This prospective pragmatic interventional study was conducted at the affiliated hospital of Qingdao university (Qingdao, China; approval no. QYFY-WZLL-30165). We searched patients pathologically diagnosed with NPC from March 1, 2023, to November 25, 2024, who planning to receive definitive CCRT, and a total of 108 patients attended.

The inclusion criteria were as follows: (1) Newly diagnosed with histologically confirmed stage III-IVa NPC according to the 8th edition of the American Joint Committee on Cancer (AJCC) staging system; (2) No prior surgical intervention for nasopharyngeal tumor other than diagnostic biopsy; (3) Aged between 18 and 70 years; (4) Karnofsky Performance Status (KPS) score ≥ 70; (5) Provided written informed consent. The exclusion criteria included the following: (1) Clinically relevant organ dysfunction or comorbidities with the potential to compromise radiotherapy tolerance or bias outcome evaluation, including metabolic syndrome requiring intensive medical management; end-stage renal disease requiring hemodialysis; solid organ transplantation; diabetes mellitus with severe complications (diabetic ketoacidosis, hyperosmolar hyperglycemic state, or established end-organ damage); ischemic heart disease with unstable symptoms; active infectious diseases such as human immunodeficiency virus (HIV) infection or tuberculosis; clinically uncontrolled thyroid or other endocrine disorders; and autoimmune diseases requiring systemic treatment; (2) History of long-term use of immunosuppressive agents or corticosteroids (≥3 months); (3) Presence of another active malignancy; (4) Patients with severe pre-existing oral conditions (including active oral infection, uncontrolled periodontitis, ≥2 non–radiotherapy-related oral ulcers, oral abscesses, or other conditions requiring dental intervention); (5) Upper respiratory tract infection within the past 2 weeks or antibiotics used during follow-up; (6) History of head and neck radiation; (7) Presence of cognitive impairment or inability to cooperate with follow-up. (Figure 1).

The target volume and radiotherapy plan for all enrolled patients were delineated and designed based on international consensus guidelines (38–40). The radiotherapy was delivered following the International Commission on Radiation Units and Measurements (ICRU) Reports No. 83 (41). The concurrent chemotherapy regimen consisted of cisplatin (100 mg/m², administered every 3 weeks).

Before radiotherapy initiation, all patients completed comprehensive and professional oral pre-treatment by dentists to be appropriated for radiotherapy and received oral hygiene instructions with reference to multiple international guidelines on radiation. They maintained basic oral hygiene with Bass Method of toothbrushing, non-medicated mouthwashes, and a hospital-prepared Chinese herbal formula, Qingdi Mixture, which contained a blend of nine herbs (e.g., Rehmannia, Ophiopogon, Scutellaria) designed to provide anti-inflammatory, antioxidant, and immunomodulatory benefits against radiation-induced mucositis.

Before the initiation of radiotherapy, all eligible patients received standardized oral-care education. To minimize physician- or patient-driven selection, eligible patients were sequentially assigned to one of three recommendation strategies in a 1:1:1 rotating sequence: no probiotic recommendation, recommendation of SsK12, or recommendation of SsM18, independent of clinical characteristics, contraindications, or patient preferences. Probiotic use was recommendation-based rather than mandated. Adherence was monitored prospectively throughout treatment. Patients who did not adhere to the assigned recommendation were excluded from the analysis cohort. The natural microbiota exposure group (reference group) served as a routine-care comparator, representing endogenous oral microbiota exposure in the absence of probiotic intervention, while The SsK12 and SsM18 groups represent routine-care-based probiotic supplementation strategies. This design represents a prospective pragmatic interventional study, and baseline characteristics were compared across groups to evaluate comparability.

The probiotic lozenges employed in this study contained SsK12 or SsM18, supplied as commercially available oral lozenges (Over-the-Counter dietary supplement). The products were manufactured by BLIS Technologies Ltd. (Dunedin, New Zealand) and marketed under the names BLIS K12™ and BLIS M18™. Each over-the-counter lozenge contained 2.5 × 10⁹ CFU of viable bacteria. From the start to the end of radiotherapy, patients were instructed to dissolve one lozenge in the oral cavity twice daily, followed by refraining from eating or drinking for at least 1 hour to promote colonization. The supplements were stored in a dry condition at 0–5 °C in a household refrigerator. Furthermore, the natural microbiota exposure group was instructed to rinse their mouths with normal saline twice a day. With routine supportive care, including Chinese patent medicines and analgesics, standardized across all groups per institutional protocols, probiotic exposure served as the main distinguishing intervention. However, patients who used antibiotics were withdrawn from the study.

