Discovery of clinical and demographic determinants of symptom burden in primary brain tumor patients using network analysis and unsupervised clustering

Abstract

Background

Precision health approaches to managing symptom burden in primary brain tumor (PBT) patients are imperative to improving patient outcomes and quality of life, but require tackling the complexity and heterogeneity of the symptom experience. Network Analysis (NA) can identify complex symptom co-severity patterns, and unsupervised clustering can unbiasedly stratify patients into clinically relevant subgroups based on symptom patterns. We combined these approaches in a novel study seeking to understand PBT patients’ clinical and demographic determinants of symptom burden.

Methods

MDASI-BT symptom severity data from a two-institutional cohort of 1128 PBT patients were analyzed. Gaussian Graphical Model networks were constructed for the all-patient cohort and subgroups identified by unsupervised clustering based on co-severity patterns. Network characteristics were analyzed and compared using permutation-based statistical tests.

Results

NA of the all-patient cohort revealed 4 core dimensions that drive the overall symptom burden of PBT patients: Cognitive, physical, focal neurologic, and affective. Fatigue/drowsiness was identified as pivotal to the symptom experience based on the network characteristics. Unsupervised clustering discovered 4 patient subgroups: PC1 (n = 683), PC2 (n = 244), PC3 (n = 92), and PC4 (n = 109). Moderately accurate networks could be constructed for PC1 and PC2. The PC1 patients had the highest interference scores among the subgroups and their network resembled the all-patient network. The PC2 patients were older and their symptom burden was driven by cognitive symptoms.

Conclusions

In the future, the proposed framework might be able to prioritize symptoms for targeting individual patients, informing more personalized symptom management.

Primary brain tumor (PBT) patients often experience debilitating disease and treatment-related symptom burdens. Over half of these patients endure 10 or more concurrent symptoms throughout their disease course,¹ and patients consistently describe alleviating symptoms as one of the issues most important to them.² However, choosing which symptoms to prioritize for treatment or prevention in individual patients remains challenging. Symptom presentation varies across patients, and co-occurring symptoms can interact with and reinforce each other.³ Unfortunately, traditional methods for analyzing symptom burden are inadequate for understanding such complex symptom interactions and capturing the heterogeneity of PBT patients’ symptom experiences. Analyzing the occurrence or severity of individual symptoms in isolation does not give insight into how one symptom may trigger or exacerbate other symptoms, how a group of symptoms may frequently co-occur due to a shared etiology, or how a symptom may differentially impact subgroups of patients.⁴ To address these limitations, this study uses network analysis (NA) and unsupervised clustering methods to discover and understand patients’ complex symptom interaction patterns, identify distinct subgroups of patients with unique symptom patterns, and inform tailored, patient-centered symptom care based on such patterns.

NA is a graph theory-based methodology that can be used to model symptom co-occurrence and co-severity patterns.^5–7 It has been extensively used to understand the relationships between symptoms in psychiatric disorders,⁸ but it has only recently been applied to the cancer context.^9–19 The only application of NA to PBT patients to date is a 2019 study from Coomans et al. who analyzed symptom co-severity patterns in 4307 newly diagnosed glioma patients, constructed an all-patient Gaussian Graphical Model (GGM) symptom network, and discovered 4 symptom clusters.²⁰ This study had an impressive sample size and was foundational in establishing the use of NA in PBT patient symptom care, but it did not address the statistical accuracy and robustness of the constructed symptom network and the discovered symptom clusters. The approach that Coomans et al. used to discover symptom clusters also resulted in half of the considered symptoms remaining isolated outside of the clusters, complicating the interpretation of the interplay among symptoms. Building off the foundational work by Coomans et al., our study addresses these limitations in an independent cohort of 1128 PBT patients, providing a unified framework for rigorous application of NA to PBT symptom data in the future.

To the best of our knowledge, this is the first study in any cancer type to combine unsupervised clustering with NA to identify subgroups of patients with potential clinical relevance. Specifically, we combine GGM analysis with a generalization of the second-order unsupervised clustering method proposed in Henry et al.⁴ Second-order unsupervised clustering is distinct from traditional first-order clustering methods, such as hierarchical clustering and latent class analysis,²¹ because it uses symptom co-severity or co-occurrence measurements to cluster patients, rather than direct symptom measurements. As a result, whereas traditional analyses group patients largely based on low and high symptom burden,²² second-order clustering groups patients based on their unique symptom patterns, potentially providing greater insight into the distinct symptom experiences of subgroups of patients.

