Abstract
Objective
To characterize symptom interactions and identify candidate intervention targets in patients with esophageal cancer using network analysis and computer-simulated symptom perturbations.
Methods
This two-center cross-sectional study recruited 255 patients with esophageal cancer via convenience sampling between May 2020 and May 2023. Participants completed the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 (EORTC QLQ-C30) and the esophageal cancer-specific module (EORTC QLQ-OES18). A Gaussian graphical model (GGM) was used to characterize symptom associations and centrality; node predictability was estimated using a mixed graphical model (MGM); and a bootstrap-averaged Bayesian network was constructed to explore directed probabilistic associations; thresholds applied for network modeling were defined as analytic settings rather than validated clinical cut-offs. Ising model-based simulations (NodeIdentifyR algorithm) were then used to explore the potential dynamic effects of symptom-targeted strategies.
Results
Fatigue showed the highest static centrality, and appetite loss showed a directed association with sleep disturbance and taste disturbance. Simulated interventions—model-based projections rather than observed effects—indicated that alleviating appetite loss or fatigue reduced the expected number of active symptoms by 2.61 and 2.58, respectively. In the aggravation scenario, pain produced the largest projected increase; choking when swallowing produced the second-largest increase despite its low static centrality, indicating divergence between static and simulation-based findings.
Conclusions
Appetite loss and fatigue emerged as candidate targets associated with the largest projected changes in global network activation, while pain emerged as the candidate target associated with the largest projected increase, consistent with its structural prominence in the network; choking when swallowing emerged as an additional simulation-informed target despite modest static centrality. As model-based projections, these hypothesis-generating findings warrant prospective evaluation in future interventional studies.
Keywords: Esophageal cancer, Symptom network analysis, Computer-simulated interventions, Ising model
Introduction
According to GLOBOCAN 2022 statistics,1 esophageal cancer ranks as the 11th most commonly diagnosed malignancy and the seventh leading cause of cancer-related mortality worldwide, with approximately 511,000 new cases and 445,000 deaths reported annually. In China,2 more than 224,000 new cases were recorded every year, accounting for 43.8% of the global total. Consequently, the disease burden in China remains substantially higher than the global average, posing a serious public health challenge.3 A considerable proportion of patients with esophageal cancer are diagnosed at an advanced stage, and the overall five-year survival rate is approximately 30%.4 Surgical resection remains the primary treatment modality and is frequently combined with radiotherapy and chemotherapy. However, owing to tumor progression, the anatomical characteristics of the esophagus, postoperative gastrointestinal reconstruction, and the adverse effects of adjuvant therapies, patients commonly experience a constellation of symptoms, including dietary alterations, fatigue, reflux, and pain.5, 6, 7 These symptoms often interact, impair quality of life, and may adversely affect prognosis and survival.8,9 Understanding the complex patterns underlying symptom interactions and identifying potential targets for intervention are therefore essential for improving symptom management and alleviating patient burden.10 Network modeling11 provides a novel framework for symptom research by enabling the quantification of complex inter-symptom relationships and the identification of potentially central symptoms that may be associated with broader symptom cascades.
Undirected network analysis can be used to characterize associations among symptoms and to estimate node centrality, thereby facilitating the identification of potential intervention targets within complex symptom systems.12 Complementarily, Bayesian network analysis offers an additional framework for exploring potential directional dependencies among symptoms.13 This approach was not intended to establish deterministic causality but rather to allow the representation of directed pathways as probabilistic dependency structures. The principal advantage of these approaches lies in their ability to accommodate complex multivariate data and to identify potentially important nodes and pathways within the network.14
However, while the capacity of network analysis to identify potential intervention targets and nodes with directed probabilistic associations is clear, the clinical impact of targeting these symptoms remains unclear. Computational simulation, as an advanced analytic approach, transcends the limitations of traditional static models by enabling the exploration of dynamic processes within symptom networks.15 Early simulation studies typically implemented ‘targeted deactivation’ by constraining the activation probability of specific symptoms to zero in order to evaluate their influence on the overall network.16 However, the assumption of complete symptom elimination is often unrealistic in clinical settings and fails to account for the heterogeneity of symptom expression. In this context, an in silico network17 intervention based on the NodeIdentifyR algorithm for Ising networks offers a potentially useful alternative research strategy. This approach conceptualizes intervention as a dynamic regulatory process, simulating therapeutic effects by modifying the activation probabilities of target symptoms (e.g., through threshold adjustments). Such simulations enable systematic evaluation of how changes in a single symptom may propagate throughout the network and help identify nodes that exert a disproportionate influence on global network activation, thereby providing a scientific basis for precision symptom management.
In the present study, we applied NodeIdentifyR-based simulation to evaluate the effects of alleviating or exacerbating symptoms among patients with esophageal cancer and to identify potential candidate intervention targets. Specifically, the objectives of this study were to: (1) estimate the prevalence of symptoms in this population and characterize the symptom network structure, identifying central symptoms and symptom predictability; (2) model the probabilistic dependencies and directional associations among symptoms using a Bayesian network; and (3) simulate symptom-specific interventions using an Ising model with the NodeIdentifyR algorithm to identify potential candidate targets for clinical management and prevention.
Methods
Study design, setting, and participants
This two-center cross-sectional study was conducted at two tertiary hospitals in Jiangsu Province — Jiangnan University Affiliated Hospital and The First People’s Hospital of Changzhou — between May 2020 and May 2023. Of the 255 participants, 171 were recruited from Jiangnan University Affiliated Hospital and 84 from The First People’s Hospital of Changzhou, with identical data-collection and quality-control protocols applied at both sites.