On the basis of our institutional data and previous reports, the average duration of OM in patients receiving head and neck chemoradiotherapy is approximately 90 days, and a reduction of about 30 days was considered clinically meaningful for patients receiving SsM18. Assuming a common standard deviation of 30 days (Cohen’s d ≈ 1.0), the required sample size for detecting this difference in a two-sample comparison of means (two-sided α = 0.05, 1 − β = 0.80) was calculated using the standard formula:

where σ is the common standard deviation and Δ is the expected mean difference. Substituting Z_1−α/2 = 1.96, Z_1–β = 0.84, and Δ = 30 yields an estimated requirement of approximately 16 patients per group. As this exploratory cohort ultimately enrolled 23 patients per group, the achieved sample size exceeds this threshold and provides an estimated statistical power of approximately 90% to detect a 30-day difference in OM duration. Smaller or more modest differences may remain underpowered.

Assessments were performed at baseline, weekly during radiotherapy, and at predefined post-treatment time points.

Prior to radiotherapy, all patients underwent assessment using the modified Beck Oral Assessment Scale (BOAS, Supplementary Table 1), by five domains: lips, gingiva, tongue, teeth, and saliva (42–44).

During radiotherapy, patients were followed weekly by two experienced radiation oncologists to document OM severity (WHO criteria), xerostomia, probiotic adherence, and concomitant medications, with post-radiotherapy follow-up at 2 weeks and 6 months for late OM degree (WHO criteria, Supplementary Table 2). SOM was defined as OM with a WHO grade 3–4 (5–7, 9, 45). OM and SOM duration were defined as the total number of consecutive days with WHO grade ≥1 and ≥3, respectively. For both, the start date was the first day the grade reached the respective threshold (≥1 for OM, ≥3 for SOM); the end date was the first day thereafter that the grade fell below that threshold without recurrence. The onset and resolution dates were determined based on patient-reported symptom timing and confirmed by trained investigators during radiotherapy or additional clinic visits. Both the calendar date and the corresponding radiotherapy fraction at onset and resolution were recorded. All durations were calculated based on continuous day counts rather than weekly intervals. Weekly grades were used exclusively for trajectory analysis and did not affect day-level duration calculations.

Blood samples were collected in the early morning at the time of the 15th radiotherapy fraction. After centrifugation at 1,400 × g for 15 minutes at 4 °C, serum levels of IL-6, Interleukin 8 (IL-8), Interferon-alpha (IFN-α), Interferon-gamma (IFN-γ), and TNF were quantified using ELISA kits (ThermoFisher Scientific, United States of America, USA), with absorbance measured at 450 nm on a Cytation multimode reader (BioTek, USA). Weekly routine blood tests, including complete blood count and albumin, were performed serially during treatment in hospital.

Acute radiotherapy-related toxicities were graded using Common Terminology Criteria for Adverse Events version 5.0 (CTCAE v5.0, Supplementary Table 3) at the 15th and 30th fractions and summarized for between-group comparisons. CTCAE v5.0 was also used to record treatment-emergent adverse events throughout radiotherapy for safety evaluation only including gastrointestinal disorders and hematologic toxicities (Supplementary Table 4). Hepatic and renal toxicities were defined by standard laboratory abnormalities (serum transaminases, serum creatinine, blood urea nitrogen, and estimated glomerular filtration rate). Late tissue toxicity was assessed using the Radiation Therapy Oncology Group (RTOG, Supplementary Table 5) criteria with post-radiotherapy follow-up at 6 months.

Additional parameters recorded throughout treatment included nutritional support, weight changes, radiotherapy interruptions, infections, and antibiotic use.

All assessments were independently performed and then jointly reviewed by two radiation oncologists, each with over five years of experience in head and neck radiotherapy–related toxicity assessment. Examiners underwent standardized training based on the established scoring scale and institutional guidelines prior to study initiation. Any discrepancies resolved through discussion to achieve a consensus grade.

