Signal stratification of antineoplastic drugs associated with interstitial lung disease: A multi-method signal detection analysis using the FAERS database

Abstract

This study evaluated the signal of association of interstitial lung disease (ILD) associated with antineoplastic drugs using a multi-method signal detection approach and the FDA Adverse Event Reporting System (FAERS) database. ILD reports linked to antineoplastic drugs from the first quarter of 2004 through the second quarter of 2024 were retrieved from FAERS and analyzed using 5 pharmacovigilance methods: reporting odds ratio (ROR), proportional reporting ratio, MHRA, Multi-item Gamma Poisson Shrinker, and Bayesian Confidence Propagation Neural Network. Drugs were classified according to their primary mechanism of action. A total of 38,950 reports of ILD were analyzed. Of these, 47.66% were in male patients, and 37.57% were aged 65 to 80 years. Geographically, Japan accounted for 43.74% of the reports. Serious outcomes were documented in 99.29% of cases, predominantly hospitalization (51.64%) and death (26.36%). Signal detection analysis identified 1210 potential drug-ILD associations. Among these, 605 demonstrated robust evidence, satisfying the criteria for 3 or more methods. Notably, non-small cell lung cancer was associated with a high burden of ILD, particularly among elderly patients. The strongest signal was observed for trastuzumab deruxtecan, with a ROR of 41.3 (95% CI: 35.8–47.7), followed by gefitinib and lenvatinib. Analysis of treatment duration revealed that immunotherapy agents exhibited the shortest median durations, ranging from 32 to 43 days. Targeted therapies showed a broader range of 24 to 379 days, while chemotherapy drugs displayed the greatest variability, with median durations spanning 4 to 76 days. ILD signal strength varies across antineoplastic drug classes, with targeted therapies and immune checkpoint inhibitors showing heightened signals. Multi-method analysis suggests the potential need for adequate monitoring.

1. Introduction

Antineoplastic agents constitute a cornerstone in the management of malignant tumors, functioning primarily to suppress the proliferation of cancerous cells or induce their apoptosis.^[1,2] Despite their therapeutic efficacy, these agents are frequently associated with a spectrum of adverse reactions.^[3] Among these, interstitial lung disease (ILD) is a serious complication characterized by inflammation and fibrosis of the lung parenchyma, which can lead to impaired gas exchange, respiratory failure, and death.^[4,5]

Clinically, ILD manifests through symptoms such as dyspnea, persistent cough, and fatigue, with severe cases potentially leading to hypoxemia and substantial deterioration of pulmonary function. The association between ILD and specific antineoplastic drugs, such as fluorouracil, capecitabine, and crizotinib, has been documented in case reports and previous pharmacovigilance studies.^[6–10] The underlying mechanisms are diverse and may involve direct cytotoxicity, immune-mediated injury, or pro-fibrotic pathways.^[11,12] In recent years, the widespread adoption of immune checkpoint inhibitors (ICIs) and molecular targeted therapies has further heightened awareness of this adverse event. However, the ILD signal profile and underlying mechanisms appear to vary across different agents.

The FDA Adverse Event Reporting System (FAERS) serves as a comprehensive repository for monitoring and documenting drug-related adverse events, compiling submissions from healthcare providers, pharmaceutical manufacturers, and patients.^[13] Several studies have leveraged FAERS to investigate ILD signal strength associated with specific drug classes, such as tyrosine kinase inhibitors.^[14] However, a comprehensive, multi-method analysis that systematically evaluates and compares ILD signals across all major categories of antineoplastic drugs is currently lacking. Furthermore, a detailed exploration of how these signals vary by key demographic and treatment characteristics remains underexplored.

This study aims to address this gap by performing a multi-method signal detection analysis of the FAERS database to identify and stratify ILD signals associated with a broad spectrum of antineoplastic agents. We further seek to characterize these associations by patient age, sex, and treatment duration. We hypothesize that significant differences in ILD signal strength exist across drug classes and patient subgroups. By providing a holistic pharmacovigilance perspective, this analysis aims to generate hypotheses for further clinical and mechanistic investigation into drug-induced ILD.

