This retrospective cohort study was conducted at Xingtai People's Hospital, a tertiary general hospital in Hebei Province, China, using routinely collected data from the CHPS. The study period covered all consecutive adult patients who received at least one course of systemic antineoplastic therapy between January 2020 and December 2024 and met the predefined inclusion and exclusion criteria. As a single-center analysis based on one institutional CHPS implementation, the study was designed primarily as a hypothesis-generating evaluation of active monitoring vs. spontaneous reporting and not as a definitive multi-center validation of pharmacovigilance performance.

No random sampling or convenience sampling was conducted; the final sample of 500 patients reflects the entire eligible population during the study period. This approach minimizes selection bias and ensures that the cohort is representative of real-world oncology practice at our institution.

The study protocol was reviewed and approved by the Ethics Committee of Xingtai People's Hospital, and all procedures were carried out in accordance with the Declaration of Helsinki and relevant national regulations on biomedical research ethics.

Patients were eligible for inclusion if they met the following criteria: (1) Age ≥18 years at the time of diagnosis or treatment; (2) Received at least one course of antineoplastic therapy, including chemotherapy, targeted therapy, or immune checkpoint inhibitors; (3) Had complete electronic medical records and medication records available in the Chinese Hospital Pharmacovigilance System (CHPS).

Patients were excluded if they met any of the following conditions: (1) Incomplete or severely missing clinical information, including missing baseline demographics, treatment details, or outcome data; (2) Use of non-oncologic medications as the primary therapy, without clear documentation of antineoplastic drug exposure; (3) Lack of follow-up information, making it impossible to assess ADR occurrence, severity, or clinical outcomes.

All data were obtained from the Chinese Hospital Pharmacovigilance System (CHPS), which integrates multi-dimensional information from electronic health records. The dataset included:

Adverse drug reactions (ADRs) were defined according to the criteria of the World Health Organization (WHO) and the National Adverse Drug Reaction Monitoring Center (NADRMC) of China. Severe ADRs were classified as events of grade 3 or higher based on the Common Terminology Criteria for Adverse Events (CTCAE), or those resulting in hospitalization or death. ADRs were categorized by system organ class (SOC), including but not limited to hematologic disorders, gastrointestinal events, skin and mucosal reactions, immune-related toxicities, and hepatic or renal impairment. Clinical outcomes were graded as complete recovery, survival with sequelae, or death.

All suspected ADR events identified through CHPS triggers or clinical documentation were independently reviewed by two trained clinical pharmacists who applied standardized CTCAE criteria to assign severity grades. Any discrepancies were adjudicated by a senior oncologist to ensure consistency. To evaluate inter-rater reliability, a random 20% subsample of ADR cases was reassessed, yielding a Cohen's κ coefficient of 0.82, indicating strong agreement. This multi-reviewer process was implemented to minimize subjective variability and enhance the robustness of ADR classification.

Two monitoring approaches were applied in this study. Active monitoring was conducted through the CHPS platform, which automatically captured abnormal prescription records, laboratory test results, and imaging findings. These data were further reviewed and confirmed by clinical physicians, enabling near real-time identification and early warning of ADRs. In contrast, spontaneous reporting relied on voluntary submissions by healthcare providers or patients. Reports were typically made only for severe events, which resulted in well-recognized issues such as underreporting and delayed reporting, particularly for mild and moderate ADRs.

To evaluate variable-level completeness within CHPS before analysis. Demographic and medication-exposure variables had <2% missingness, whereas selected laboratory parameters (ALT, AST, eGFR) showed 5%–12% missingness. Based on the observed patterns and clinical workflow, missingness was considered most compatible with a missing-at-random (MAR) mechanism. Variables with <5% missingness were analyzed using complete-case analysis. For variables with ≥5% missingness, multiple imputation by chained equations (m = 5) was performed, incorporating demographics, comorbidities, treatment regimen, and ADR status as predictors. Results from the imputed datasets were pooled using Rubin's rules, and sensitivity analyses confirmed consistency with complete-case findings.

Given that CHPS is based on routinely collected clinical data, several procedures were implemented to reduce potential ADR misclassification. All candidate ADR events—identified via abnormal laboratory values, medication changes, or clinician-entered notes—were independently reviewed by two trained clinical pharmacists using standardized CTCAE criteria, with disagreements resolved by a senior oncologist. Event timestamps were cross-checked against medication administration records to verify temporal plausibility. ADR categories were assigned according to a predefined SOC-based taxonomy to minimize subjective variation. Despite these safeguards, some degree of misclassification remains possible due to variability in routine clinical documentation and the inherent limitations of retrospective EHR extraction.