Ultimately, the aim of this study is to use NA and unsupervised clustering to inform precision health-based symptom care for PBT patients and provide a framework for future analyses.

Methods

Patients and Settings

This secondary analysis used cross-sectional data from a two-institutional cohort of 1128 PBT patients enrolled at the National Institutes of Health (NIH; Bethesda, MD, USA) in the Neuro-Oncology Branch’s Natural History Study (NHS; NCI 16-C-0151), and clinical studies from the MD Anderson Cancer Center (MDACC; Houston, TX, USA) published from 2006 to 2015.^3,23 The studies were approved by the ethical committees of their institutions, patients gave informed consent to participate, and principal investigators gave permission for use of the de-identified data in this retrospective study.

Materials

Demographic and clinical information (overviewed in Table 1) were collected at patients’ first visit using standardized forms. Each patient also completed the MD Anderson Symptom Inventory – Brain Tumor (MDASI-BT) module, reporting on the severity of 22 symptoms, and 6 measures of interference with daily life, experienced in the previous 24 hours on a 0–10 severity scale, with 0 being “not present” and 10 being “as bad as you can imagine” (Supplementary Table 1).³

Network Analysis

MDASI-BT 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 symptom severities, and the weights of the edges (line thickness) represent their degree of association after controlling for the contribution of the remaining symptoms.^6,7

We determined the optimal number of symptoms to include in networks using the Unique Variable Analysis (UVA) approach as implemented in the R package EGAnet.²⁴ UVA assesses topological overlap of variables to identify potentially redundant pairs; these can be consolidated into a single underlying variable using the Maximum Likelihood with Robust standard errors estimator.²⁵

Following redundant variable consolidation, GGMs were constructed using the R packages qgraph²⁶ and EGAnet.²⁴ All variables were assumed to be continuous and were determined to be non-normally distributed using the Shapiro-Wilk normality test (Supplementary Figure 1A). As recommended, nonparanormal transformation was applied prior to GGM construction.^7,27 GGMs were regularized using the Graphical Least Absolute Shrinkage and Selection Operator based on Extended Bayesian Information Criterion (EBIC-GLASSO). This step removes potentially spurious edges from the network, leading to more accurate network architectures.^5–7

Symptom clusters in each GGM were identified using the R package EGAnet, which uses the walk trap algorithm to identify node communities within networks.²⁸ These symptom clusters were compared with clusters discovered using principal component analysis with varimax rotation, which was the clustering method used during initial construct validation of the MDASI-BT.³ Symptom importance in each GGM was assessed by calculating network centrality measures including strength, betweenness, and 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 (that is, how often a symptom serves as the common factor between other symptoms); and closeness measures how well a node (symptom) is indirectly connected to other nodes (symptoms).⁶ These centrality measures reflect which symptoms play a key role in exacerbating other symptoms and contribute the most to overall symptom burden. They were calculated and plotted with the R package qgraph.²⁶ Additionally, bridge centrality measures were calculated using the R package networktools.²⁹ Bridge centrality measures are calculated between nodes in different communities, and they provide insight into which symptoms are responsible for the connection between different symptom clusters.³⁰

Network accuracy (how sensitive to sampling variation a network is) and stability (whether the interpretation of a network remains consistent when fewer observations are considered) were assessed using permutation-based statistical tests: Tests assessing edge weight accuracy,⁶ the centrality and edge weight stability,⁶ and the stability of symptom clusters.³¹ GGMs from different patient subgroups were compared to determine differences in overall network structure and individual symptom centrality measurements.³²

Second-Order Unsupervised Patient Clustering

Patients were divided into subgroups using unsupervised clustering on co-severity of symptoms. The clustering used here generalizes the concordance network clustering approach described in Henry et al.,⁴ and uses continuous symptom co-severity (as opposed to binary co-occurrence) data to identify communities of patients based on the similarity of their co-severity patterns.

Clinical and Demographic Characteristics Comparison

The clinical and demographic characteristics of different patient communities were compared using the independent t-test for continuous data and Fisher’s exact test for categorical data. Bonferroni correction was applied to account for multiple comparisons.

A graphical schematic of the study design is shown in Figure 1.