We applied convenience sampling from inpatient departments at both sites. The inclusion criteria were as follows: (1) a pathological diagnosis of esophageal cancer and receipt of surgery, chemotherapy, or radiotherapy, (2) age 18 years or older; and (3) adequate cognitive and communication abilities to provide informed consent and complete the study questionnaires. The exclusion criteria were as follows: (1) presence of another concurrent malignancy, (2) advanced disease with an estimated life expectancy of ≤ 3 months.
Sample size estimation was based on the number of parameters to be estimated in the symptom network. With 18 symptoms included (n = 18), the threshold parameter = 18, the pairwise correlation parameter = 18 × (18 − 1) / 2 = 153, yielding an approximate guideline of 171 based on the heuristic n + n (n−1) / 2 rule commonly applied in exploratory symptom network studies,18, 19, 20 rather than a formal power analysis forGaussian graphical model (GGM), mixed graphical model (MGM), Bayesian network, or Ising simulation models (see Limitations). To account for an anticipated 20% invalid response rate, the target sample size was increased to 214.
This cohort partially overlaps with a previously published study examining symptom clusters and core symptoms in the same patient population using exploratory factor analysis and undirected network centrality metrics.21 The present analysis differs in both research aims and methodology: it employs a Bayesian network to characterize directed probabilistic associations among symptoms and Ising-model simulations to project the effects of symptom-targeted interventions, neither of which was addressed in the earlier study.
Measures
Demographic and clinical characteristics
Demographic and clinical information was collected using a self-developed general information questionnaire. Demographic variables included sex, age, marital status, educational attainment, per capita monthly household income, and type of medical insurance. Clinical characteristics, including tumor stage, disease duration, surgical history, treatment modality, and time since diagnosis, were extracted from medical records by trained researchers.
Esophageal cancer-related symptoms
Symptoms were assessed using the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 (EORTC QLQ-C30)22 and its esophageal-specific module (QLQ-OES18).23 The EORTC QLQ-C3022,24 consists of 30 items comprising a global health status scale, five functional scales, three multi-item symptom scales (fatigue, pain, and nausea and vomiting), five single-item symptom measures (dyspnea, appetite loss, sleep disturbance, constipation, and diarrhea), and one item assessing financial difficulties. The QLQ-OES18 module includes four symptom scales (dysphagia, eating, reflux, and esophageal pain) and six single items (trouble swallowing saliva, choking, dry mouth, taste disturbance, coughing, and speech difficulties). Except for the global health status scale, which uses a seven-point response format, all items are rated on a four-point Likert scale ranging from 1 (not at all) to 4 (very much). Raw scores (RS) are calculated as the mean of the constituent items (RS = ∑ Qn / n) and then linearly transformed into standardized scores (SS) ranging from 0 to 100. Transformation formulas are as follows: for functional scales, SS = [1 - (RS - 1) / R] × 100; for symptom scales and global health status, SS = [(RS - 1) / R] × 100, where R denotes the range of the scores (R = 6 for global health status; R = 3 for all other scales). Higher scores on the symptom scales indicated greater symptom severity. To construct the symptom network, only the eight symptom measures from the QLQ-C30 and all 10 symptom measures from the QLQ-OES18 were included as nodes, resulting in a total of 18 symptom nodes for the analysis. The Chinese versions of the EORTC QLQ-C30 and QLQ-OES18 have both demonstrated good reliability and validity in prior psychometric validation studies, with Cronbach’s α exceeding 0.70 for the respective multi-item subscales.22,23
Data collection
Data were collected through face-to-face interviews conducted by trained researchers. Participants received a standardized explanation of the study purpose and instructions for questionnaire completion, and written informed consent was obtained prior to distribution. Completed questionnaires were retrieved immediately and reviewed for completeness; any items found to be unanswered were clarified and completed by participants on the spot. All data were independently double-entered and cross-checked by two researchers to ensure accuracy. A total of 280 questionnaires were distributed, of which 25 (8.9%) were subsequently excluded during data cleaning: 15 due to patterned responses (e.g., identical or straight-line answers across items) and 10 due to logically inconsistent answers. Exclusion decisions were independently adjudicated by two trained researchers, with disagreements resolved by a third senior researcher. Data cleaning was completed prior to any outcome analysis, resulting in a final analytical sample of 255 participants.
Statistical analysis
All statistical analyses were conducted using R software (version 4.5.3) using the packages “qgraph”, “bootnet”, “mgm" (v1.2.14), “bnlearn" (v4.9), and “IsingFit".
We first estimated an undirected GGM to characterize interrelationships among the 18 symptoms. Graphical LASSO combined with EBIC (γ = 0.5) was applied to obtain a sparse and interpretable network.19 Node strength was calculated as the primary centrality metric. Betweenness, closeness, and expected influence were additionally computed. Edge-weight accuracy was assessed via 95% bootstrap confidence intervals, and the stability of strength centrality was evaluated using the correlation stability coefficient (CS-coefficient), with values over 0.25 considered acceptable and over 0.50 indicating strong stability.20 To assess potential node redundancy, a goldbricker analysis was conducted using the networktools package (Hittner 2003 method, threshold = 0.25, corMin = 0.50).25 Results indicated no redundancy among the 18 symptom nodes, supporting the retention of all nodes in the final network. Node predictability was quantified using the R2 metric derived from the MGM framework, representing the proportion of variance in each node explained by its neighboring nodes. In the network visualization, the colored ring surrounding each node represents this predictability value, with the shaded proportion corresponding to the percentage of variance explained. A correlation heatmap with hierarchical clustering (complete linkage, Euclidean distance) was additionally generated based on the same Spearman correlation matrix used for GGM estimation, to provide a complementary visualization of pairwise symptom associations.
