Network-derived symptom architecture and simulated leverage points during radiotherapy for oral cancer

Abstract

Objective

To characterize the symptom network among patients with oral cancer undergoing radiotherapy and explore potential symptom leverage points using network-based intervention simulation.

Methods

Patients with oral cancer undergoing radiotherapy were recruited from a tertiary hospital in China between January and December 2024 and assessed using the M.D. Anderson Symptom Inventory-Head and Neck module. The primary symptom network was estimated from the original 0–10 symptom scores using a regularized Gaussian graphical model. Strength centrality, bridge strength and network stability were assessed. For the simulation analysis, symptom scores were dichotomized as absent or present, and a binary Ising model was estimated. Symptom-specific intervention simulations were implemented through threshold perturbations and repeated 1000 times, with additional sensitivity analyses performed to assess robustness.

Results

Among the 508 participants, dry mouth, fatigue, and mouth or throat mucus were the most common symptoms. Four symptom clusters were obtained through factor analysis. Disturbed sleep showed the highest raw strength centrality (1.182), followed by mouth or throat soreness (1.084) and fatigue (1.077), whereas choking exhibited the highest bridge strength in the four-factor symptom network. No statistically significant differences in network structure or global strength were detected across subgroups defined by sex, anatomical site, or treatment modality, although these findings should be interpreted cautiously. In repeated simulations, pain showed the largest mean modeled reduction in the number of symptoms present under the alleviating intervention condition (mean Δ = −0.696), followed closely by mouth or throat soreness (−0.686). Vomiting showed the largest mean modeled increase under the aggravating intervention condition (mean Δ = +0.514) and ranked first in 75.7% of repetitions. Sensitivity analyses showed consistent effect directions and broadly similar overall ranking patterns.

Conclusions

The findings identify several symptoms with relatively high network connectivity or greater modeled influence in intervention simulations implemented through symptom-specific threshold perturbations. These simulation results represent changes in the mean modeled number of symptoms present rather than changes in symptom severity and should be interpreted as exploratory and hypothesis-generating rather than causal or therapeutic effects.

Introduction

Oral cancer represents a major subtype of head and neck malignancies, with a steadily increasing global incidence over recent decades, including among younger populations.^1,2 Radiotherapy remains an important component of treatment, either as a definitive modality in selected patients or as adjuvant therapy after surgery, contributing to locoregional control and survival outcomes.^3,4 However, radiotherapy is associated with a substantial and persistent symptom burden arising from cumulative tissue toxicity, systemic inflammatory responses, and treatment-related functional impairment. Patients with oral cancer undergoing radiotherapy commonly experience multiple concurrent symptoms, including mucosal injury, xerostomia, fatigue, sleep disturbance, gastrointestinal symptoms, and psychological distress.5, 6, 7, 8 These symptoms rarely occur in isolation; instead, they cluster and interact,⁹ forming interconnected symptom patterns that may evolve throughout treatment and collectively impair treatment tolerance, functional recovery, and quality of life.¹⁰ Despite this complexity, routine symptom management is largely guided by symptom prevalence or severity, with priority typically given to the most severe or distressing symptoms at a given time point. However, this approach assumes that symptoms exert independent effects and may therefore overlook the complex interrelationships among symptoms during radiotherapy.

A symptom cluster is defined as a group of three or more concurrent symptoms that may or may not share a common etiology or underlying mechanism.¹¹ Compared with single-symptom management, symptom cluster management may achieve greater synergistic and reinforcing effects.¹² Although this approach is useful for simplifying symptom dimensions, it remains insufficient for characterizing the complex interrelationships among multiple symptoms.^13,14 Symptom network analysis offers a data-driven solution, visualizing these interdependencies by representing symptoms as nodes and their conditional associations as edges within a network structure.¹⁵ Richard et al.¹⁶ reported that symptom network structures in patients with cancer varied according to cancer type, treatment phase, and survivorship duration, highlighting the potential importance of examining clinically relevant subgroups. Accordingly, informed by clinical expertise and the characteristics of Chinese patients, we performed exploratory subgroup network analyses based on sex, anatomical site, and treatment modality.

Although network analysis can characterize the contemporaneous structure of interrelated symptoms, it does not account for dynamic interactions among symptoms or predict the effects of interventions.^17,18 Traditional clinical intervention studies, particularly randomized controlled trials, may require substantial resources and can face practical and ethical constraints.¹⁹ Intervention simulation provides a complementary approach by modeling how hypothetical symptom-specific changes may influence the broader symptom network.^20,21 Within this framework, interventions are implemented computationally through controlled changes in symptom parameters, allowing the modeled system-level consequences of symptom-specific manipulation to be examined. This approach provides a framework for moving beyond descriptive network analysis toward identifying and comparing potential symptom leverage points based on their modeled effects.²² However, empirical applications of network analysis combined with intervention simulation in patients with oral cancer undergoing radiotherapy remain limited, and their added value beyond traditional severity-based approaches has not been systematically evaluated.

In this study, we aimed to characterize the symptom network structure among patients with oral cancer undergoing radiotherapy and to explore potential symptom leverage points using network-based intervention simulation. Specifically, we sought to (1) identify symptom clusters using exploratory factor analysis and estimate a regularized symptom network to examine central and bridge symptoms; (2) apply symptom-specific intervention simulations, operationalized through threshold perturbations, to quantify relative changes in the modeled number of symptoms present; and (3) examine whether symptom network structures differed across clinically relevant subgroups defined by sex, anatomical site, and treatment modality. By integrating symptom network analysis with intervention simulation, we aimed to generate effect size-informed, hypothesis-generating insights that may help inform future symptom-management research in this patient population.