Results

Patient Population

A total of 1128 PBT patients from clinical studies conducted at National Institutes of Health and MD Anderson Cancer Center (MDACC) were analyzed (Table 1). All patients completed the MDASI-BT in full, except for 15 patients who did not complete the interference scores survey. These patients were included in symptom analysis but excluded from interference score analysis. Most patients were male (58.4%), white non-Hispanic (81.6%), had a KPS $≥$ 90 (68.7%), had primary (not recurrent) disease (57%), and were currently off treatment (69.9%). 48.0% of patients had Grade 4 tumors, as defined by the 2016 World Health Organization classification of tumors of the central nervous system,³³ whereas 25.5%, 21.7%, and 3.5% of patients had Grade 3, 2, and 1 tumors, respectively. The most common diagnosis was glioblastoma (31.4%), however, information on specific diagnosis was not available for 26.4% of patients due to the legacy status of the MDACC sample, which also precluded reclassification according to the 2021 World Health Organization guidelines.³⁴

Symptom Prevalence/Severity and Interference Scores

In total, 1054 patients (93.4%) reported having at least one symptom, and 675 patients (59.8%) reported having at least one symptom with a severity ≥5. A total of 14.5% of patients reported between 1 and 4 concurrent symptoms, 35.7% reported between 5 and 10, 40.3% reported between 10 and 20, and 32 patients (2.8%) reported over 20 concurrent symptoms. Fatigue was the most prevalent and severe symptom, followed by drowsiness, difficulty remembering, distress/feeling upset, and disturbed sleep (Figure 2). 
The severity of the 22 symptoms was assessed to determine redundancy between symptoms (that is, whether any pair of symptoms measured the same underlying construct). Following this assessment, 2 variable pairs, fatigue & drowsiness, and nausea & vomiting, met the criteria for redundancy because they exhibited significant topological overlaps, conceptual redundancy, and floor/ceiling occurrence effects (Supplementary Figure 1B, Figure 2). Each pair was then consolidated into a single latent variable; for example, the pair fatigue & drowsiness was consolidated into the symptom fatigue/drowsy, whose severity was computed as a mixture of the severity of fatigue and drowsiness. Thus, in the proceeding NA, 20 symptoms were analyzed (18 original, and 2 consolidated pairs).

In total, 875 patients (77.6%) reported at least one aspect of interference with daily living caused by their symptoms, with 788 (69.9%) reporting at least one activity-related interference and 791 (70.1%) reporting at least one mood-related interference. Across the entire cohort, the mean activity and mood-related interference scores were 2.63 and 2.21, respectively, and the median activity and mood-related interference scores were 1.67 and 1.33, respectively (Table 1).

Symptom Clusters

A regularized Gaussian Graphical Model (GGM) was constructed using the 20-symptom severity data from the entire cohort (Figure 3A). Permutation-based statistical tests verified that the network edge weights were both accurate and stable (Supplementary Figure 2A, B). The walk trap algorithm, which identifies the underlying set of dimensions that explain most of the variability in a network, discovered 4 distinct symptom clusters, indicating that the symptom experience of PBT patients spans 4 core symptom dimensions. Based on the symptoms in each cluster, these 4 dimensions were identified as cognitive, physical, focal neurologic, and affective (Figure 3A).

Permutation-based testing was used to evaluate the stability of the symptom clusters—randomly resampling the original dataset with replacement a large number of times, repeating the symptom cluster discovery procedure for each permutation, and calculating how often each symptom occurred in the same cluster as in the original analysis.³¹ Indeed, the 4 symptom clusters were stable, as exactly 4 dimensions were identified in 86% of the 10 000 randomly sampled permutations; all symptoms except for impaired vision were found in the same cluster in over 80% of permutations (Figure 3B); and the cognitive, physical, focal neurologic, and affective clusters were exactly replicated in 99.8%, 72.5%, 50.6%, and 82.7% of permutations, respectively. Despite the focal neurologic cluster appearing relatively unstable compared to the remaining clusters, the replication of the cluster for over 50% of the permutations is highly significant compared to the probability of randomly selecting the exact 4 member symptoms out of the 20. Moreover, this instability appears driven by the instability of impaired vision, which had only 58% stability and substantial cross-loadings in the physical and cognitive clusters (Supplementary Table 2). Given the impact impaired vision can have on both cognitive and physical functioning, this cross-loading both makes conceptual sense and led us to hypothesize that impaired vision may serve as an important connection between the focal neurologic, cognitive, and physical clusters. This finding suggests that impaired vision may be a bridge by which focal neurologic symptoms impact cognitive and physical symptoms.