We then constructed a directed acyclic graph using a Bayesian network26 with the Hill-Climbing algorithm and Akaike Information Criterion (AIC) scoring criterion to model potential directional pathways among symptoms.27 Continuous symptom scores were trichotomized into three levels (Low: ≤ 34; Medium: > 34–≤ 56; High: > 56) prior to analysis. Network robustness and edge-direction stability were assessed over 1000 bootstrap resamples, and only edges with bootstrap strength ≥ 0.50 were retained in the final averaged network. Edge directions with values close to 0.50 were interpreted as exploratory findings.
Finally, to simulate symptom-targeted interventions, standardized symptom scores were dichotomized (inactive: ≤ 50; active: > 50) using the scale midpoint as an analytic cut-off, and an Ising model was fitted using the NodeIdentifyR algorithm.17 Two intervention scenarios were modeled using the general simulation framework of Lunansky et al.,17 by shifting each node’s threshold parameter by ± 5 units of the standard deviation of the estimated threshold parameters across all 18 nodes (a single perturbation magnitude applied uniformly across nodes; alleviating: threshold decreased; aggravating: threshold increased).17 Changes in global network activation relative to baseline were compared across nodes to identify candidate intervention targets.28,29
Full details of the analytical workflow, including model specifications, package versions, and parameter settings, are provided in the Supplementary Statistical Analysis Report.
Results
Demographic and clinical characteristics
A total of 255 participants were included in the analysis. The majority were male (n = 223, 87.5%) and married (n = 251, 98.4%), and most had a secondary school education or lower (n = 196, 76.9%). The median duration since diagnosis was 6.00 months with an interquartile range (IQR) of 2.00–20.00, ranging from 1 to 234 months. Detailed participant characteristics are presented in Table 1.
Table 1.
Participant demographic and clinical characteristics (N = 255).
| Variables | n (%) / Value |
|---|---|
| Age (years) | |
| Mean ± SD | 63.83 ± 7.48 |
| Range | 42.00–87.00 |
| Sex | |
| Male | 223 (87.5) |
| Female | 32 (12.5) |
| BMI (kg/m2) | |
| Mean ± SD | 19.73 ± 2.78 |
| Range | 13.06–27.04 |
| < 18.5 | 94 (36.9) |
| 18.5–24.9 | 150 (58.8) |
| ≥ 25.0 | 11 (4.3) |
| Educational level | |
| Middle school (junior high) or below | 196 (76.9) |
| High school / technical secondary school | 40 (15.7) |
| University or above | 19 (7.5) |
| Marital status | |
| Married | 251 (98.4) |
| Single | 1 (0.4) |
| Divorced/widowed | 3 (1.2) |
| Medical insurance | |
| Medical insurance for non-working urban residents | 159 (62.4) |
| Medical insurance for urban workers | 89 (34.9) |
| No insurance | 7 (2.7) |
| Residence | |
| Rural area | 165 (64.7) |
| Urban area | 90 (35.3) |
| Employment status | |
| Employed | 71 (27.8) |
| Not employed | 122 (47.8) |
| Retired | 62 (24.3) |
| Per capita monthly household income (CNY) | |
| < 3000 | 138 (54.1) |
| 3000-8000 | 96 (37.6) |
| > 8000 | 21 (8.2) |
| Stage of disease | |
| I | 8 (3.1) |
| II | 77 (30.2) |
| III | 94 (36.9) |
| IV | 76 (29.8) |
| Surgery history | |
| Yes | 177 (69.4) |
| No | 78 (30.6) |
| Treatment modality | |
| Chemotherapy alone | 197 (77.3) |
| Chemotherapy and radiotherapy | 27 (10.6) |
| Radiotherapy alone | 30 (11.8) |
| Chemotherapy and esophageal stent placement | 1 (0.4) |
| Duration since cancer diagnosis (months) | |
| Median (IQR) | 6.00 (2.00, 20.00) |
| Range | 1.00–234.00 |
SD, standard deviation; BMI, body mass index; IQR, interquartile range.
Network analysis
Symptom associations were characterized using an undirected GGM network (Fig. 1a). The network structure was relatively sparse: out of 153 possible edges, 11 (7.2%) were estimated as nonzero (Fig. 1a). Trouble with eating (A10) was the most prevalent symptom (99.6%), with a median score of 25.00 (IQR: 16.67–41.67) (Table 2). Node predictability, quantified using the R2 metric derived from the MGM, is also reported in Table 2 and is visually represented by the colored ring surrounding each node in Fig. 1, with the shaded proportion corresponding to the percentage of variance explained. Fatigue (A6) showed the highest predictability (66.6%), followed by pain (A8, 63.0%) and appetite loss (A3, 43.8%). Within the overall network, fatigue (A6) was identified as the most central symptom, with the highest strength centrality (Str = 0.7465), followed by pain (A8, Str = 0.5445) and appetite loss (A3, Str = 0.3802), consistent with the ranking of node predictability. Bootstrap 95% CIs for edge weights indicated high accuracy, and a case-dropping bootstrap yielded a CS-coefficient of 0.671 for strength, confirming acceptable network stability (Supplementary Fig. S1). Node redundancy check using the goldbricker function confirmed that all 18 symptom nodes were statistically distinct, supporting the retention of all nodes in the final network. Other centrality metrics (Betweenness, Closeness, and Expected Influence) are provided in Supplementary Table S1 for reference.
Fig. 1.
Symptom network structure and centrality in patients with esophageal cancer. (a) EBICglasso-regularized Gaussian graphical model of symptom associations, with node predictability (MGM-derived R2) integrated as colored rings; (b) correlation heatmap and hierarchical clustering among symptoms, based on the same Spearman correlation matrix used for network estimation in (a); (c) z-standardized strength centrality of symptoms.
Table 2.