Methods

Study design and settings

This cross-sectional study was conducted between January and December 2024 at a tertiary hospital in China. The study was reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement. Patients with oral cancer undergoing radiotherapy were recruited from the radiotherapy outpatient department using convenience sampling. This hospital is a national referral center for oral and maxillofacial-head and neck oncology, providing comprehensive surgical and radiotherapeutic care for patients with diverse tumor sites, disease stages, and treatment trajectories. As a major clinical base of the National Center for Oral Diseases, the hospital receives patients from multiple regions across China, encompassing a broad spectrum of demographic and clinical characteristics commonly encountered in radiotherapy practice for oral cancer.

Participants

The inclusion criteria for the participants were as follows: (1) age ≥ 18 years; (2) pathologically confirmed oral cancer; (3) currently receiving radiotherapy; and (4) intact cognitive capacity with provided informed consent. Participants were excluded if they met any of the following criteria: (1) history of other concurrent malignancies; (2) known history of severe mental health disorders (e.g., major depressive episode with suicidal ideation, bipolar disorder, schizophrenia) based on medical records and a brief clinical interview at enrollment; and (3) voluntary withdrawal or incomplete information.

Measurements

Sociodemographic and clinical data

The questionnaire on patient sociodemographic and clinical characteristics was self-developed by the research team based on accumulated clinical experience and review of relevant literature. Sociodemographic data included age, sex, body mass index, smoking and alcohol history, comorbidities, and health insurance coverage status. Clinical data included tumor site, indwelling tube status, flap reconstruction history, and concomitant treatment modalities. These data were collected by a trained research assistant through either direct patient interviews or review of electronic medical records.

Self-reported symptoms

We used the M.D. Anderson Symptom Inventory-Head and Neck (MDASI-H&N) to assess symptom presence and severity. The scale was developed and clinically validated by Rosenthal et al.²³ to evaluate the self-reported symptom burden specifically for patients with head and neck cancer. The MDASI-H&N comprises 13 core items, 9 head-and-neck-specific items, and 6 interference items. In the present study, the 13 core symptom items and 9 head-and-neck-specific symptom items were used for the network analyses. Each item was rated on a 0–10 numeric rating scale, with 0 indicating “not present” and 10 representing “the worst imaginable severity.” Higher scores reflected greater symptom severity. The MDASI-H&N has demonstrated good psychometric properties in Chinese populations with head and neck cancers, including those with oral cancer.^24,25 In this study, internal consistency was good, with Cronbach's α values of 0.835 for the 13 core symptom items, 0.805 for the nine head and neck-specific items, and 0.890 for the full 22-item scale.

Data collection

All symptom assessments were paper-based. Two trained researchers collected the data and used standardized explanations to clarify any misunderstandings. Participants completed questionnaires independently or with caregiver assistance within 15–20 min. Completed questionnaires were checked immediately for completeness by the researchers. Of 515 eligible patients, 508 provided valid data, yielding a 98.6% response rate.

Sample size estimation

The symptom network comprised 22 symptom nodes and 231 possible pairwise associations (22 × 21/2). For sample-size planning, the 22 node-level parameters and 231 possible pairwise associations were considered, yielding 253 parameters in total. Using this parameter count as a pragmatic planning benchmark and allowing for an anticipated 20% rate of non-response or incomplete data, the target sample size was set at no fewer than 317 participants. Ultimately, 508 participants were included in the analysis, exceeding the planned recruitment target. The empirical accuracy and stability of the estimated network were subsequently evaluated using non-parametric and case-dropping bootstrap procedures.

Data analysis

All analyses were performed using SPSS 27.0 (IBM Corp., Armonk, NY, USA) and R version 4.4.2 (R Foundation for Statistical Computing, Vienna, Austria), with a two-sided significance level of 0.05 applied where applicable. Two analytically distinct network models were used in this study: a Gaussian graphical model (GGM) based on the original continuous 0–10 symptom severity scores for the primary symptom network analysis, and a binary Ising model based on dichotomized symptom indicators for the subsequent simulation analysis. The overall study workflow is summarized in Fig. 1.

Fig. 1.

Overview of the study design and analytical workflow.

Descriptive analysis and identification of symptom clusters

Descriptive analyses were used to characterize the study population and symptom prevalence and severity. Continuous variables are reported as mean ± standard deviation (SD) for normally distributed data or as median and interquartile range (IQR) for non-normally distributed data. Categorical variables are reported as frequencies and percentages.

Exploratory factor analysis was performed to identify symptom clusters. Sampling adequacy and factorability were assessed using the Kaiser-Meyer-Olkin (KMO) measure and Bartlett's test of sphericity. Factors were extracted using principal axis factoring with oblique (Promax) rotation, as we assumed that underlying symptom factors might be correlated. The number of factors to retain was determined by multiple criteria: (1) parallel analysis; (2) visual inspection of the scree plot; (3) clinical interpretability; and (4) eigenvalues ≥ 1 as reference. Because these criteria did not necessarily converge, alternative two-, three-, and four-factor solutions were additionally examined and compared according to explained variance, factor loading patterns, and clinical interpretability. When statistical criteria diverged, priority was given to a solution showing a clearer simple structure and greater clinical interpretability. Factor loadings ≥ 0.30 were retained; items loading on multiple factors were assigned to the factor with the largest absolute loading.²⁶ Each retained factor was required to contain at least two symptoms. The internal consistency of each symptom cluster was assessed using Cronbach's alpha coefficient.