Comparing these symptom clusters to clusters determined using principal component analysis with varimax rotation, which was used by Armstrong et al. in the original validation of the MDASI-BT,³ we found substantial overlap between the 2 methods (Supplementary Table 3). Only changes in vision, fatigue/drowsiness, and disturbed sleep were placed in different clusters. This is likely explained by the relatively low item stability of change in vision and the high cross-loadings of fatigue/drowsiness and disturbed sleep between the physical and affective symptom clusters (Supplementary Table 2). Thus, we also hypothesized that fatigue/drowsiness and disturbed sleep may serve as bridge symptoms between the physical and affective dimensions.

Symptom Centralities

To determine which symptoms are most important to patients’ overall symptom burden and may be most important to target, we calculated the symptom centrality measures of strength, closeness, and betweenness. All centrality measures were found to be stable using permutation-based statistical tests (Supplementary Figure 2C). We found that fatigue/drowsiness had the highest strength, closeness, and betweenness centrality of all symptoms in the network (Figure 3C), indicating that fatigue/drowsiness 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 in all 3 centrality measures included distress/feeling upset, difficulty concentrating, and difficulty remembering.

Next, we calculated bridge centralities to determine which symptoms connect the different symptom clusters, and to formally test the bridge symptom hypotheses posed in the previous section. As the term indicates, bridge centrality gives insight into which symptoms may provide links between the core dimensions and likely precede or exacerbate the severity of multiple symptom dimensions. All bridge centrality measurements were highly stable according to the permutation-based test (Supplementary Figure 2D). Fatigue/drowsiness had the highest bridge strength, bridge closeness, and bridge betweenness (Figure 3D), and its symptom cluster cross-loadings suggest that it primarily serves to bridge the physical, affective, and cognitive dimensions (Supplementary Table 2). Other symptoms that demonstrated high values across all 3 bridge centrality measurements included irritability (clusters with substantial cross-loading: Affective, cognitive, and physical), change in appearance (physical, focal neurologic, and affective), distress/feeling upset (affective and physical), difficulty concentrating (cognitive and affective), and difficulty remembering (cognitive and physical). Furthermore, as hypothesized above, problems with vision (focal neurologic, cognitive, and physical) and disturbed sleep (physical, affective) had relatively high bridge strength and bridge closeness, respectively, suggesting that they also serve as important bridge symptoms.

Patient Subgroups

After analyzing symptom network properties across the entire patient cohort, we asked whether patients could be divided into clinically relevant subgroups based on unique symptom patterns. We hypothesized that the heterogeneity in PBT patients’ symptom presentation would also be reflected in their symptom severity and co-severity patterns. To test this hypothesis, we used unsupervised clustering based on symptom co-severity patterns and identified 4 patient subgroups, or communities, referred to as PC1 (n = 683), PC2 (n = 244), PC3 (n = 92), and PC4 (n = 109) (Figure 4A). Patients in PC1 had the highest average severity and occurrence across all symptoms, and PC4 had the lowest (Figure 4B). As we had hypothesized, the 4 groups demonstrated unique symptom severity and occurrence patterns. PC1 had high severity and occurrence across all 4 symptom cluster categories. On the other hand, PC2 had particularly high spikes in the severity and occurrence of symptoms in the cognitive (difficulty understanding, difficulty concentrating, difficulty speaking, and difficulty remembering) and focal neurologic clusters (numbness/tingling, problems with vision, and weakness on one side of the body); PC3 had spikes in the physical (change in bowel pattern, pain, dry mouth, disturbed sleep, and fatigue/drowsiness) and focal neurologic clusters (numbness/tingling, problems with vision, and weakness on one side of the body); and PC4 had very low severity and occurrence across all symptoms.

Regularized networks could be constructed for PC1 and PC2 (Figure 5A, B), but not for PC3 or PC4. The small sample sizes of PC3 and PC4 led to networks in which all edges were removed upon regularization. Permutation-based testing was used to assess the accuracy and stability of each network. PC1 had moderate to high accuracy and stability, and PC2 had low to moderate accuracy and stability due to its smaller sample size (Supplementary 
Figure 3).

