Symptom Network Analysis and Unsupervised Clustering of Oncology Patients Identifies Drivers of Symptom Burden and Patient Subgroups With Distinct Symptom Patterns

2.1. Patients and Settings

This secondary analysis used cross‐sectional data from 3088 cancer patients who participated in one of three studies (i.e., Chemotherapy‐Induced Peripheral Neuropathy (CIPN) study [25], Symptom Clusters Study [26, 27], COVID‐19 study [28, 29]), whose details were described previously and are summarized below.

2.2. Materials

Patients completed a demographic questionnaire, the Karnofsky Performance Status (KPS) scale [30], and the Self‐Administered Comorbidity Questionnaire [31]. For the CIPN and Symptom Clusters studies, medical records were reviewed for disease and treatment information. For the COVID‐19 study, patients reported their clinical information (Table 1).

Patients completed the Memorial Symptom Assessment Scale (MSAS) [32] that evaluated the occurrence, frequency, severity, and distress of 32 symptoms experienced in the past week. Severity scores that ranged from 0 to 4 (i.e., 0 = did not report the symptom, 1 = slight, 2 = moderate, 3 = severe, 4 = very severe) were used in the current analyses. The MSAS has well‐established validity and reliability [33, 34].

In addition, they completed the Multidimensional Quality of Life Scale‐Patient Version (MQOLS‐PV) that assesses four domains of QOL (i.e., physical, psychological, social, and spiritual well‐being) and overall QOL. Subscale and total scores range from 0 to 10 with higher scores indicating better QOL [35]. The MQOLS‐PV has well‐established validity and reliability [36, 37, 38, 39].

2.3. Network Analysis

MSAS symptom severity data were used to construct Gaussian Graphical Model (GGM) networks. In a GGM symptom network, the nodes of the network (circles) represent symptoms; the edges (lines) represent the existence of conditionally dependent relationships between the severity of the symptoms; and the weights of the edges (line thickness) represent their degree of association after controlling for the contribution of the remaining symptoms [40, 41].

The optimal number of symptoms to include in networks was determined using unique variable analysis (UVA), which assesses the topological overlap of variables to identify potentially redundant pairs. Then, these redundant pairs were consolidated into a single underlying variable using the maximum likelihood with robust standard errors estimator [42].

Following redundant variable consolidation, GGMs were constructed [43, 44]. All variables were assumed to be continuous and were determined to be non‐normally distributed using the Shapiro–Wilk normality test (Figure S1B). As recommended, nonparanormal transformation was applied prior to GGM construction [41, 45]. GGMs were regularized using the Graphical Least Absolute Shrinkage and Selection Operator based on Extended Bayesian Information Criterion. This step removes potentially spurious edges from network models and leads to more accurate network architectures [40, 41, 46].

Symptom clusters in each GGM were identified using the walk trap algorithm [47]. Symptom importance in each GGM was assessed by calculating network centrality measures (i.e., strength, betweenness, closeness). Strength is a measure of how well a node (symptom) is directly connected to other nodes (symptoms). Betweenness measures how important a node is in connecting other nodes (i.e., how often a symptom serves as the common factor between other symptoms). Closeness measures how well a node (symptom) is indirectly connected to other nodes (symptoms) [40]. These centrality measures suggest which symptoms play a key role in activating or exacerbating other symptoms and contribute the most to overall symptom burden. In addition, bridge centrality measures were calculated [48]. Bridge centrality measures are calculated between nodes in different clusters. They provide insights into which symptoms are responsible for the connections between different symptom clusters [49]. Network accuracy and stability were assessed using permutation‐based statistical tests [40, 50].