Of the 80 eligible patients enrolled, 69 completed the follow-up and were included in the analysis. Comprehensive baseline data were extracted from the Hospital Information System (HIS) and the Varian radiotherapy system. Data included demographic characteristics (sex, age, body mass index, BMI), medical history (diabetes, hypertension), lifestyle factors (heavy smoking, weekly standard drinks), and Nutritional Risk Screening 2002 (NRS2002) score (46) (Supplementary Table 6). Laboratory parameters consisted of complete blood cell counts (neutrophils, lymphocytes, monocytes, eosinophils, basophils, hemoglobin, platelet), albumin, and Epstein-Barr virus DNA load. Pathological characteristics included tumor differentiation, TNM stage, and clinical stage. Treatment-related variables covered the number of chemotherapy cycles, application of Tomotherapy, the dose to 50% of the oral cavity volume (ORC, D50%), and the mean dose to 30% of the bilateral parotid glands (PG, mean bilateral parotid dose, D30%). Heavy smoking was defined as a cumulative exposure of ≥ 10 pack-years (47–50). Alcohol consumption was assessed as the average number of standard drinks per week and modeled as a continuous variable, with one standard drink defined as 14 g of pure ethanol (51). Δ Albumin = Albumin (before radiotherapy) - Albumin (after radiotherapy). Radiotherapy interruption is defined as any unplanned suspension of treatment that lasts for 3 or more consecutive working days during the course of radiotherapy. Inflammatory-based indices were calculated as follows: neutrophil-to-lymphocyte ratio (NLR = absolute neutrophil count/absolute lymphocyte count) (52), platelet-to-lymphocyte ratio (PLR = platelet count/absolute lymphocyte count), lymphocyte-to-monocyte ratio (LMR = absolute lymphocyte count/absolute monocyte count), systemic immune-inflammation index (SII = platelets × neutrophils/lymphocytes) (53), systemic inflammation response index (SIRI = neutrophils × monocytes/lymphocytes) (54, 55). Nutritional indicator s such as the prognostic nutritional index (PNI) and body mass index (BMI) were calculated as follows: PNI = albumin + 5 × lymphocytes (56) and BMI = weight (kg)/height squared (m²).

Data were presented as mean (standard deviation, SD), median (interquartile range, IQR), or number (percentage, %) as appropriate. For comparisons of baseline characteristics across the three study groups (natural microbiota exposure, SsK12, and SsM18), the Kruskal–Wallis test was used for skewed continuous variables, one-way ANOVA for normally distributed continuous variables, and the Chi-square or Fisher’s exact test for categorical variables.

To evaluate the effects of probiotic supplementation on OM prevention and other treatment-related outcomes, follow-up data–including WHO oral toxicity score, patient quality of life scores, relevant clinical indicators, and medication use–were compared across the three groups using the same suite of non-parametric and parametric tests (Kruskal–Wallis test, one-way ANOVA, Chi-square test, or Fisher’s exact test). These analyses were conducted to demonstrate the efficacy of probiotics in mitigating radiotherapy-induced adverse events, particularly OM, and to investigate potential safety concerns. Post-hoc pairwise comparisons were performed using the Mann-Whitney U test for continuous variables and Fisher’s exact test for categorical variables to identify the specific sources of any significant overall differences.

Group-based trajectory modeling (GBTM) was employed to identify distinct longitudinal trajectories of OM severity during and after radiotherapy. We followed a two-step procedure based on Nagin’s work to determine the optimal trajectory model. The first step involved selecting the number of latent trajectories: starting with a single-group cubic model, we incrementally increased the number of groups and evaluated model fit using the Bayesian Information Criterion (BIC), average posterior probability of assignment (APPA > 0.70), odds of correct classification (OCC > 5.0), and a minimum of 5% of participants per group. The second step optimized the shape of trajectories by iteratively removing nonsignificant higher-order polynomial terms while retaining linear terms. Weekly OM assessments were prospectively collected, and patients with incomplete follow-up for OM trajectory or duration estimation were excluded from the corresponding analyses. No imputation or interpolation was performed.

Univariate logistic regression was used to identify clinical factors associated with assignment to an OM trajectory group. Variables with p< 0.1 and variance inflation factor (GVIF^(1/(2*Df)))< 5 in univariate analysis were included in a multivariable Firth penalized logistic regression model. This model aimed to identify independent predictors of trajectory group assignment, with the probiotic intervention variable excluded from this step. The Firth correction was applied both to reduce small-sample bias and to address potential separation due to sparse data. To reduce the risk of overfitting, the number of variables included in multivariable models was limited relative to the number of outcome events, and covariates were selected based on univariate screening and clinical relevance rather than automated selection procedures. Due to marked right-skewness, log-transformed values of IL-6 and weekly alcohol consumption (measured in standard drinks) were used in the multivariable models to enhance stability and limit the influence of extreme observations. Because BOAS was a priori selected as the primary oral health indicator, it was retained in multivariable models to ensure interpretability and consistency with the study design. The Teeth variable, being a sub-score of BOAS, was not entered into the same multivariable model to avoid conceptual overlap, structural multicollinearity, and potential overadjustment. Instead, its role was explored through stratified and interaction analyses within the BOAS framework.