The network for PC1 closely resembled the network for the entire patient cohort, organized along the 4 core dimensions of cognitive, physical, focal neurologic, and affective symptoms (Figure 5A). As in the all-patient network, the symptoms with high strength, closeness, and betweenness centralities were fatigue/drowsiness, difficulty concentrating, distress/feeling upset, and difficulty remembering (Figure 5C). Furthermore, as in the all-patient network, symptoms with high bridge centralities included fatigue/drowsiness, irritability, difficulty concentrating, change in appearance, distress/feeling upset, and difficulty remembering (Figure 5D). These findings suggest that the conclusions from the all-patient network largely represent the symptom experience of patients in PC1.

PC2, on the other hand, appeared to represent a subset of patients with distinct symptom dimensions and overall symptom patterns. Only 3 symptom clusters were identified in the network for PC2, and the clusters did not correspond to any clear symptom pattern (Figure 5B). The symptoms in PC2 with the highest centralities and bridge centralities were difficulty understanding, difficulty remembering, and difficulty speaking, which are all cognitive symptoms. Compared to PC1, PC2 had significantly higher centralities and bridge centralities for symptoms in the cognitive (difficulty understanding, difficulty remembering, and difficulty speaking) and focal neurologic clusters (numbness, weakness, and impaired vision), but lower centralities and bridge centralities for symptoms in the affective cluster (distress/feeling upset, sadness, and irritability) (Figure 5C, D). Furthermore, when the 2 networks were compared statistically, they were shown to have significantly different network architectures (p = 0) (Supplementary Figure 4A, C) and significantly different global strength (overall interconnectivity among the symptoms), with PC1 having stronger interconnectivity (p = 0) (Supplementary Figure 4B, D). Thus, PC1 and PC2 had significantly different symptom dimensions and co-severity patterns, with the overall symptom burden for patients in PC1 driven by 4 core symptom dimensions, and the overall symptom burden for patients in PC2 driven by primarily cognitive symptoms, but also partly by focal neurologic symptoms.

Comparison of Patient Subgroups’ Clinical and Demographic Characteristics

Given the differences between the symptom experiences of the patient subgroups, we next asked whether the subgroups differed in their clinical and demographic determinants. We compared age, Karnofsky Performance Status (KPS), sex, ethnicity, education status, employment status, income, diagnosis, tumor grade, tumor recurrence status, treatment status, tumor side, lobe location of tumor, and MDASI-BT interference item scores for patients with available data (Table 1). Significant differences were found between patient subgroups for the MDASI-BT interference item scores, age, and KPS. Most notably, patients in PC1 had significantly higher overall, activity-related, and mood-related interference scores than all other subgroups (Figure 4D); patients in PC2 were significantly older than patients in PC1 and PC4 (Figure 4C); and patients in PC4 had significantly lower interference scores and higher KPS than patients in all other subgroups (Figure 4C, D).

Discussion

This novel study provides a framework by which NA and unsupervised clustering can be used to better understand complex symptom patterns in brain tumor patients, stratify patients into clinically relevant subgroups, and inform precision-health-based symptom care. The analysis of symptom severity data from 1128 PBT patients showed that most patients experienced symptoms across 4 core dimensions (cognitive, physical, focal neurologic, and affective). These dimensions substantially overlapped with factor groupings discovered using principal component analysis with varimax rotation, which was the method used by Armstrong et al. during initial construct validation of the MDASI-BT.³ The symptom fatigue/drowsiness scored the highest in all centrality measures, indicating that it had the greatest impact on the severity of the other symptoms and may have served as a bridge by which physical symptoms led to the onset or exacerbation of affective and cognitive symptoms. Thus, for most patients, fatigue/drowsiness may be important to prioritize during symptom management, especially early in care to avoid triggering or exacerbating other symptoms. Other symptoms with high centrality and bridge centrality included distress/feeling upset, difficulty concentrating, difficulty remembering, and irritability. Additionally, change in appearance and nausea/vomiting had relatively high centrality even though they had low severity and occurrence, suggesting that these symptoms have a disproportionate impact on patients’ overall symptom burden and that targeting these symptoms may be effective in preventing symptom clusters from exacerbating each other. Other studies have independently reported the presence of similar symptom clusters,^3,35,36 the driving effect fatigue/drowsiness has in PBT patients,^37,38 and the disproportionate impacts change in appearance and nausea/vomiting have in cancer patients,^12,39 but this study represents the most comprehensive integration of all these findings using a single mode of analysis.