2. Methods

2.1. Data sources

This investigation utilized data derived from the FAERS. Reports of ILD associated with commonly used antineoplastic agents were extracted, spanning from the first quarter of 2004 through the second quarter of 2024. Data retrieval was performed using the OpenVigil 2.1 platform, integrated with the Medical Dictionary for Regulatory Activities (MedDRA) version 24 (https://openvigil.sourceforge.net), and the AERSMine platform.^[15] Search parameters encompassed adverse drug reaction (ADR) reporting country, patient gender, age, and outcome measures. Subgroup analyses for ILD were conducted via AERSMine by stratifying data according to age, gender, and additional categorical variables.

2.2. Data processing

ILD-related ADR codes and classifications were identified using preferred terms and system organ class designations within the MedDRA framework. Data extraction was facilitated by the analytical tools OpenVigil 2.1 and AERSMine. OpenVigil 2.1 enabled direct access to structured ADR reports from FAERS via an application programming interface. The search term “ILD” was input into the adverse event field, with the time frame defined from the first quarter of 2004 through the second quarter of 2024, without pre-specifying target drugs, to capture all relevant ILD reports. Similarly, in AERSMine, “ILD” was queried, with demographic filters set to “Adults (18–64)” and “Elderly (65 and above),” and display options configured to list drugs as rows and aggregate by pharmacologic class. The study workflow is detailed in Figure 1.

Figure 1.

Workflow for detecting and analyzing drug-associated interstitial lung disease (ILD) signals based on the FAERS database. FAERS = FDA Adverse Event Reporting System.

Antineoplastic agents were classified into 3 broad categories, cytotoxic chemotherapy, targeted therapy, and ICIs. This was consistent with the structural categorization often employed in the FAERS database and prevalent clinical practice. Furthermore, a refined classification system based on the primary mechanism of action (MOA) was subsequently implemented. This could provide a more precise pharmacological perspective. Agents with multimodal mechanisms were classified according to their dominant MOA to enhance the accuracy of the analysis and mitigate classification bias.

Treatment durations were calculated as the difference between the reported start and stop dates for each drug-ILD case pair in the FAERS datasets. To ensure robustness, the analysis was restricted to cases containing valid entries for both dates.

2.3. Signal detection and statistical analysis

Signal detection was performed using a multi-method approach encompassing both frequentist (proportional reporting ratio (PRR) and reporting odds ratio (ROR)) and Bayesian (Multi-item Gamma Poisson Shrinker (MGPS), Bayesian Confidence Propagation Neural Network (BCPNN)) disproportionality analysis techniques. This strategy was employed to enhance the robustness of our findings by seeking concordance across methods with different statistical underpinnings.^[16,17] The ROR and PRR, recognized globally for their high sensitivity and specificity in post-marketing drug safety evaluations, were primary metrics.^[16] A signal was deemed significant if the adverse event count exceeded 2, the PRR surpassed 2, and the χ² statistic exceeded 4, with higher χ² values indicating increased likelihood of drug-related adverse events. Additionally, 3 supplementary pharmacovigilance methods were employed: the Medicines and Healthcare products Regulatory Agency (MHRA) score, MGPS, and BCPNN. Positive signals were defined by method-specific thresholds: MHRA (absolute count of ILD reports ≥ 3, PRR ≥ 2, χ² ≥ 4); MGPS (Empirical Bayes Geometric Mean 5th percentile [EBGM05] > 2); BCPNN (Information Component lower bound [IC025] > 0 and IC minus 2 standard deviations > 0). A signal was considered robust and worthy of further investigation only if it was detected by 3 or more of these methods, a conservative threshold designed to minimize false positives that may arise from the idiosyncrasies of any single algorithm.

Correlation analyses between absolute ILD event counts and relative risk (based on ROR values) across age and gender subgroups were performed using GraphPad software, applying Spearman rank correlation coefficient for non-normally distributed data. Continuous variables exhibiting non-normal distribution were summarized as medians with interquartile ranges, while normally distributed variables were reported as means with standard deviations; categorical variables were presented as counts with percentages. Statistical analyses were performed using GraphPad software, with chi-square tests applied to assess significance, where P < .05 was considered statistically significant.