2.4. Unsupervised Patient Clustering

Patients were divided into subgroups using unsupervised clustering on co‐severity of symptoms. This type of clustering generalizes the concordance network clustering approach described by Henry and colleagues [22]. It uses continuous symptom co‐severity data to identify communities of patients based on shared co‐severity patterns. Briefly, for each patient with reported symptom severity vector X, it calculates a concordance matrix XX ^T /100 that represents the co‐severity of symptoms for that patient. For each pair of patients, the similarity between their concordance matrices is then calculated using the Adjusted Rand Index (ARI) [51] between the two matrices. This generates a similarity matrix which is then passed into the walktrap community detection algorithm from the R‐package igraph [52] (using default parameters) to identify patient communities. Additional details on the NA steps and unsupervised clustering were described previously [15].

2.5. Demographic and Clinical Characteristic Comparisons

Multinomial logistic regression was used to assess demographic and clinical characteristics of subgroup membership. Patient community 1 (PC1) was used as the reference group because it was the most demographically similar to the entire patient cohort (Table 1) Models were constructed with cancer diagnosis, age, sex, current treatment status, ethnicity, income, and KPS scores as covariates. Odds‐ratios and p values with Bonferroni correction were calculated to assess the individual covariates' impact on subgroup membership. The R package nnet was used for these analyses [53].

3.1. Patient Sample

Of the 3088 cancer patients (Table 1), most were female (80.8%), non‐Hispanic White (76.9%), and currently on treatment (64.9%). Breast cancer was the most common diagnosis (53.4%), followed by gastrointestinal (14.6%), gynecological (8.3%), and lung (5.3%) cancer. The mean and median ages of patients across all studies were 59.7 and 60.9 years, respectively. Mean and median KPS scores were 85.3 and 90, respectively.

3.2. Symptom Prevalence and Severity and QOL Scores

A total of 3019 patients (97.8%) reported having at least one symptom and 1687 patients (54.6%) reported having at least one symptom that was severe or very severe (severity ≥ 3). In the sample, 23.3% reported between 1 and 5 concurrent symptoms, 34.1% reported between 6 and 10, 41.9% reported between 10 and 20, and 6.5% reported over 20 concurrent symptoms. Across all of the patients, lack of energy was the most severe symptom, followed by difficulty sleeping, pain, numbness/tingling in hands and feet, worrying, feeling drowsy, difficulty concentrating, and problems with sexual interest or activity (Figure 1). Potential redundancy between the 32 symptoms was assessed (i.e., whether any pair of symptoms likely measured the same underlying construct) by evaluating their interdependence overlap, conceptual redundancy, and outlying occurrence frequency effects. The symptom pair of nausea and vomiting met the criteria for redundancy (Figure S1A, and Figure 1) and was subsequently consolidated into a single latent variable, whose severity was computed as a mixture of the severity of the two individual symptoms. Therefore, in the NA, 31 symptoms were analyzed (30 original and 1 consolidated pair).

Across the entire cohort, the mean total QOL score was 6.0 (median = 6.1, range = 0.8–10). Mean subscale scores were as follows: physical well‐being of 7.2 (median = 7.4, range = 0.3–10), psychological well‐being of 5.6 (median = 5.7, range = 0.3–10), social well‐being of 6.2 (median = 6.4, range = 0–10), and spiritual well‐being of 5.2 (median = 5.1, range = 0–10; Table 1).

3.3. Symptom Clusters

A regularized GGM network model was constructed using the 31‐symptom severity scores for the entire sample (Figure 2A). Permutation‐based statistical tests verified the accuracy and stability of the network edge weights. The symptom organization in the network was assessed for the presence of physical and psychological clusters or dimensions that explained the symptom experience. Three distinct symptom clusters were identified and labeled as constitutional, gastrointestinal‐epithelial, and psychological (Figure 2A).

