantify patients' adherence to treatment.

Social listening offers the advantage of capturing spontaneous, top-of-mind insights from hundreds or thousands of patients from target populations with diverse backgrounds without introducing interviewer bias. This approach enables a breadth of patient experiences to be captured from a large, diverse patient population and still allows for in-depth exploration of specific research questions, particularly for rarer events or experiences that may be missed in small interview or focus group studies. For example, we were able to identify and examine over 50 posts specifically about severe hot flush related to HT and found it is a key factor in the decision to pause or switch treatment and can lead to treatment discontinuation. A traditional qualitative study on this topic would typically involve 15-30 patients and likely recruit only patients currently on HT to avoid recall bias.^48,49 Among these patients, it is possible that only a handful would have experienced severe hot flush and even fewer would have discontinued treatment because of hot flush. Furthermore, NLP and ML methodology can complement traditional patient interviews or focus groups by providing an initial readout of patients' perspectives, thus informing clinical outcome assessment measurement strategies for clinical trial development.¹⁶ Social listening may also offer unique overarching insights that may not be obtained via qualitative approaches due to challenges in identifying patients with specific experiences who can answer a specific research question.

Methodologically, NLP/ML allows us to (1) scale analysis to unseen before volumes; (2) detect rare events usually not seen in limited subsets of patients in interviews; (3) remove interviewer bias and capture purely spontaneous patient reports; (4) quickly develop solutions answering new questions; (5) incorporate new data sources that cannot be processed by analysts because of volume; and (6) significantly speed up the process of feedback collection compared with human involved process.

Although social listening is acknowledged as an acceptable approach to offer supplementary information about the patient's experience of disease and treatment,^16,50 this method has a number of limitations. Limited information is available on the sociodemographic and clinical characteristics of patients posting on social media, and posts cannot be validated to ensure they are from patients with eBC nor can ambiguous posts be followed up on to clarify meaning. In our study, the posts came from a reliable BC-specific community forum, increasing our confidence that the posts were legitimate, and if descriptions were unclear or challenging to interpret, they were excluded from the analysis. Additionally, patient identification may be withheld to increase data anonymization, making associations between posts belonging to the same patient impossible, as was the case in our study.

Using the NLP and ML methodology, real-world patient experiences from eBC-specific and, among those, HT-specific posts were captured from a large and diverse group of patients that allowed specific research questions to be explored, informing future clinical research in the field of eBC. The use of NLP and ML presents a valuable contribution in demonstrating how real-world data combining with an AI approach can provide insights and further understanding and improve patients' experience with eBC through clinical research. To our knowledge, this is the first social listening study that captures the patients' perspective on eBC and offers actionable insights into real-world patient experiences of HT that can be used to improve patient care.

AUTHOR CONTRIBUTIONS

Conception and design: Sameet Sreenivasan, Emuella M. Flood, Natasha Markuzon, Jasmine Y.Y. Sze

Collection and assembly of data: Chao Fang, Natasha Markuzon

Data analysis and interpretation: Sameet Sreenivasan, Emuella M. Flood, Natasha Markuzon, Jasmine Y.Y. Sze

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).