All statistical analyses were performed using SPSS software and R software. A two-sided P value of <0.05 was considered statistically significant. Missing data were handled by either multiple imputation or listwise deletion, depending on the extent and distribution of missingness. Signal detection was performed using three widely adopted disproportionality methods in pharmacovigilance: the proportional reporting ratio (PRR), the reporting odds ratio (ROR), and the information component (IC). A signal was considered positive if the predefined thresholds were met, namely PRR ≥ 2 with χ² ≥ 4, lower limit of the 95% confidence interval (CI) for ROR > 1, or lower limit of the 95% CI for IC > 0.

Potential risk factors associated with ADRs were first assessed by univariate analysis using the chi-square test for categorical variables and the Student's t test for continuous variables. Variables showing statistical significance in univariate analysis were subsequently entered into a multivariate logistic regression model. Covariates included age, sex, body mass index (BMI), comorbidities, number of concomitant medications, and hepatic/renal function status. Results were reported as odds ratios (ORs) with 95% confidence intervals (CIs).

A predictive model for severe ADRs was developed using multivariate logistic regression. Model performance was evaluated by calculating the area under the receiver operating characteristic (ROC) curve (AUC), as well as sensitivity, specificity, and the Youden index at the optimal cut-off point. Model comparisons were conducted using the DeLong test to assess differences in AUC. To address potential overfitting from using the same dataset for model development and evaluation, we conducted bootstrap internal validation with 1,000 resamples, generating optimism-corrected AUC estimates. Calibration was assessed using calibration plots comparing predicted vs. observed probabilities and the Hosmer–Lemeshow goodness-of-fit test. In addition to the multivariable model, comparative analyses were performed for (1) single-factor models (e.g., age, number of concomitant drugs), (2) the multivariable combined model, and (3) models based on active monitoring vs. spontaneous reporting.

A total of 500 patients were enrolled in this study, including 280 males (56.4%) and 220 females (43.6%). The overall mean age was 58.7 ± 9.6 years, with males being slightly older than females (60.1 ± 9.4 vs. 56.8 ± 9.7 years). Regarding cancer type, lung cancer was the most frequent diagnosis (31.7%), followed by breast cancer (23.5%), gastric cancer (18.5%), and other tumors (26.2%). As expected, breast cancer occurred almost exclusively in females (54.0%), whereas lung cancer (39.8%) and gastric cancer (24.7%) were more common among males. With respect to treatment regimens, chemotherapy was the predominant modality (40.4%), followed by targeted therapy (28.2%), immunotherapy (19.6%), and combination regimens (12.4%). Chemotherapy and targeted therapy were more frequently administered in male patients, while immunotherapy and combination therapies were relatively more common among females. Detailed baseline characteristics are presented in Table 1.

Among the 500 patients included in the analysis, a total of 185 adverse drug reactions (ADRs) were identified, corresponding to an overall incidence of 37.00% (185/500). Of these, 52 cases (28.11%) were classified as severe ADRs (CTCAE grade ≥ 3). When comparing different surveillance strategies, active monitoring captured 160 cases, whereas spontaneous reporting identified only 50 cases during the same period. The distribution of ADR incidence and severity levels is illustrated in Figure 1.

Among the 185 ADR cases, hematologic disorders were the most frequently reported (29.73%), predominantly presenting as leukopenia, anemia, and thrombocytopenia. Gastrointestinal reactions ranked second (25.95%), including nausea, vomiting, diarrhea, and mucositis. Skin and mucosal reactions accounted for 19.46%, with rash, pruritus, and stomatitis being the most common presentations.

Immune-related events represented 12.97% of all ADRs and were primarily associated with PD-1/PD-L1 inhibitors, manifesting as thyroid dysfunction, hepatitis, and pneumonitis. Hepatic and renal impairments were observed in 8.11% of cases, most commonly drug-induced hepatotoxicity and transient renal function decline. Other less frequent manifestations (3.78%) included infusion-related reactions and cardiovascular adverse events. The proportional distribution of ADR categories is depicted in Figure 2.