Estimation of the contemporaneous symptom network

We used the estimateNetwork function from the bootnet package to construct the primary Gaussian graphical symptom networks for the total sample and subgroups. These networks were estimated from the original continuous 0–10 symptom severity scores of the MDASI-H&N using regularized partial correlations via the graphical lasso (GLASSO) with the extended Bayesian information criterion (EBIC; tuning parameter γ = 0.5).¹³ Pearson correlations were used as the input correlation method (corMethod = “cor”). No missing data were present in the analyzed dataset; therefore, no missing-data imputation procedure was required. We applied the Fruchterman-Reingold algorithm with a spring-electrical layout, which positions nodes with stronger edge weights closer together in the network.²⁷

Centrality indices were calculated using the qgraph package.²⁸ Strength centrality, defined as the sum of absolute edge weights directly connected to a node, represents network connectivity. A symptom with higher strength centrality was more likely to co-occur with other symptoms.²⁹ Bridge strength centrality was calculated to identify bridge symptoms, which are broadly defined as symptoms that may link distinct symptom clusters, facilitating the spread or amplification of symptom burden across domains.³⁰ Strength and bridge strength were reported as raw centrality indices rather than standardized z scores.

We assessed network accuracy and stability of the Gaussian graphical network using the bootnet package. Accuracy was evaluated via non-parametric bootstrapping (nBoots = 1000) to compute 95% confidence intervals for edge weights; narrower intervals indicated more precise estimates.³¹ Stability was assessed using case-dropping bootstrap to derive correlation stability (CS) coefficients for centrality, with CS > 0.25 considered acceptable and CS > 0.50 ideal.³²

To explore potential differences in symptom networks across clinically relevant subgroups, separate networks were estimated for sex (male vs female), anatomical site (oral cavity proper vs oral vestibule), and treatment modality (radiotherapy alone vs radiotherapy plus other treatments) using the same EBIC-glasso procedure and analytical parameters as the main network. Network comparison tests (NCT) were conducted using the NetworkComparisonTest package with 1000 permutations.³³ Network structure invariance and global strength invariance were assessed, with statistical significance determined using two-sided P values < 0.05.

Simulation-based intervention analysis

Following the NodeIdentifyR framework, symptom-specific intervention simulations were conducted using a binary Ising model. The original 0–10 symptom scores were dichotomized as symptom absent (score=0) or present (score > 0), and a binary Ising model was estimated using the IsingFit package.¹⁸ The estimated edge-weight matrix and symptom-specific threshold parameters were extracted for the subsequent intervention simulations.

The binary Ising model was evaluated separately from the Gaussian graphical network. Specifically, observed symptom activation probabilities were compared with probabilities obtained from binary symptom states simulated using the estimated Ising model parameters. Agreement was summarized using the mean absolute differences between observed and simulated activation probabilities across the 22 symptoms. The bootstrap accuracy and centrality stability procedures used for the Gaussian graphical network were not applied to the binary Ising model.

Two simulation conditions were examined for each symptom: an alleviating intervention and an aggravating intervention. These interventions were operationalized as symptom-specific threshold perturbations, with the target threshold decreased or increased by two SD of the full set of estimated symptom thresholds, respectively.³⁴ During each simulation, the edge-weight matrix and the thresholds of all other symptoms were kept unchanged. These threshold changes represented a modeled decrease or increase in the activation propensity of the target symptom.

For each intervention condition, 5000 binary symptom profiles were generated using the fixed edge-weight matrix and the corresponding threshold values. Within each repetition, a common baseline sample generated from the original, unmodified Ising model parameters was used for all alleviating and aggravating intervention comparisons. No post-intervention networks were re-estimated.

For each simulated participant, the 22 binary symptom indicators were summed to obtain the modeled number of symptoms present. The intervention effect (Δ) was defined as the difference in the mean modeled number of symptoms present between the intervention condition and the common baseline condition.

To evaluate the stability of the primary ± 2 SD intervention results, the complete simulation procedure was repeated 1000 times using a prespecified random seed. For each symptom and intervention direction, we summarized the distribution of Δ across repetitions using the mean and median Δ, the empirical 2.5th–97.5th percentile simulation interval, the mean and median rank, the percentage of repetitions in which the symptom ranked first, and the percentage in which it ranked among the top three.

Sensitivity analyses were additionally performed using ± 1 SD and ± 1.5 SD threshold shifts, with the primary ± 2 SD analysis retained for comparison. For each perturbation magnitude, the intervention simulations were repeated 1000 times. Directional consistency was assessed by whether the sign of Δ remained unchanged across perturbation magnitudes, and ranking consistency was quantified using Spearman rank correlation coefficients (ρ) based on the mean ranks of all 22 symptoms.

Results

Characteristics of the participants

A total of 508 participants were included in the analysis. The mean age was 56.63 ± 13.00 years, and most patients were male (n = 315, 62.01%) and of Han ethnicity (n = 506, 99.61%). Spouses were the most common primary caregivers (n = 331, 65.16%). Most tumors originated from the oral cavity proper (n = 383, 75.39%), and most participants had primary tumors (n = 453, 89.17%). Flap reconstruction had been performed in 399 participants (78.54%). Over 80% of participants were receiving radiotherapy alone at the time of assessment. Additional sociodemographic and clinical characteristics are summarized in Table 1.

Table 1.

Characteristics of the participants (N = 508).