Symptom prevalence, severity, and predictability (N = 255).
| Symptom | Median (IQR) | n (%) | MGM-derived R2 | |
|---|---|---|---|---|
| A1 | Dyspnea | 0.00 (0.00, 33.33) | 88 (34.5) | 28.70% |
| A2 | Sleep disturbance | 33.33 (0.00, 33.33) | 144 (56.5) | 6.90% |
| A3 | Appetite loss | 33.33 (0.00, 66.67) | 173 (67.8) | 43.80% |
| A4 | Constipation | 0.00 (0.00, 33.33) | 90 (35.3) | 17.90% |
| A5 | Diarrhea | 0.00 (0.00, 0.00) | 63 (24.7) | 8.90% |
| A6 | Fatigue | 33.33 (22.22, 55.56) | 236 (92.5) | 66.60% |
| A7 | Nausea and vomiting | 0.00 (0.00, 16.67) | 95 (37.3) | 39.90% |
| A8 | Pain | 33.33 (16.67, 33.33) | 217 (85.1) | 63.00% |
| A9 | Dysphagia | 33.33 (22.22, 55.56) | 213 (83.5) | 37.20% |
| A10 | Trouble with eating | 25.00 (16.67, 41.67) | 254 (99.6) | 24.20% |
| A11 | Reflux | 16.67 (0.00, 16.67) | 153 (60.0) | 18.00% |
| A12 | Esophageal pain | 0.00 (0.00, 11.11) | 84 (32.9) | 29.30% |
| A13 | Trouble swallowing saliva | 0.00 (0.00, 0.00) | 45 (17.6) | 12.50% |
| A14 | Choking when swallowing | 33.33 (0.00, 33.33) | 139 (54.5) | 26.20% |
| A15 | Dry mouth | 33.33 (0.00, 33.33) | 128 (50.2) | 10.40% |
| A16 | Trouble with taste | 33.33 (0.00, 33.33) | 141 (55.3) | 21.10% |
| A17 | Trouble with coughing | 0.00 (0.00, 0.00) | 38 (14.9) | 9.80% |
| A18 | Trouble with talking | 0.00 (0.00, 0.00) | 44 (17.3) | 12.90% |
EORTC, European Organization for Research and Treatment of Cancer; IQR, interquartile range; MGM, mixed graphical model. “Symptom presence” was defined as an EORTC-transformed item/scale score > 0. Predictability was quantified as MGM-derived R2.
Sensitivity analysis: subgroup network comparison by tumor stage
A subgroup sensitivity analysis was conducted to examine whether network findings were consistent across clinical subgroups. Patients were stratified into Early/Intermediate Stage (Stages I–II, n = 85) and Advanced Stage (Stages III–IV, n = 170) groups based on tumor stage recorded at diagnosis. Partial correlation networks were constructed for each subgroup using a fixed layout to facilitate topological comparison. A Network Comparison Test (NCT) based on 1000 permutations revealed no statistically detectable between-group differences in either Global Strength (S = 0.491, P = 0.784) or Network Structure Invariance (M = 0.177, P = 0.639); however, given the limited sample size in the Early/Intermediate Stage subgroup (n = 85), this subgroup analysis may have had limited power to detect true differences. Visual inspection of the subgroup networks was also consistent with these results, with central symptom nodes maintaining similar centrality and connectivity patterns across both disease stages (Supplementary Fig. S2).
Bayesian network model
Fig. 2 presents the Bayesian network structure among 18 symptoms, based on bootstrap-averaged edges (1000 resamples, threshold ≥ 0.50); edge directions and bootstrap strength values are reported in Supplementary Table S2. Directed edges represent potential conditional dependencies among symptoms. Fatigue (A6) had the greatest number of outgoing edges, showing directed probabilistic associations with pain (A8), appetite loss (A3), dysphagia (A9), nausea and vomiting (A7), and trouble with eating (A10), suggesting a hub-like position within the network; pain and appetite loss each showed additional directed probabilistic associations with other symptom-related nodes (pain: esophageal pain, dry mouth, and constipation; appetite loss: trouble with taste and sleep disturbance). Bootstrap validation supported the robustness of the core network structure, with 16 edges retained at a threshold ≥ 0.50, two of which showed relatively weak directional evidence (direction < 0.55) and should be interpreted as exploratory findings. Four symptoms — diarrhea (A5), reflux (A11), trouble swallowing saliva (A13), and trouble with talking (A18) — showed no bootstrap-supported edges and appeared as isolated nodes with no retained connections. Full details of edge directions and bootstrap strength values are provided in Supplementary Table S2.
Fig. 2.
Bayesian network structure of symptoms in patients with esophageal cancer (N = 255), based on bootstrap-averaged edges (1000 resamples; strength threshold ≥ 0.50). Edge directions shown correspond to the majority-supported direction among resamples in which the edge was present. Four symptoms — diarrhea (A5), reflux (A11), trouble swallowing saliva (A13), and trouble with talking (A18) — showed no bootstrap-supported edges at the specified threshold and are displayed as isolated nodes with no retained connections; this does not imply that these symptoms are unrelated to the network but reflects the conservative bootstrap threshold applied. Full details of edge strength and direction values are provided in Supplementary Table S2. ∗A1 Dyspnea, A2 Sleep disturbance, A3 Appetite loss, A4 Constipation, A5 Diarrhea, A6 Fatigue, A7 Nausea and vomiting, A8 Pain, A9 Dysphagia, A10 Trouble with eating, A11 Reflux, A12 Esophageal pain, A13 Trouble swallowing saliva, A14 Choking when swallowing, A15 Dry mouth, A16 Trouble with taste, A17 Trouble with coughing, A18 Trouble with talking.