2.4. Ethical review

This study was a retrospective, observational pharmacovigilance analysis utilizing data from the FAERS database. The research did not involve any direct contact with or intervention in human subjects. No new personal information or biological samples were collected; all data were derived from the publicly available, anonymized, and de-identified reports within the FAERS database. Therefore, this study did not require review or approval by an Institutional Review Board or research ethics committee.

3. Results

3.1. Demographic characteristics

A total of 38,950 ILD-related ADR reports were identified (Table 1). Males accounted for 47.66% (n = 18,562), females for 40.49% (n = 15,771), with 11.85% (n = 4617) missing gender data. Patients aged 65 to 80 years constituted the largest age group (37.57%, n = 14,632), followed by those aged 51 to 64 years (30.18%, n = 11,757). The median age was 69 years (IQR: 59–76), mean 65.8 ± 16.1 years; 22.08% lacked age information. Mean body weight was 64.7 ± 33.8 kg. Most reports came from healthcare professionals (physicians: 50.49%, n = 19,667; other professionals: 24.79%, n = 9657), while nonprofessional reports made up 12.92% (n = 5032).

Table 1.

Baseline characteristics of the ILD reports identified from FAERS database (2004 Q1–2024 Q2).

Factors		Counts or mean	Percentages (%)
Gender
Female		15,771	40.49
Male		18,562	47.66
Unknown		4617	11.85
Age group
Missing	8600		22.08
Adult (18–64 yr)	11,757		30.18
Elderly (≥65 yr)	18,074		46.41
Age
Mean (SD)	65.80 (16.12)		100
Median (Q1, Q3)	69.00 (59.00, 76.00)		100
Weight
Mean (SD)	64.66 (33.76)		100
Median (Q1, Q3)	62.00 (52.00, 74.00)		100
Occupation
Physician	19,667		50.49
Other health professional	9657		24.79
Consumer or non-health professional	5032		12.92
Pharmacist	2937		7.54
Unspecified	1489		3.82
Lawyer	168		0.43
Outcomes
Non-serious	276		0.71
Serious	38,674		99.29

FAERS = FDA Adverse Event Reporting System, ILD = interstitial lung disease, SD = standard deviation.

3.2. Geographic and temporal distribution

Japan contributed the highest number of reports (43.74%, n = 17,035), followed by the U.S. (14.04%, n = 5470) and France (12.88%, n = 5015), collectively accounting for 70.66% of global cases (Fig. 2A). By continent, Asia (47.89%), Europe (25.55%), and North America (21.95%) were most represented (Fig. 2B). Reports rose steadily from 2004 to 2019 (peak in 2019: 9.30%, n = 3623), with a slight decline thereafter (Fig. 3A).

Figure 2.

Overall Geographic distributions of all ILD reports. (A) Country distribution. (B) Continental distribution. ILD = interstitial lung disease.

Figure 3.

Trends, word cloud, and correlation assessment of drug-associated interstitial lung disease (ILD) based on FAERS data. (A) Temporal trend of ILD reports in the FAERS database from Q1 2004 to Q2 2024, showing a steady increase in the number of cases over time. (B) Distribution of various outcomes of ILD reports in the FAERS database from Q1 2004 to Q2 2024. (C) Scatter plot of absolute report counts versus relative risk (ROR) for ILD, highlighting drugs with both high reporting frequency and high ROR. The solid line indicates the trend, while dotted lines represent confidence bounds. FAERS = FDA Adverse Event Reporting System.

3.3. Clinical outcomes

Serious outcomes were reported in 99.29% (n = 38,674) of cases. Hospitalization (51.64%, n = 19,972) and death (26.36%, n = 10,196) were most frequent, followed by life-threatening events (11.22%, n = 4338) and disability (2.99%, n = 1156). Only 0.71% (n = 276) were non-serious (Table 1, Fig. 3B, Table S1, Supplemental Digital Content, https://links.lww.com/MD/Q917).