Characteristics		n (%)
Age (years), mean ± SD		56.63 ± 13.00
Body mass index (kg/m2), mean ± SD		22.35 ± 3.16
Sex	Male	315 (62.01)
	Female	193 (37.99)
Ethnicity	Han	506 (99.61)
	Minority	2 (0.39)
Employment status	Employed	163 (32.09)
	On sick leave	47 (9.25)
	Unemployed	101 (19.88)
	Retired	197 (38.78)
Marital status	Married	456 (89.76)
	Divorced	19 (3.74)
	Single	19 (3.74)
	Widowed	14 (2.76)
Education level	Primary school or below	89 (17.52)
	Secondary school	191 (37.60)
	Post-secondary	153 (30.12)
	University or above	75 (14.76)
Medical payment method	Insurance	422 (83.07)
	Self-pay	34 (6.69)
	Other	52 (10.24)
Primary caregiver	Self	88 (17.32)
	Spouse	331 (65.16)
	Parents or children	69 (13.58)
	Other	20 (3.94)
Comorbidities	No comorbidities	295 (58.07)
	Hypertension	142 (27.95)
	Diabetes	49 (9.65)
	Cardiovascular disease	18 (3.54)
	Other	70 (13.78)
Smoking history	No	275 (54.13)
	Former smoker	209 (41.14)
	Current smoker	24 (4.72)
Alcohol history	No	313 (61.61)
	Former drinker	186 (36.61)
	Current drinker	9 (1.77)
Location of tumor	Oral cavity proper	383 (75.39)
	Oral vestibule	125 (24.61)
Cancer status	Primary	453 (89.17)
	Recurrence	55 (10.83)
Flap reconstruction	Yes	399 (78.54)
	No	109 (21.46)
With indwelling tubes	Gastric feeding tube	247 (48.62)
	Tracheostomy tube	162 (31.89)
	None	238 (46.85)
Diet type	Regular diet	210 (41.34)
	Therapeutic diet	298 (58.66)
Treatment modality	Radiotherapy alone	410 (80.71)
	Radiotherapy plus other treatments	98 (19.29)

Multiple responses were allowed for comorbidities and indwelling tubes. SD, standard deviation.

Symptom prevalence and severity

Symptom prevalence across the 22 assessed symptoms ranged from 24.80% to 89.37% (Fig. 2). The most prevalent symptoms were dry mouth (n = 454, 89.37%), followed by fatigue (n = 441, 86.81%), and mouth or throat mucus (n = 426, 83.86%). In terms of symptom severity, dry mouth had the highest mean severity score (4.80 ± 2.81), followed by taste changes (3.90 ± 3.06) and mouth or throat mucus (3.66 ± 2.77).

Fig. 2.

Symptom prevalence among patients with oral cancer undergoing radiotherapy.

Q1, pain; Q2, fatigue; Q3, nausea; Q4, disturbed sleep; Q5, distress; Q6, shortness of breath; Q7, forgetfulness; Q8, lack of appetite; Q9, drowsiness; Q10, dry mouth; Q11, sadness; Q12, vomiting; Q13, numbness or tingling; Q14, mouth or throat mucus; Q15, difficulty swallowing or chewing; Q16, choking; Q17, difficulty speaking; Q18, skin pain; Q19, constipation; Q20, taste changes; Q21, mouth or throat soreness; Q22, teeth or gum problems.

Symptom cluster analysis

Sampling adequacy was supported by a KMO value of 0.88 and a significant Bartlett's test of sphericity (χ² = 3932.72, P < 0.001). The parallel analysis and scree plot figures are presented in Supplementary Fig. S1. Although parallel analysis suggested a two-factor solution, alternative two-, three-, and four-factor solutions were further compared based on explained variance, factor loading patterns, and clinical interpretability. The four-factor solution was retained because it demonstrated a clearer symptom structure and greater clinical interpretability, accounting for 42.02% of the total variance. Cronbach's α values across the four clusters ranged from 0.673 to 0.829. Detailed comparisons of the alternative factor solutions are provided in Supplementary Table S1. Based on the symptom characteristics, the four factors were named as follows. The oral local toxicity symptom cluster included difficulty swallowing or chewing and choking. The chronic psychoneurological symptom cluster comprised fatigue, disturbed sleep, distress, shortness of breath, forgetfulness, drowsiness, sadness, and constipation. The head and neck mucosal-physical symptom cluster encompassed pain, lack of appetite, dry mouth, numbness or tingling, mouth or throat mucus, difficulty speaking, skin pain, taste changes, mouth or throat soreness, and teeth or gum problems. The gastrointestinal symptom cluster included nausea and vomiting (Table 2).

Table 2.

Factor structure and loadings of symptom clusters in patients with oral cancer undergoing radiotherapy (N = 508).

Symptom cluster	Cluster composition	Factor loading				Cronbach's alpha coefficient
		Factor 1	Factor 2	Factor 3	Factor 4
Oral local toxicity symptom cluster	Q15: Difficulty swallowing or chewing	0.891				0.673
	Q16: Choking	0.430
Chronic psychoneurological symptom cluster	Q2: Fatigue		0.301			0.795
	Q4: Disturbed sleep		0.616
	Q5: Distress		0.590
	Q6: Shortness of breath		0.821
	Q7: Forgetfulness		0.515
	Q9: Drowsiness		0.382
	Q11: Sadness		0.378
	Q19: Constipation		0.413
Head and neck mucosal-physical symptom cluster	Q1: Pain			0.872		0.829
	Q8: Lack of appetite			0.616
	Q10: Dry mouth			0.615
	Q13: Numbness or tingling			0.392
	Q14: Mouth or throat mucus			0.388
	Q17: Difficulty speaking			0.369
	Q18: Skin pain			0.629
	Q20: Taste changes			0.604
	Q21: Mouth or throat soreness			0.627
	Q22: Teeth or gum problems			0.306
Gastrointestinal symptom cluster	Q3: Nausea				0.704	0.744
	Q12: Vomiting				0.735
Variance explained		7.03%	11.58%	15.87%	7.54%
Cumulative variance		7.03%	18.61%	34.48%	42.02%