Computer-simulated interventions
Results of the computer-simulated interventions are presented in Fig. 3. Reductions in appetite loss (which showed directed probabilistic associations with sleep disturbance and taste disturbance in the Bayesian network) and fatigue (the node with the highest strength centrality and exhibiting directed probabilistic associations with several other symptoms) were associated with decreases in global network activation, with sum scores declining from 10.36 to 7.75 and 7.78, respectively (reductions of 2.61 and 2.58 points, respectively), indicating that these central symptoms represent potential candidate targets for intervention. In contrast, pain markedly increased overall severity, with the total activation score rising from 10.36 to 17.68 (an increase of 7.33 points). Choking when swallowing (A14) showed the second-largest projected increase in the sum score (+6.65 points). These findings suggest that pain may represent a potential priority symptom for preventive management.
Fig. 3.
Projected effects of computer-simulated interventions on the esophageal cancer symptom network. (a) Alleviating intervention; (b) Aggravating intervention. Y-axis: Sum Score (0–100). Baseline = 10.36 (horizontal dashed line); perturbation = 5 SDs applied to one node’s threshold at a time. Nodes ordered by post-intervention Sum Score.
Discussion
Main findings
This study integrated static symptom network analysis with computer-simulated interventions to explore potential candidate management targets in patients with esophageal cancer. Computer-simulated results showed that perturbation of appetite loss (A3) was associated with the largest projected decrease in overall symptom burden under the simulated alleviation scenario (from 10.36 to 7.75, a decrease of 2.61 points), while perturbation of fatigue (A6) was associated with a comparable decrease (from 10.36 to 7.78, a decrease of 2.58 points). Under the simulated aggravation scenario, pain (A8) was associated with the largest projected increase in overall severity (from 10.36 to 17.68, an increase of 7.33 points); notably, pain also ranked second in static strength centrality (Str = 0.5445), indicating general consistency between its structural prominence and its simulated impact. However, a divergence between static network metrics and the simulation results was observed for choking when swallowing (A14), which was associated with the second-largest projected increase (an increase of 6.65 points) despite showing only low strength centrality in the static network (Str = 0.0411, one of the lowest among the 18 nodes). This pattern is broadly consistent with observations by Lunansky et al.,17 who suggested that candidate intervention targets identified through simulation may be associated with, but are not necessarily identical to, central nodes identified through static centrality metrics. These findings suggest that static network metrics alone may not fully capture the complexity of symptom interdependencies, and that computer simulation may offer a complementary perspective for exploring potential intervention targets.30,31
In the Bayesian network, fatigue had the greatest number of outgoing edges, showing directed probabilistic associations with pain (A8), appetite loss (A3), dysphagia (A9), nausea and vomiting (A7), and trouble with eating (A10); appetite loss, in turn, showed further associations with sleep disturbance (A2) and taste disturbance (A16). This is consistent with longitudinal evidence from Sun et al.,32 who found that fatigue progressively became the most influential symptom during the mid-to-late stages of chemotherapy in patients with esophageal cancer, and previous research suggesting that fatigue-related symptoms may share underlying pathophysiological mechanisms, including tumor- and treatment-related immune activation and the release of pro-inflammatory cytokines such as interleukin (IL)-1, IL-6, and tumor necrosis factor (TNF)-α.33 The association between appetite loss and sleep disturbance is broadly consistent with a symptom-network study in patients with colorectal cancer receiving chemotherapy, which reported that fatigue demonstrated strong associations with both sleep disturbance and anorexia,34 and the association between appetite loss and taste disturbance is consistent with prior evidence that taste alterations in patients with cancer are associated with appetite, fatigue, and quality of life.35 Appetite loss has also been linked to a broader gastrointestinal symptom cluster — comprising nausea, bloating, diarrhea, dysphagia, and poor appetite — in patients with esophageal cancer,36 is common within the first six months following esophageal cancer surgery,37 and has been associated with progressive muscle wasting, immune dysfunction, and increased mortality risk.38,39 These findings suggest that fatigue may warrant longer-term monitoring in clinical practice, and that early nutritional and pharmacological strategies targeting appetite loss may warrant consideration in this population.
Computer-simulated results identified appetite loss and fatigue as the two candidate targets associated with the largest projected reductions in overall symptom burden under the alleviation scenario, consistent with their prominent positions in the Bayesian network. In contrast, pain (A8) was identified as the candidate target associated with the largest projected increase under the aggravation scenario, with choking when swallowing (A14) associated with the second-largest projected increase. Pain in patients with esophageal cancer may arise from multiple sources, including direct tumor invasion of surrounding tissues and nerves and treatment-related tissue injury. Emerging evidence suggests that peripheral sensory neurons, particularly nociceptors, can be activated within the tumor microenvironment and release neuropeptides that engage in bidirectional crosstalk with immune cells, a mechanism distinct from the cytokine-mediated pathways implicated in cancer-related fatigue.40 Tumor-related nerve damage through perineural invasion has also been described as a contributing mechanism to neuropathic pain in cancer. In the Bayesian network, pain also showed directed probabilistic associations with esophageal pain, dry mouth, and constipation, suggesting a broader symptom footprint consistent with its prominent role in the simulated aggravation scenario. Choking when swallowing, the node with the second-largest simulated increase despite its markedly lower static centrality (Str = 0.0411), further illustrates this divergence between static and simulation-based findings noted above;17 it is a common postoperative complication of esophageal cancer surgery,41 with proposed mechanisms including reduced hyoid displacement, diminished vocal fold mobility, and sarcopenia-related weakness of swallowing-related muscles;42, 43, 44 proactive swallowing function screening and rehabilitation43, 44, 45, 46, 47, 48, 49 may therefore also warrant consideration, particularly among patients who have undergone surgery.