Clinically relevant subgroups of patients with unique symptom patterns were discovered through unsupervised clustering. The PC1 subgroup encompassed the majority (61%) of patients and exhibited the highest average symptom severity and occurrence, as well as the highest activity and mood-related interference scores, across all the subgroups. This correlation highlights the bi-directional relationships between symptom burden and physical impairment, and symptom burden and mood, that have been reported in other studies^40,41 and were recently emphasized by a Response Assessment in Neuro-Oncology (RANO) collaborative report between the NCI, FDA, EORTC, and EMA.² On the other hand, the PC2 subgroup, which comprised 22% of patients, represented older patients for whom cognitive symptoms drove the overall network structure. This was indicated by the significantly higher centrality and bridge centrality for cognitive symptoms (difficulty understanding, difficulty remembering, and difficulty speaking), and to a lesser extent focal neurologic symptoms (numbness, weakness, and impaired vision), than the PC1 symptom network, suggesting that these symptoms should be prioritized for management in this subgroup. This finding is consistent with reporting that age is a risk factor for severe cancer-related cognitive impairment,⁴² and it suggests that both pharmacologic and non-pharmacologic interventions that preserve or enhance cognitive functioning^43,44 may be important to prioritize in the PC2 subgroup. Moreover, further studies should evaluate whether medications for non-cognitive symptoms that have cognitive side effects (such as antiepileptics, anti-inflammatory therapies, and pain medications⁴⁴) may unintentionally exacerbate the overall symptom burden within this subgroup.

Symptom networks could not be constructed for the smaller patient subgroups PC3 and PC4, which represented 8% and 10% of patients respectively, but analysis of the symptom occurrence and severity in these clusters suggested that they had unique symptom patterns. Patients in PC3 exhibited particularly high occurrence and severity for treatment-related symptoms such as problems with vision, weakness, numbness, dry mouth, pain, and changes in bowel patterns, raising the hypothesis that side effects of treatment may drive overall symptom burden in PC3. Furthermore, patients in PC4 exhibited low severity across all symptoms, identifying a subpopulation of PBT patients with minimal overall symptom burden, high KPS, and low interference scores. Interestingly, 30.3% of patients in PC4 had Grade 3 tumors, 34.9% had Grade 4 tumors, and 28.4% were currently on treatment, indicating that PC4 does not simply represent patients with low-grade tumors who are off treatment.

Several challenges and limitations point to the need for further validation of our preliminary findings. First, the patient population was highly heterogeneous with respect to diagnosis and timing of assessment during disease course, limiting evaluation of their impact on the symptom patterns. Second, clinical and histopathology data were missing for a large part of the sample, further complicating assessment of the determinants of symptom patterns. Third, the lack of information on comedications makes it impossible to disambiguate side effects of comedications from treatment or disease-related symptoms. Fourth, the cross-sectional design of the study precluded analysis of symptom networks’ changes over time and whether these changes can be predicted or prevented. Nevertheless, our study undertook several steps to ensure the robustness of the computational results. These included the use of network regularization (which reduced the likelihood of false positive associations between symptoms and led to discarding of unreliable networks for the smaller subgroups of patients), the employment of a series of permutation-based statistical tests to formally assess the stability of the network structures and the predicted importance of the symptoms, and the use of a robust community detection algorithm to determine the symptom clusters and patient communities.

Ultimately, the proposed framework has the potential to inform precision-health-based symptom care by giving patients and providers the ability to interpret symptom severity as the result of the complex interplay of multiple interacting and reenforcing symptoms. Over the past 2 decades, symptom management research has undergone a paradigm shift from studying symptoms in isolation to trying to understand how symptoms synergistically affect the patient quality of life and outcomes; however, clinical practice lags far behind, as symptoms are often still treated in isolation.⁴⁵ Our study provides a framework by which to bridge that gap, but highlights the need for longitudinal large-scale studies that would accrue frequent reporting of patient symptom severity alongside detailed clinical and demographic information. Only through such initiatives, which may require leveraging multi-institutional collaborations, partnership with local community clinics, adoption of novel applications that facilitate symptom reporting, and utilization of powerful computational frameworks, can precision-health-based symptom care be fully realized.

Funding

This research was supported by the Intramural Research Program of the National Cancer Institute, National Institutes of Health.

Conflict of interest statement. None declared.

Authorship statement

Study design and manuscript writing: BHB and OC. Analysis: BHB, OC, and OG. Contributions to study design and manuscript writing: TSA and MRG. Data collection: EV, A Christ, ALK, AA, A Choi, HEL, TM, LB, EB, NL, MP, MPP, TP, LP, JW, MRG, and TSA. All authors read and approved the manuscript.

References