Network analysis

Network structure

The Gaussian graphical network estimated from the original 0–10 symptom severity scores is shown in Fig. 3. Both positive and negative conditional associations were observed among symptoms, although most retained edges were positive. The strongest regularized partial correlation was observed between nausea (Q3) and vomiting (Q12) (edge weight = 0.418), followed by distress (Q5) and sadness (Q11) (0.295) and pain (Q1) and mouth or throat soreness (Q21) (0.293). Detailed edge weights for all symptom pairs are provided in Supplementary Table S2.

Fig. 3.

Symptom network of patients with oral cancer undergoing radiotherapy.

Blue and red edges indicate positive and negative conditional associations, respectively. Edge thickness reflects the absolute strength of the regularized partial association. Node positions were determined by the Fruchterman–Reingold layout.

Node centrality and bridge strength

Based on the Gaussian graphical network estimated from the original continuous symptom scores, raw strength centrality values are presented in Fig. 4. Disturbed sleep (Q4) showed the highest strength centrality (1.182), followed by mouth or throat soreness (Q21) (1.084) and fatigue (Q2) (1.077). Raw bridge strength estimates are presented in Fig. 5. Choking (Q16) showed the highest bridge strength (0.670), followed by difficulty swallowing or chewing (Q15) (0.646) and fatigue (Q2) (0.566). A sensitivity analysis comparing the four-factor and two-factor solutions showed that bridge strength rankings varied according to the underlying factor structure. In the four-factor solution, choking ranked first in bridge strength, whereas difficulty swallowing or chewing showed the highest bridge strength in the two-factor solution. Detailed results are provided in Supplementary Table S3.

Fig. 4.

Raw strength centrality values of the Gaussian graphical symptom network.

Q1, pain; Q2, fatigue; Q3, nausea; Q4, disturbed sleep; Q5, distress; Q6, shortness of breath; Q7, forgetfulness; Q8, lack of appetite; Q9, drowsiness; Q10, dry mouth; Q11, sadness; Q12, vomiting; Q13, numbness or tingling; Q14, mouth or throat mucus; Q15, difficulty swallowing or chewing; Q16, choking; Q17, difficulty speaking; Q18, skin pain; Q19, constipation; Q20, taste changes; Q21, mouth or throat soreness; Q22, teeth or gum problems.

Fig. 5.

Bridge strength of symptoms in the four-factor symptom network.

The accuracy and stability of the estimated symptom network are shown in Fig. 6. The CS coefficient for strength centrality was 0.594, indicating good stability, whereas closeness and betweenness showed lower stability and should be interpreted cautiously.

Fig. 6.

Accuracy and stability of the symptom network (A) Non-parametric bootstrap results for edge-weight accuracy (B) Case-dropping bootstrap results for centrality stability.

Subgroup analysis

NCTs were conducted to explore potential differences in symptom network structures across clinically relevant subgroups, including sex, anatomical site, and treatment modality. No statistically significant differences were identified in network structure invariance or global strength invariance across any subgroup comparisons (all P > 0.05). Detailed subgroup network structures and NCT results are presented in Supplementary Fig. S2–S4 and Supplementary Table S4.

Simulation-based intervention effects and robustness analyses

Binary Ising model evaluation

The intervention simulation analyses were based on the binary Ising model estimated from the dichotomized symptom data. The estimated edge-weight matrix and symptom-specific threshold parameters are presented in Supplementary Table S5 and S6. The mean absolute difference between the observed and simulated symptom activation probabilities across the 22 symptoms was 0.026, indicating that the model reasonably reproduced the observed marginal symptom activation patterns.

Stability of the primary ± 2 SD intervention effects

In the primary ± 2 SD analysis, the threshold perturbation procedure was repeated 1000 times, and the mean intervention effects with their empirical 95% simulation intervals are shown in Fig. 7. In the alleviating simulations, pain (Q1) showed the largest mean effect (mean Δ = −0.696), followed closely by mouth or throat soreness (Q21; mean Δ = −0.686). Pain ranked first in 51.3% of repetitions and among the top three in 92.7%, whereas mouth or throat soreness ranked first in 38.3% and among the top three in 87.4%. These results indicate that pain and mouth or throat soreness were both among the leading potential leverage points, while the exact first-place ranking was not fully stable.

Fig. 7.

Simulated intervention effects across 1000 repetitions. Points represent the mean intervention effect (Δ), and horizontal error bars represent the empirical 2.5th–97.5th percentile simulation interval across 1000 repetitions of the primary ± 2 SD analysis. Δ denotes the difference in the mean modeled number of symptoms present between the intervention and common baseline conditions (A) Aggravating interventions; more positive Δ values indicate larger modeled increases in the mean number of symptoms present (B) Alleviating interventions; more negative Δ values indicate larger modeled reductions in the mean number of symptoms present. Dashed vertical lines indicate Δ = 0.

In the aggravating simulations, vomiting (Q12) showed the largest mean effect (mean Δ = +0.514), ranked first in 75.7% of repetitions, and ranked among the top three in 98.5%. Shortness of breath (Q6) and forgetfulness (Q7) followed in mean effect (mean Δ = +0.474 and + 0.436, respectively). Thus, vomiting showed the largest mean aggravating effect across repeated simulations, although it did not rank first in every repetition. Detailed stability statistics are provided in Supplementary Table S7 and S8.