Multiplicative interactions between different domains of BOAS (especially Teeth domain) and inflammatory biomarkers were evaluated to assess their joint effects on OM trajectories, which was done by including relevant cross-product terms in multivariable logistic regression models and assessing their significance with Wald tests. Specifically, we added cross-product terms (e.g., Teeth × IL-6) to the models. All interaction models were adjusted for potential covariates identified from univariate analyses (p< 0.10), including IL-8, weekly standard drinks, IFN-α, PNI, albumin, PLR, TNF, Nutritional Risk Screening (NRS) 2002 score, D50 of ORC, and the number of induction chemotherapy (IC) cycles. Classical additive interaction indices were not estimated because the outcome was a categorical trajectory rather than a time-to-event endpoint, and the limited sample size resulted in sparse data and separation across exposure combinations, precluding reliable estimation on the additive risk scale.

Cox proportional hazards analyses were conducted to further investigate factors associated with the timing of OM onset (the key difference between trajectories). For these analyses, the composite BOAS score was decomposed into its individual subscales. The model was adjusted for covariates that showed an association at p< 0.10 in univariable logistic regression. The number of radiotherapy fractions at OM onset was used as the time-to-event variable. The treatment variable was excluded from subsequent multivariable analyses to reduce confounding from intervention. Kaplan–Meier curves were generated to visualize the time to OM onset across different levels of key risk factors. Differences between groups were evaluated using the log-rank test. Kaplan-Meier curves were plotted according to the Teeth subdomain categories only for visual comparison (for descriptive purposes only).

Causal mediation analysis was performed based on multiple linear regression models to investigate the potential mechanisms through which clinical predictors influence OM-related outcomes. This analysis aimed to assess whether clinical predictors, such as probiotic supplementation (SsK12 or SsM18 groups versus the natural microbiota exposure reference group) and teeth status (BOAS dental domain scores 2 or 3 versus score 1), on OM duration and SOM risk were mediated by a panel of inflammatory biomarkers. The biomarkers examined included IL-6, IL-8, IFN-α, PNI. Mediation was considered significant if: (1) the exposure variable was significantly associated with the mediator; (2) the exposure variable was significantly associated with the outcome in a model without the mediator; and (3) the magnitude of the direct exposure–outcome association was attenuated after adjusting for the mediator in the model. The statistical significance of the indirect mediation effect was tested using the bias-corrected bootstrap method with 5,000 resamples to derive 95% confidence intervals (CIs). All mediation models were adjusted for weekly standard drinks, IFN-α, PNI, albumin, PLR, TNF, NRS 2002 score, D50 of ORC, and the number of IC cycles. The analyses relied on the standard assumptions of causal mediation analysis, including sequential ignorability (no unmeasured confounding). We tested for exposure–mediator interaction in preliminary models; as none were significant, final mediation models did not include interaction terms.

All regression, interaction and mediation models were adjusted for potential confounders. When IL-6 (or IL-8) was modeled as the mediator, the other cytokine (IL-8 or IL-6, respectively) was not adjusted for, to avoid overadjustment due to their biological coupling within the same inflammatory cascade. Other covariates were included for confounding control. Multicollinearity was assessed and confirmed absent (GVIF^(1/(2*Df))< 5). All tests were two-tailed with a significance level of p< 0.05. Statistical analyses were conducted using R (version 4.3.3) and Stata. R analyses were performed with the stats, car, MASS, logistf, survival, survminer, interactionR, and mediation packages. GBTM was performed using the traj package in Stata. All other visualizations were conducted using R version 4.3.3 or OriginLab 2024.

Among the 108 patients initially screened, 80 met all inclusion and exclusion criteria and provided written informed consent. After follow-up, 23 patients from each group were included in the final analysis (Figure 1). The final cohort comprised predominantly male patients (69.6%), with a median age of 55 years. Common comorbidities and risk factors included hypertension (21.7%), diabetes mellitus (10.1%), and heavy smoking (26.1%). All patients were diagnosed with stage III–IV NPC, most with poorly differentiated histology (92.8%), and all received 2–4 cycles of induction chemotherapy prior to radiotherapy. Baseline characteristics were well balanced across the three groups (Figure 1; Table 1).

The overall incidence of OM was 95.7% (66/69). SOM occurred in 42.4% of patients. OM onset occurred at a median of 11th radiotherapy fraction, and the median duration was 71 days. OM increased rapidly during radiotherapy, with 87.0% occurring within the first three weeks of treatment, followed by prompt resolution after radiotherapy completion. All cases of SOM resolved within one week after radiotherapy, and 76.8% achieved complete mucosal recovery at the six-month follow-up. Radiotherapy adherence was high, with 91.7% of patients completing treatment as scheduled. Brief interruptions (<3 days) were rare and did not materially affect the analysis.