Implications for nursing practice and research
Overall, fatigue and pain were identified as central symptoms within the symptom network of patients with esophageal cancer and may represent candidate targets for clinical assessment and intervention, with their role as intervention targets warranting confirmation in future longitudinal and interventional studies. Appetite loss was identified as the candidate target associated with the largest projected decrease in overall symptom burden under the simulated alleviation scenario, consistent with its prominent structural position within the network. Nurses and public health practitioners may consider these three central symptoms — fatigue, appetite loss, and pain — alongside choking when swallowing, a secondary simulation-identified target despite its comparatively low static centrality, as starting points for developing symptom-targeted nursing interventions, although whether such interventions would reduce overall symptom burden remains to be established.
Limitations
Several limitations should be acknowledged. First, the cross-sectional design precludes definitive causal inference; relationships identified through network modeling and Bayesian analysis should be interpreted as probabilistic dependencies requiring confirmation through future longitudinal research. Second, participants were recruited from two tertiary hospitals in Jiangsu Province via convenience sampling; patients from the same geographical region may share similar clinical and sociodemographic characteristics, limiting generalizability, and no formal center-level comparisons were conducted between sites. Third, although a subgroup sensitivity analysis by tumor stage showed no statistically detectable differences, this should not be interpreted as confirming network stability, particularly given the limited sample size in the Early/Intermediate Stage subgroup (n = 85); the sample’s heterogeneity in treatment modality (surgery, chemotherapy, radiotherapy) and disease duration since diagnosis may have further shaped symptom profiles and network configurations, and the extended recruitment window (2020–2023) may have introduced additional temporal heterogeneity from evolving treatment protocols. Fourth, the sample size estimation used a heuristic parameter-count approach rather than a formal power analysis for GGM, MGM, Bayesian network, or Ising simulation models; future studies may consider simulation-based power analysis. Finally, the computer-simulated intervention findings are exploratory and hypothesis-generating; dichotomizing and trichotomizing symptom scores at fixed analytic thresholds also entails some loss of information—for instance, the medium category was structurally unreachable for most single-item symptoms in the Bayesian network—and results may be threshold-dependent, warranting confirmation through future longitudinal and interventional studies.
Conclusions
This study combined network analysis with computer-simulated interventions to characterize the symptom network among patients with esophageal cancer. Fatigue (A6) had the highest strength centrality in the undirected network. In the computer simulations, perturbations involving appetite loss and fatigue were associated with the largest projected decreases in the global activation metric, whereas activation of pain (A8) was associated with the largest projected increase. Bayesian network analysis further suggested that fatigue and appetite loss showed directed probabilistic associations with several other symptoms. These model-based findings are exploratory in nature and may inform the design of future studies evaluating symptom-targeted interventions in patients with esophageal cancer. The clinical efficacy of these potential targets requires confirmation through prospective longitudinal and interventional research before informing clinical practice recommendations.
CRediT authorship contribution statement
Xiaomeng Wen: Data curation, Formal analysis, Methodology, Writing - original draft. Yan Liu: Investigation, Software. Yujia Huang: Investigation, Software. Yuan Lin: Investigation, Software. Yue Qiao: Data curation. Xiaoyun Han: Supervision, Writing - review & editing. Xianwei Guo: Formal analysis, Methodology, Software. Danfeng Gu: Investigation, Data curation, Resources. All authors have read and approved the final manuscript.
Ethics statement
This study was approved by the Institutional Review Board of The First People’s Hospital of Changzhou (Approval No. 2020-CL003-01) and the Affiliated Hospital of Jiangnan University (Approval No. LS2020010), and was conducted in accordance with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. All participants provided written informed consent.
Data availability
The data that support the findings of this study are available from the corresponding author, Xiaoyun Han, upon reasonable request.
Declaration of generative AI and AI-assisted technologies in the writing process
During the preparation of this manuscript, the authors used DeepSeek for language editing and grammar checking. The authors reviewed all changes and take full responsibility for the final content.
Funding
This study was supported by the Science and Technology Project of Changzhou Municipal Health Commission (Grant No. ZD202407). The funders had no role in the study design, data collection, analysis, interpretation, writing of the report, or decision to submit the article for publication.
Declaration of competing interest
The authors declare no conflict of interest.
Acknowledgments
The authors appreciate all the participants in this study.
Footnotes
Supplementary data to this article can be found online at https://doi.org/10.1016/j.apjon.2026.101019.
Appendix. Supplementary materials
The following is the Supplementary data to this article.