Sensitivity to perturbation magnitude

Sensitivity analyses using ± 1 SD, ± 1.5 SD, and ± 2 SD threshold shifts showed complete consistency in the direction of the simulated effects. All 22 alleviating effects remained negative and all 22 aggravating effects remained positive across the three perturbation magnitudes, corresponding to directional consistency in all 44 symptom-by-direction combinations.

The overall symptom rankings were also broadly similar across perturbation magnitudes. For the alleviating simulations, Spearman correlations between the mean ranks were ρ = 0.980 for ± 1 SD vs. ± 1.5 SD, ρ = 0.937 for ± 1 SD vs. ± 2 SD, and ρ = 0.977 for ± 1.5 SD vs. ± 2 SD. For the aggravating simulations, the corresponding correlations were ρ = 0.957, 0.914, and 0.989. Nevertheless, the exact ordering of the leading symptoms varied somewhat with the perturbation magnitude. In the alleviating simulations, skin pain (Q18) had the best mean rank at ± 1 SD, whereas pain (Q1) had the best mean rank at both ± 1.5 SD and ± 2 SD. In the aggravating simulations, vomiting (Q12) showed the largest mean Δ at all three perturbation magnitudes, whereas the relative rankings of other highly ranked symptoms, including shortness of breath (Q6), forgetfulness (Q7), and skin pain (Q18), varied across settings. These findings indicate consistent effect directions and broadly similar overall ranking patterns, but do not support an invariant ordering of the leading symptoms. Detailed sensitivity results are provided in Supplementary Table S9–S11.

Discussion

Main findings

To our knowledge, this is the first study to characterize the contemporaneous symptom network and explore potential symptom leverage points using intervention simulation implemented through symptom-specific threshold perturbations in patients with oral cancer undergoing radiotherapy.

We identified four symptom clusters, and their composition was largely consistent with prior expectations. One exception was the inclusion of “lack of appetite” in the head and neck mucosal-physical symptom cluster. Some variability in cluster composition is expected because of differences in symptom assessment tools, analytical methods, and clustering criteria across studies.³⁵ This finding may also reflect the anatomical complexity and treatment-related characteristics of our study population. Moreover, more than half of the participants had at least one indwelling tube, and nearly half had a gastric feeding tube, which may reflect treatment-related changes in oral intake. To date, no study has examined symptom clusters specifically in patients with oral cancer undergoing radiotherapy. However, a systematic review by Mathew et al.³⁶ identified three major symptom clusters in patients with head and neck cancer (general, cancer-specific, and gastrointestinal), which is broadly consistent with our findings. The cluster structure observed here may therefore reflect both common symptom patterns in head and neck cancer and characteristics specific to this clinical population. Overall, identifying symptom clusters provides a clinically meaningful framework that may help clinicians better understand symptom co-occurrence and facilitate integrated symptom management.

Disturbed sleep showed the highest strength centrality in our network, suggesting that it was one of the most highly connected symptoms among patients with oral cancer undergoing radiotherapy. Sleep disturbance may occupy a central position because it may reflect the combined effects of physical discomfort, psychological distress, and treatment-related burden during cancer therapy. Although disturbed sleep does not consistently emerge as the symptom with the highest centrality across cancer populations, previous network analyses have demonstrated its important role within interconnected symptom structures. Zhu et al.¹⁴ reported that disturbed sleep showed one of the highest predictability values among symptoms in cancer survivors, indicating close integration with neighboring symptoms. Furthermore, studies conducted among patients receiving radiotherapy have highlighted the importance of sleep disturbance within treatment-related symptom networks. In patients with head and neck cancer undergoing intensity-modulated radiotherapy, sleep disturbance was closely involved in the psychoneurological symptom network, suggesting an important role in symptom interactions during active treatment.³¹ Beyond sleep-related symptoms, mouth or throat soreness showed the second highest strength centrality in our network. This may reflect the prominent role of oral symptom burden during radiotherapy in patients with oral cancer. Consistent with our findings, Li et al.³⁷ showed that mouth or throat sores occupied an important position in the dynamic symptom network among patients with oral cancer undergoing postoperative radiotherapy and were associated with subsequent pain, dental problems, and difficulty swallowing or chewing. Previous symptom network studies have also suggested that the relative importance of individual symptoms may vary according to cancer type and treatment context.²⁹ Thus, the high strength centrality of mouth or throat soreness may reflect its close connections with both local oral symptoms and broader treatment-related symptoms. Fatigue ranked third in strength centrality, which is consistent with previous symptom network studies among patients with cancer. A large network analysis involving 1330 cancer survivors reported that fatigue consistently emerged as a highly connected symptom across different cancer types, treatment modalities, and survivorship stages, and was strongly associated with multiple emotional symptoms.³⁸ However, the centrality of fatigue may vary according to cancer type, treatment phase, and disease trajectory. A longitudinal study by Rha and Lee involving 249 patients demonstrated that the network position of fatigue changed across treatment cycles, suggesting that its importance may fluctuate during cancer treatment.³⁹ In our study, the active radiotherapy setting may partly explain why fatigue remained highly connected within the symptom network. Subgroup analyses showed generally consistent network structures across sex, anatomical site, and treatment modality groups, and no significant differences were detected in network structure or global strength. These findings suggest broadly similar network organization across the examined subgroups, although interpretation should remain cautious because subgroup sizes were unequal and some comparisons may have had limited statistical power.