Among the 185 reported ADRs, 52 cases (28.11%) were classified as severe (CTCAE grade ≥ 3). Of these, 18 patients (34.62%) required hospitalization, 12 patients (23.08%) discontinued treatment, and 5 patients (9.62%) died due to ADR-related complications. Regarding clinical outcomes, 31 patients (59.62%) fully recovered after appropriate management, 16 patients (30.77%) improved but remained with residual symptoms, and 5 patients (9.62%) ultimately died. Detailed outcomes of severe ADRs are presented in Table 2.

Univariate analysis identified several factors significantly associated with ADRs, including age ≥65 years, female sex, concomitant use of ≥3 medications, hepatic/renal dysfunction, BMI ≥25 kg/m², history of cardiovascular disease, and drug exposure ≥14 days (all P < 0.05). Multivariate logistic regression confirmed that age ≥65 years (OR = 1.84, 95% CI: 1.13–2.99, P = 0.015), concomitant use of ≥3 medications (OR = 2.27, 95% CI: 1.39–3.71, P = 0.001), hepatic/renal dysfunction (OR = 1.90, 95% CI: 1.07–3.38, P = 0.028), and drug exposure ≥14 days (OR = 1.76, 95% CI: 1.05–2.94, P = 0.031) were independent risk factors for ADRs. Female sex, BMI ≥25 kg/m², and cardiovascular comorbidity showed increased risk in univariate analysis but were not statistically significant after adjustment (Table 3). Detailed estimates for both univariate analyses and the expanded multivariate model are provided in Supplementary Table S1.

Disproportionality analyses were performed using proportional reporting ratio (PRR), reporting odds ratio (ROR), and information component (IC). Several conventional signals, consistent with current drug labeling, were detected, including hematologic toxicities associated with platinum compounds and gastrointestinal adverse events associated with taxanes. PD-1/PD-L1 inhibitors demonstrated a potential association with skin hyperpigmentation, which has not been comprehensively documented in existing product information. In addition, tyrosine kinase inhibitors (TKIs) showed a signal for cardiotoxicity, specifically QT prolongation and arrhythmia, suggesting a need for enhanced monitoring in clinical practice. The detailed results of signal detection are presented in Table 4, and the relative strength of drug–ADR associations is illustrated in Figure 3.

A comparative analysis demonstrated that active monitoring substantially outperformed spontaneous reporting in detecting ADRs. Active monitoring identified 160 cases, whereas spontaneous reporting captured only 50 cases during the same observation period. Moreover, active monitoring achieved a wider coverage of ADR types, including mild and moderate events that were frequently underreported in spontaneous systems.

On average, active monitoring detected ADRs 6.5 days earlier than spontaneous reporting, highlighting its practical value for timely pharmacovigilance and clinical intervention. Detailed comparisons are presented in Table 5, and the differences in ADR detection between the two methods are illustrated in Figure 4.

To evaluate the predictive value for severe ADRs, receiver operating characteristic (ROC) curve analyses were performed. The multivariable logistic regression model demonstrated good discriminatory ability, with an AUC of 0.82 (95% CI: 0.75–0.88). At the optimal cut-off, the sensitivity and specificity were 78.0% and 74.5%, respectively. After bootstrap internal validation (1,000 iterations), the optimism-corrected AUCs were 0.82 and 0.80, respectively, indicating limited overfitting. Calibration analysis demonstrated good agreement between predicted and observed risks. The calibration plot showed no systematic deviation across risk strata, and the Hosmer–Lemeshow test was non-significant (P = 0.41), indicating acceptable model fit. Calibration slope, intercept, Brier score, and optimism-corrected AUC metrics are summarized in Supplementary Table S2.

Compared with single-variable models (age ≥65 years: AUC = 0.63; concomitant use of ≥3 drugs: AUC = 0.68), the multivariable model showed significantly improved predictive performance (P < 0.05, DeLong test). Furthermore, the active monitoring–based model (AUC = 0.84) outperformed the spontaneous reporting model (AUC = 0.72), underscoring the added value of proactive pharmacovigilance. Detailed model performance metrics are provided in Table 6, and ROC curves are shown in Figure 5.

Variable-level missingness and the full specification of the MICE imputation model, including predictors used for chained equations, are summarized in Supplementary Table S3. To evaluate the influence of additional confounder adjustment, a direct comparison between the original and expanded multivariate logistic regression models is presented in Supplementary Table S4, showing minimal changes in effect estimates and preserved significance for key predictors.