References
- 1.Bray F., Laversanne M., Sung H., et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74(3):229–263. doi: 10.3322/caac.21834. [DOI] [PubMed] [Google Scholar]
- 2.Han B., Zheng R., Zeng H., et al. Cancer incidence and mortality in China, 2022. J Natl Cancer Cent. 2024;4(1):47–53. doi: 10.1016/j.jncc.2024.01.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Yang S., Lin S., Li N., et al. Burden, trends, and risk factors of esophageal cancer in China from 1990 to 2017: an up-to-date overview and comparison with those in Japan and South Korea. J Hematol Oncol. 2020;13(1):146. doi: 10.1186/s13045-020-00981-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Hsu P.K., Chien L.I., Huang C.S., et al. Treatment patterns and outcomes in patients with esophageal cancer: an analysis of a multidisciplinary tumor board database. Ann Surg Oncol. 2022;29(1):572–585. doi: 10.1245/s10434-021-10568-z. [DOI] [PubMed] [Google Scholar]
- 5.Ukegjini K., Vetter D., Fehr R., et al. Functional syndromes and symptom-orientated aftercare after esophagectomy. Langenbecks Arch Surg. 2021;406(7):2249–2261. doi: 10.1007/s00423-021-02203-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Fu L., Feng X., Jin Y., et al. Symptom clusters and quality of life in gastric cancer patients receiving chemotherapy. J Pain Symptom Manag. 2022;63(2):230–243. doi: 10.1016/j.jpainsymman.2021.09.003. [DOI] [PubMed] [Google Scholar]
- 7.Wang Y., Xie Z., Liu Y., et al. Symptom clusters and impact on quality of life in esophageal cancer patients. Health Qual Life Outcomes. 2022;20(1):168. doi: 10.1186/s12955-022-02084-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Choi E., Kim S.H., Lee Y.W., et al. Supportive care needs and health-related quality of life of esophageal cancer survivors. Asia Pac J Oncol Nurs. 2021;8(2):164–171. doi: 10.4103/apjon.apjon_60_20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Wei L., Lv F., Luo C., et al. Study on sentinel symptoms and influencing factors of postoperative chemotherapy in patients with gastric cancer. Eur J Oncol Nurs. 2023;64 doi: 10.1016/j.ejon.2023.102318. [DOI] [PubMed] [Google Scholar]
- 10.Fang J., Wong C.L., Liu C.Q., et al. Identifying central symptom clusters and correlates in children with acute leukemia undergoing chemotherapy: a network analysis. Front Oncol. 2023;13 doi: 10.3389/fonc.2023.1236129. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Robinaugh D.J., Hoekstra R.H.A., Toner E.R., et al. The network approach to psychopathology: a review of the literature 2008-2018 and an agenda for future research. Psychol Med. 2020;50(3):353–366. doi: 10.1017/S0033291719003404. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Triolo F., Belvederi Murri M., Calderon-Larranaga A., et al. Bridging late-life depression and chronic somatic diseases: a network analysis. Transl Psychiatry. 2021;11(1):557. doi: 10.1038/s41398-021-01686-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.McNally R.J., Mair P., Mugno B.L., et al. Co-morbid obsessive-compulsive disorder and depression: a Bayesian network approach. Psychol Med. 2017;47(7):1204–1214. doi: 10.1017/S0033291716003287. [DOI] [PubMed] [Google Scholar]
- 14.Tennant P.W.G., Murray E.J., Arnold K.F., et al. Use of directed acyclic graphs (DAGs) to identify confounders in applied health research: review and recommendations. Int J Epidemiol. 2021;50(2):620–632. doi: 10.1093/ije/dyaa213. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Borsboom D., Robinaugh D.J., The Psychosystems Group, et al. Robustness and replicability of psychopathology networks. World Psychiatry. 2018;17(2):143–144. doi: 10.1002/wps.20515. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Robinaugh D.J., Millner A.J., McNally R.J. Identifying highly influential nodes in the complicated grief network. J Abnorm Psychol. 2016;125(6):747–757. doi: 10.1037/abn0000181. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Lunansky G., Naberman J., van Borkulo C.D., et al. Intervening on psychopathology networks: evaluating intervention targets through simulations. Methods. 2022;204:29–37. doi: 10.1016/j.ymeth.2021.11.006. [DOI] [PubMed] [Google Scholar]
- 18.Papachristou N., Barnaghi P., Cooper B., et al. Network analysis of the multidimensional symptom experience of oncology. Sci Rep. 2019;9(1):2258. doi: 10.1038/s41598-018-36973-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Epskamp S., Fried E.I. A tutorial on regularized partial correlation networks. Psychol Methods. 2018;23(4):617–634. doi: 10.1037/met0000167. [DOI] [PubMed] [Google Scholar]
- 20.Epskamp S., Borsboom D., Fried E.I. Estimating psychological networks and their accuracy: a tutorial paper. Behav Res Methods. 2018;50(1):195–212. doi: 10.3758/s13428-017-0862-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Wen X., Sun C., Han X., Wang X., et al. Network analysis of core symptoms and symptom clusters in patients with esophageal cancer. J Nurs Sci. 2025;40(5):20–25. doi: 10.3870/j.issn.1001-4152.2025.05.020. [in Chinese] [DOI] [Google Scholar]
- 22.Wan C., Chen M., Zhang C., et al. Application of EORTC QLQ-C30 (V3.0) in measuring the quality of life of cancer patients. J Pract Oncol. 2005;20(4):353–355. doi: 10.3969/j.issn.1001-1692.2005.04.028. [in Chinese] [DOI] [Google Scholar]
- 23.Dai Z., Lang W., Yang H., et al. Validation of EORTC QLQ-OES18 for Chinese patients with esophageal cancer. Dis Esophagus. 2017;30(10):1–7. doi: 10.1093/dote/dox046. [DOI] [PubMed] [Google Scholar]
- 24.Guo M., Yin X.M., Wang C., et al. Symptom clusters in esophageal cancer patients after surgery. Chin J Nurs. 2019;54(8):1189–1193. doi: 10.3761/j.issn.0254-1769.2019.08.014. [in Chinese] [DOI] [Google Scholar]
- 25.Levinson C.A., Brosof L.C., Vanzhula I., et al. Social anxiety and eating disorder comorbidity and underlying vulnerabilities: using network analysis to conceptualize comorbidity. Int J Eat Disord. 2018;51(7):693–709. doi: 10.1002/eat.22890. [DOI] [PubMed] [Google Scholar]