Dry mouth, specifically radiation-induced xerostomia, was the most prevalent symptom and had the highest mean severity score. Clinically, it impairs chewing and swallowing, increases the risk of dental caries and secondary oral infections, and substantially diminishes quality of life.⁴⁰ Despite its high prevalence and severity, xerostomia did not emerge as one of the most central symptoms in the network. This finding is consistent with previous reports showing that the most prevalent or severe symptoms are not necessarily the most central.^14,41 In network analysis, centrality reflects a symptom's statistical connectivity and its potential covariation with other symptoms,²⁷ rather than its prevalence alone. A highly prevalent symptom, such as xerostomia, may exhibit limited interindividual variation, which can reduce its covariation with other symptoms and, consequently, its network connectivity. In contrast, disturbed sleep, mouth or throat soreness, and fatigue showed stronger connectivity with multiple symptoms. Thus, prevalence, severity, and network centrality capture different aspects of clinical importance and should be considered together rather than interchangeably.

Choking demonstrated the highest bridge strength in the four-factor symptom network, indicating its potential role in connecting different symptom clusters. Jones et al.⁴² suggested that bridge symptoms may play an important role in linking otherwise distinct symptom communities. In our network, choking was connected with difficulty swallowing or chewing (Q15) and difficulty speaking (Q17). Cough is an airway-protective reflex that helps clear the airway.⁴³ Previous studies have shown that swallowing impairment increases the risk of aspiration and choking, whereas speech difficulties may reflect impaired tongue-pharyngeal coordination and weakened airway protective reflexes, further increasing susceptibility to choking.^44,45 These coexisting functional impairments may interfere with eating and drinking and restrict participation in social mealtimes, thereby affecting patients’ social interaction and psychosocial well-being.⁴⁶ Accordingly, choking may reflect an important functional link between local swallowing-related symptoms and broader symptom domains. Moreover, sensitivity analyses indicated that bridge symptom rankings varied across alternative factor structures, suggesting that the identification of bridge symptoms should be interpreted within the context of the underlying symptom cluster framework.

The simulation-based intervention analyses identified several symptoms with relatively large modeled effects. In the aggravating simulations, vomiting showed the largest mean modeled increase in the number of symptoms present across 1000 repetitions and ranked first in 75.7% of repetitions. Clinically, vomiting remains an important symptom requiring prompt assessment and etiological investigation, particularly because it may accompany treatment-related toxicity or other clinically significant conditions.⁴⁷ Within the simulation framework, this finding indicates that increasing the modeled activation propensity of vomiting was associated with the largest mean increase in the number of symptoms present.

In the alleviating simulations, pain showed the largest mean modeled reduction in the number of symptoms present, followed closely by mouth or throat soreness. This pattern may be clinically plausible because radiation-induced oral mucositis is a common and debilitating treatment-related toxicity that can cause substantial oral pain.⁴⁸ Mouth or throat soreness may also reflect mucosal injury involving the oral cavity and pharynx. Such mucosal toxicity is frequently accompanied by dysphagia, impaired oral intake, and nutritional deterioration and may therefore coexist with multiple functional and treatment-related symptoms.⁴⁹ These interrelated consequences may help explain why perturbations of pain and mouth or throat soreness were associated with relatively large modeled reductions in the mean number of symptoms present.

The repeated-simulation and sensitivity analyses further showed that the exact ordering of the leading symptoms was not invariant. In the primary ± 2 SD analysis, pain and mouth or throat soreness showed closely similar alleviating effects, whereas vomiting showed the largest mean aggravating effect but did not rank first in every repetition. Across ± 1 SD, ± 1.5 SD, and ± 2 SD perturbations, the direction of all simulated effects remained unchanged, and the overall symptom rankings were strongly correlated across settings. However, the exact positions of several highly ranked symptoms varied with the perturbation magnitude, particularly in the alleviating simulations. These findings support broadly consistent overall patterns while suggesting that the leading symptoms are better considered as a group of potential leverage points than as a fixed rank order.

The symptoms identified as influential in the simulation analysis did not fully overlap with those showing the highest centrality in the Gaussian graphical network. Previous work has similarly emphasized that node centrality and simulated intervention impact capture related but non-equivalent properties of a networked system.¹⁸ This discrepancy is therefore not unexpected because the two approaches quantify different aspects of the symptom system. Strength centrality reflects the extent to which a symptom is directly connected to other symptoms through the estimated partial-correlation structure and therefore characterizes its topological prominence within the estimated network. In contrast, the simulation analysis evaluates the modeled system-level change associated with perturbing the activation propensity of a single symptom while holding the remaining Ising model parameters fixed. Thus, centrality and simulated intervention effects should be regarded as complementary rather than interchangeable: centrality describes structural connectedness, whereas the simulation analysis provides an estimate of the modeled change in the symptom system associated with targeted perturbation.

Implications for nursing practice and research

Our findings have several implications for clinical nursing practice. Highly prevalent or severe symptoms, such as dry mouth, should continue to receive dedicated symptom-specific assessment and supportive care regardless of their network position. For disturbed sleep, mouth or throat soreness, and fatigue, which exhibited high centrality, health care professionals may routinely assess sleep quality, oral symptom burden, and fatigue levels using validated instruments and provide individualized supportive care according to patients' symptom profiles. Choking, which showed the highest bridge strength in the four-factor symptom network, may also warrant particular attention because of its position as a bridge symptom and its relevance to swallowing-related functional impairment. The simulation findings provide an additional, complementary perspective. Modeled decreases in the activation propensity of pain and mouth or throat soreness were associated with relatively large reductions in the mean number of symptoms present, whereas an increase in the modeled activation propensity of vomiting was associated with the largest mean increase. These findings suggest that pain, mouth or throat soreness, and vomiting may be useful symptoms for closer assessment and prospective evaluation, but the simulated effects should not be interpreted as establishing direct causal effects or fixed treatment priorities. In practice, symptom management may therefore benefit from an integrated framework that considers symptom prevalence and severity, patients’ immediate clinical needs, network characteristics such as strength and bridge strength, and simulation-based modeled effects rather than relying on any single metric.