Permutation‐based testing was used to evaluate the stability of the symptom clusters by repeating the symptom cluster discovery procedure for random resampling of the original dataset and calculating how often each symptom occurred in the same cluster. Exactly three clusters were identified in 74.1% of the 10,000 randomly sampled permutations. The gastrointestinal‐epithelial and psychological clusters were highly stable, as each individual symptom was found in the same cluster in over 85% of permutations (Figure 2B). The entire gastrointestinal‐epithelial and psychological clusters were replicated exactly in 87% and 86% of permutations, respectively. In contrast, the constitutional symptom cluster was not stable, as many individual symptoms in this cluster had low stability (Figure 2B) and the entire cluster was replicated exactly in only 8.6% of permutations. These results suggest that the constitutional symptom cluster serves as a “catch‐all” for generalized symptoms of cancer that do not group into the more well‐defined gastrointestinal‐epithelial or psychological symptom clusters.

3.4. Symptom Centralities

The high degree of connectivity in the network (i.e., 465 total connections with nonzero weights and 115 of these connections were deemed significant by permutation‐based testing at 95% confidence) suggests that activation of the symptom experience can be rapid and sustainable. To determine which symptoms are core to this activation and may have the most impact on deactivation of overall symptom burden when targeted, we calculated the symptom centrality measures of strength, closeness, and betweenness for each symptom and verified their stability using permutation‐based statistical tests. Of note, lack of energy had the highest strength, closeness, and betweenness centrality scores of all symptoms in the network (Figure 2C). This finding suggests that lack of energy substantially contributes to the severity of other symptoms and may drive many of the co‐severity relationships in the network. Other symptoms that had high values across all centrality measures (z‐score > 0 in all three measures) included: change in the way food tastes, nausea/vomiting, lack of appetite, difficulty concentrating, feeling that “I don't look like myself,” and hair loss (Figure 2C).

Next, we assessed which symptoms may provide activation links between the three symptom clusters and may precede or exacerbate the severity of multiple symptoms, by calculating bridge centralities and verifying their stability with permutation‐based statistical tests. Lack of energy, feeling that “I don't look like myself,” and difficulty concentrating had the highest bridge strength, bridge closeness, and bridge betweenness (Figure 2D).

Analyzing the network connections of these three symptoms, the symptoms in the network most strongly connected to (and likely activated by) lack of energy include the constitutional symptoms of feeling drowsy, pain, and difficulty sleeping; the gastrointestinal‐epithelial symptoms of nausea/vomiting and lack of appetite; and the psychological symptom of difficulty concentrating. The symptoms most strongly connected to feeling that “I don't look like myself” include the constitutional symptoms of feeling bloated and swelling of arms/legs; the gastrointestinal‐epithelial symptoms of changes in skin and hair loss; and the psychological symptoms of feeling sad, worrying, and decreased sexual interest. The symptoms most strongly connected to difficulty concentrating include the constitutional symptoms of lack of energy, difficulty sleeping, sweats, and feeling drowsy; and the psychological symptoms of feeling sad, feeling nervous, and feeling irritable (Figure 2A).

3.5. Patient Subgroups

The all‐patient sample analysis provides important insights into symptom clusters and dominant interaction patterns that occur across the entire sample of patients with heterogeneous types of cancer and who are at various points in their cancer trajectory. However, this type of analysis is likely to overshadow more nuanced symptom experiences of specific subgroups of patients. To assess and capture potential heterogeneity in patients' symptom experiences, we assessed the (dis)similarity of the symptom experience and sought to unbiasedly cluster patients based on shared symptom co‐severity patterns using random walk‐based unsupervised clustering. This assessment led to the identification of six patient subgroups or communities, referred to as PC1 (n = 851), PC2 (n = 639), PC3 (n = 198), PC4 (n = 232), PC5 (n = 995), and PC6 (n = 171) (Figure 3A). These clusters remained consistent across 11 separate runs of unsupervised clustering using different random walk starting positions. GGMs and network centrality measures were calculated for each subgroup (Figures S2 and S3), except for PC6. Because this subgroup had a minimal symptom burden (Figure 3A,G), the network was sparse and unstable. In addition, we compared the demographic and clinical characteristics of the subgroups (Table 1) and performed multinomial logistic regression to determine the characteristics that most strongly predicted subgroup membership (Table 2).