- 26.Briganti G., Scutari M., McNally R.J. A tutorial on bayesian networks for psychopathology researchers. Psychol Methods. 2023;28(4):947–961. doi: 10.1037/met0000479. [DOI] [PubMed] [Google Scholar]
- 27.Scutari M. Learning Bayesian networks with the bnlearn R package. J Stat Softw. 2010;35(3):1–22. doi: 10.18637/jss.v035.i03. [DOI] [Google Scholar]
- 28.Li H., Xiong Y., Zhang Q., et al. The interplay between sleep and cancer-related fatigue in breast cancer: a casual and computer-simulated network analysis. Asia Pac J Oncol Nurs. 2025;12 doi: 10.1016/j.apjon.2025.100692. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Chen F., Deng Y., Li S., et al. Fear of cancer recurrence and perceived pain in patients with breast cancer: a network analysis approach. Asia Pac J Oncol Nurs. 2025;12 doi: 10.1016/j.apjon.2025.100763. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Li P., Huang N., Yang X., et al. A simulation-based network analysis of intervention targets for adolescent depressive and anxiety symptoms. Asian J Psychiatry. 2024;99 doi: 10.1016/j.ajp.2024.104152. [DOI] [PubMed] [Google Scholar]
- 31.Liang M., Pan Y., Cai J., et al. Navigating specific targets of breast cancer symptoms: an innovative computer-simulated intervention analysis. Eur J Oncol Nurs. 2025;74 doi: 10.1016/j.ejon.2024.102708. [DOI] [PubMed] [Google Scholar]
- 32.Sun C., Sun D., Yang H., et al. Exploring the dynamics of symptom networks in esophageal cancer patients during chemotherapy: a cross-lagged panel network analysis. Eur J Oncol Nurs. 2026;81 doi: 10.1016/j.ejon.2026.103138. [DOI] [PubMed] [Google Scholar]
- 33.Bower J.E. Cancer-related fatigue—mechanisms, risk factors, and treatments. Nat Rev Clin Oncol. 2014;11(10):597–609. doi: 10.1038/nrclinonc.2014.127. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Xie R., Gu X.X., Wang Y. Symptom network analysis in patients undergoing chemotherapy for colorectal cancer: a cross-sectional study. Cancer Manag Res. 2025;17:3073–3085. doi: 10.2147/CMAR.S558889. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Donald M. A matter of taste: alteration in patients with cancer. Br J Nurs. 2022;31(13):S10–S14. doi: 10.12968/bjon.2022.31.13.S10. [DOI] [PubMed] [Google Scholar]
- 36.Zhu C., Wang N., Xu F., et al. Exploration of symptom cluster patterns and their trajectory in esophageal cancer surgery patients. Eur J Oncol Nurs. 2025;74 doi: 10.1016/j.ejon.2025.102801. [DOI] [PubMed] [Google Scholar]
- 37.Meng Y., Gao R., Yang H., et al. Health-related quality of life and related factors among esophageal cancer survivors after esophagectomy in the 6-month postoperative period: a multicenter cross-sectional study in north China. Asia Pac J Oncol Nurs. 2025;12(1) doi: 10.1016/j.apjon.2025.100655. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Martin L., Lagergren P. Risk factors for weight loss among patients surviving 5 years after esophageal cancer surgery. Ann Surg Oncol. 2015;22(2):610–616. doi: 10.1245/s10434-014-3973-2. [DOI] [PubMed] [Google Scholar]
- 39.Wikman A., Johar A., Lagergren P. Presence of symptom clusters in surgically treated patients with esophageal cancer: implications for survival. Cancer. 2014;120(2):286–293. doi: 10.1002/cncr.28308. [DOI] [PubMed] [Google Scholar]
- 40.Mardelle U., Bretaud N., Daher C., et al. From pain to tumor immunity: influence of peripheral sensory neurons in cancer. Front Immunol. 2024;15 doi: 10.3389/fimmu.2024.1335387. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Hayes M., Gillman A., Elliott J.A., et al. The prevalence, nature and severity of oropharyngeal dysphagia in the acute post-operative phase following curative resection for esophageal cancer. Dis Esophagus. 2025;38(4):doaf054. doi: 10.1093/dote/doaf054. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Kaneoka A., Inokuchi H., Ueha R., et al. Longitudinal analysis of dysphagia and factors related to postoperative pneumonia in patients undergoing esophagectomy for esophageal cancer. Dysphagia. 2024;39(3):376–386. doi: 10.1007/s00455-023-10618-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Su Y., Zhang J., Li Y., et al. Risk factors for dysphagia after oesophagectomy in oesophageal cancer patients: a systematic review and meta-analysis. Ann Surg Oncol. 2025;32(9):6419–6436. doi: 10.1245/s10434-025-17709-8. [DOI] [PubMed] [Google Scholar]
- 44.Oguma J., Ozawa S., Ishiyama K., et al. Clinical significance of sarcopenic dysphagia for patients with esophageal cancer undergoing esophagectomy: a review. Ann Gastroenterol Surg. 2022;6(6):738–745. doi: 10.1002/ags3.12603. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Gillman A., Kenny C., Hayes M., et al. Nature, severity, and impact of chronic oropharyngeal dysphagia following curative resection for esophageal cancer: a cross-sectional study. Dis Esophagus. 2024;37(5):doae003. doi: 10.1093/dote/doae003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Uchida M., Hashimoto K., Mukudai S., et al. Preoperative evaluation of surgery for intractable aspiration based on the prognostic nutritional index. Nihon Jibiinkoka Gakkai Kaiho. 2014;117(12):1457–1462. doi: 10.3950/jibiinkoka.117.1457. [DOI] [PubMed] [Google Scholar]
- 47.Kurita D., Utsunomiya D., Kubo K., et al. Handgrip strength predicts early postoperative dysphagia after thoracoscopic-laparoscopic esophagectomy in male patients with esophageal cancer. Esophagus. 2022;19(4):586–595. doi: 10.1007/s10388-022-00925-6. [DOI] [PubMed] [Google Scholar]
- 48.Lee S.Y., Cheon H.J., Kim S.J., et al. Clinical predictors of aspiration after esophagectomy in esophageal cancer patients. Support Care Cancer. 2016;24:295–299. doi: 10.1007/s00520-015-2776-8. [DOI] [PubMed] [Google Scholar]
- 49.Gillman A., Hayes M., Sheaf G., et al. Exercise-based dysphagia rehabilitation for adults with oesophageal cancer: a systematic review. BMC Cancer. 2022;22(1) doi: 10.1186/s12885-021-09155-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The data that support the findings of this study are available from the corresponding author, Xiaoyun Han, upon reasonable request.