Several directions for future research warrant consideration. As the present findings are based on cross-sectional data, they are hypothesis-generating and cannot capture dynamic changes in symptom relationships over time. Future studies should adopt longitudinal designs incorporating detailed treatment-phase and radiation-exposure information to characterize changes in symptom networks and symptom trajectories across the treatment course. Large-scale multicenter prospective studies are also needed to evaluate the generalizability of the present network structure and simulation findings. In particular, the symptoms showing relatively large modeled effects should be prospectively examined to determine whether changes in these symptoms are followed by broader changes in symptom occurrence. Ultimately, prospective intervention studies are required to determine whether targeting these symptoms may produce clinically meaningful benefits.

Limitations

Some limitations of this study should be acknowledged. First, we did not collect complete tumor-node-metastasis (TNM) staging data, the exact week of radiotherapy, or the cumulative radiation dose, precluding stage-stratified or phase-specific network analyses. Although we performed subgroup analyses according to treatment modality, these were not equivalent to phase- or dose-specific analyses. Therefore, the generalizability of our findings across disease stages remains uncertain. Second, although the overall sample size was adequate, some subgroup analyses involved relatively small sample sizes, which may have reduced the power to detect subtle network differences. Therefore, subgroup findings should be interpreted cautiously, and the lack of statistically significant differences should not be considered evidence of network equivalence. Future studies with larger and more balanced samples are needed to confirm the observed subgroup patterns. Third, we conducted the exploratory factor analysis using the same dataset that was used for the network analysis, which may have introduced overfitting. Consequently, the four symptom clusters identified should be regarded as exploratory. Confirmatory factor analysis in an independent cohort is needed to validate the factor structure. Fourth, given the cross-sectional design, repeated simulations and perturbation-magnitude sensitivity analyses supported the robustness of the modeled results within the fitted Ising framework, but they do not provide evidence that these simulated effects would occur over time or following real-world symptom modification. Therefore, the simulated intervention effects should be interpreted as model-based and hypothesis-generating rather than as evidence of temporal or causal effects. Finally, although participants with known severe mental disorders were excluded, several measurement limitations should be acknowledged. Our screening relied solely on medical records and brief clinical interviews without standardized instruments (e.g., Patient Health Questionnaire-9, Generalized Anxiety Disorder-7, or suicide ideation scales) or predefined referral pathways for psychological distress. In addition, symptom network estimation was based on single-item measures from the MDASI-H&N, which may introduce measurement error despite the instrument's established validity. These limitations may have affected the detection of subthreshold symptoms and the precision of the estimated symptom network. Future studies should incorporate multi-item instruments, standardized risk assessment protocols, and clearly defined referral pathways to improve the robustness and clinical applicability of the findings.

Conclusions

This study enhances understanding of symptom experiences in patients with oral cancer through an integrated application of network analysis and simulation-based approaches. Disturbed sleep, mouth or throat soreness, and fatigue emerged as symptoms with high strength centrality, while choking showed the highest bridge strength in the four-factor symptom network. In the simulation analyses, modeled decreases in the activation propensity of pain and mouth or throat soreness were associated with relatively large reductions in the mean number of symptoms present, whereas an increase in the modeled activation propensity of vomiting was associated with the largest mean increase. Taken together, these findings identify several potential symptom leverage points rather than a fixed hierarchy of intervention targets. The results are exploratory and hypothesis-generating and may inform future research on symptom management. Prospective and longitudinal studies are needed to determine whether modifying these symptoms leads to clinically meaningful downstream benefits.

CRediT authorship contribution statement

Pingping Ji: Conceptualization, Methodology, Writing - original draft, Writing - Review & Editing. Zheng Zhu: Methodology, Validation, Writing - Review & Editing, Supervision. Yayuan Tian: Investigation, Data curation. Wentao Shi: Software, Formal analysis, Visualization. Jingjing Dai: Resources, Data curation, Project administration. Lili Hou: Project administration, Funding acquisition. Guihua Hao: Conceptualization, Supervision, Project administration, Funding acquisition. All authors have read and approved the final manuscript.

Ethics statement

The study was approved by the Medical Ethics Committee of the Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine (Approval No. SH9H-2023-T215-2) 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 on request from the corresponding author, GH. The data are not publicly available due to their containing information that could compromise the privacy of research participants. The study-specific R scripts and reproducibility materials are publicly available at https://github.com/1057137713-netizen/oral-cancer-symptom-network-analysis.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT (OpenAI) for language editing and improvement of readability. The authors subsequently reviewed and edited the content and take full responsibility for the final manuscript.

Funding

This study was supported by Shanghai Jiao Tong University School of Medicine (Grant No. SJTUHLXK2024); the Nursing Fund of Ninth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (Grant Nos. JYHL2023QN01 and JYHL2025ZD03). The funders had no role in the study design; collection, analysis, or interpretation of data; writing of the report; or decision to submit the article for publication.

Declaration of competing interest

The authors declare no conflict of interest.

Appendix. Supplementary materials

The following is the Supplementary data to this article.