Because PC1 was demographically and clinically most representative of the entire cohort (Table 1), it was used as the reference group in multinomial logistic regression (Table 2). However, unlike in the network for the entire cohort, only a handful of symptoms (i.e., lack of energy, pain, numbness and tingling, difficulty sleeping, and drowsiness) were disproportionally represented in PC1 (Figure 3B). This subgroup appears to represent a cohort of cancer patients who are most affected by common constitutional symptoms of cancer but do not experience the gastrointestinal‐epithelial and psychological symptom clusters, or the more unique symptom patterns of some of the other subgroups. The symptoms with the highest centrality in PC1 (i.e., the most important to target in this subgroup) include lack of energy, pain, numbness and tingling, and lack of appetite (Figure S3A).

Multinomial logistic regression revealed that age, treatment status, and KPS score were the strongest determinants of PC2 membership (Table 2). PC2 was the youngest of the subgroups, had the most patients currently on treatment (92%), and had the lowest KPS scores. In addition, these patients had the lowest total, physical, psychological, and social QOL scores (Table 1). Patients in this subgroup had a very high symptom burden across all symptoms (Figure 3A,B) and their symptoms clustered into three groups: (1) constitutional, (2) psychological/changes in daily function, and (3) epithelial/changes in appearance (Figure S2B). Analyzing the centrality measurements, particularly important symptoms included: pain, lack of energy, multiple psychological symptoms (i.e., feeling worried, feeling irritable, difficulty concentrating), and lack of appetite (Figure S3B).

High KPS score was the strongest determinant of PC3 membership (Table 2). In addition, this group tended to be further out from their initial cancer diagnosis and less likely to be currently on treatment (42%) (Table 1). On average, these patients had low symptom severity scores except for numbness and pain (Figure 3D). The low symptom severity scores for this subgroup led to a sparse network (Figure S2C) which suggests that activation of individual symptoms is unlikely to lead to activation or exacerbation of other symptoms and that targeting of any severe individual symptoms is similarly effective. PC3 may represent patients who have a good prognosis or are recovered but continue to experience significant neuropathy.

Treatment status and ethnicity were the strongest determinants of PC4 membership (Table 2), as this group was more likely to be on treatment (90%) and was disproportionately Asian/Pacific Islander (18.1% compared to 8.0% across the entire cohort) (Table 1). This group had a high burden of treatment‐related symptoms such as hair loss, dry mouth, constipation, change in the way food tastes, and lack of appetite. However, compared to the mean severity scores across the entire sample, these patients had a much lower severity of psychological symptoms, less pain, less difficulty sleeping, and less drowsiness (Figure 3E). Their total, physical, psychological, social, and spiritual QOL scores were higher than average, despite having lower than average KPS scores and being most likely to have metastatic sites among all the groups (Table 1). The symptoms with the highest centralities in PC4 included lack of energy, dizziness, lack of appetite, and change in the way food tastes (Figure S3D).

Younger age and treatment status were the strongest determinants of PC5 membership (Table 2). This group was less likely to be currently on treatment (53%) and was mostly composed of female (87%) breast cancer patients (63%) (Table 1). Patients in PC5 experienced a disproportionately high burden of psychological symptoms compared to the overall sample (Figure 3F). They had lower than average total, psychological, social, and spiritual QOL scores, but average physical QOL and higher than average KPS scores (Table 1). This group may represent younger patients who may respond well to treatment but for whom the psychological burden of cancer is especially difficult. The symptoms with the highest centralities in PC5 included lack of energy, feeling sad, and feeling that “I don't look like myself” (Figure S3E).

The strongest determinant of PC6 membership was high KPS scores (Table 2), as patients in PC6 had the highest KPS scores out of all the subgroups (Table 1). This group experienced very low symptom burden across all symptoms (Figure 3G), was further out from diagnosis than average, was less likely to be currently on treatment (48%), was least likely to have metastatic sites, and had the highest QOL scores across all categories. A network could not be constructed for PC6 due to the sparsity of the co‐severity relationships.