3.4. Signal detection and risk assessment

Of the 1210 identified drug-ILD pairs, 789 met thresholds in at least one detection method (e.g., ROR ≥ 3, PRR ≥ 2). A total of 605 pairs met ≥ 3 criteria (including MGPS EBGM05 > 2 and BCPNN IC025 > 0), indicating robust safety signals that warrant further investigation.

3.5. ILD and antineoplastic drug associations

3.5.1. Correlation between absolute count and relative risk

In the full cohort (n = 22,023), an albeit weak correlation was observed between absolute report count and ROR (Spearman R = 0.2310, P = .0034; Fig. 3C). This suggests that while drugs with higher reporting frequencies may tend to have higher disproportionality scores, the strength of this relationship is limited, indicating that relative risk measures capture distinct information from mere report counts.

3.5.2. Analysis of drug-indication-ILD associations

A stratified analysis was performed to elucidate the association between specific drug-indication pairs and the risk of ILD. The detailed cross-tabulation of key antineoplastic agents against their primary indications, including the absolute number of ILD reports and the corresponding Observed/Expected (O/E) ratios, was provided in Table S2 (Supplemental Digital Content, https://links.lww.com/MD/Q917). This analysis revealed heterogeneity in ILD risk across different malignancies for the same drug class. To visually represent the complex interplay, we constructed a Sankey diagram (Fig. S1, Supplemental Digital Content, https://links.lww.com/MD/Q917). This visualization highlighted that the majority of ILD reports were not uniformly distributed but were instead concentrated within a limited number of key indications. Most notably, agents targeting non-small cell lung cancer (NSCLC), spanning multiple drug classes including ICIs, EGFR tyrosine kinase inhibitors, and cytotoxic chemotherapeutic agents, emerged as the predominant contributors to the overall burden of ILD reports.

3.5.3. Outcome analysis by indication for interstitial lung disease in the FAERS database

An analysis of ILD outcomes across all reported therapeutic indications was conducted, categorizing outcomes as death (de), disability (ds), hospitalization (ho), life-threatening (lt), required intervention (ri), or other (ot). The distribution of ILD-related adverse events and their severity varied substantially by indication (Table S1, Supplemental Digital Content, https://links.lww.com/MD/Q917). NSCLC was associated with the highest absolute number of ILD reports (n = 2873). Among these, a significant proportion represented serious outcomes: 1387 deaths, 1614 hospitalizations, and 433 life-threatening events. Additionally, 96 disability cases and 48 cases requiring intervention were reported. For metastatic renal cell carcinoma (mRCC), 678 ILD cases were identified. This indication was associated with 260 deaths, 450 hospitalizations, and 100 life-threatening events. A further 47 disability cases were reported. In malignant melanoma, 335 ILD cases were reported, including 109 deaths, 228 hospitalizations, and 47 life-threatening events. Fourteen disability cases were also documented. Chronic myeloid leukemia was associated with 416 ILD cases. Outcomes included 94 deaths, 271 hospitalizations, and 69 life-threatening events. Eighteen disability cases were reported. Notably, diffuse large B-cell lymphoma presented a distinct pattern among the top indications, with 791 ILD cases characterized by a very high number of “other” outcomes (n = 704) relative to other serious outcomes (58 deaths, 90 hospitalizations). Across all indications, hospitalization was the most frequently reported serious outcome, followed by death. This analysis underscores that the burden and severity of drug-associated ILD are not evenly distributed across cancer types, with NSCLC contributing the largest absolute number of severe ILD outcomes.

3.5.4. Adverse event signal detection by five methods

In a study analyzing 2992 screened drugs, 250 antineoplastic agents were evaluated for signal detection using 5 methods. Of these, 64 agents met the signal threshold in at least one method, with 59 meeting it in 3 or more (Table 2, Fig. 4). Among chemotherapy agents (Fig. 4A), paclitaxel showed strong signals with an ROR of 7.5 and PRR of 7.4, positive across all 5 methods, followed by carboplatin (ROR 6.2, PRR 6.1), cyclophosphamide (ROR 5.1, PRR 5.0), and etoposide (ROR 3.5, PRR 3.5). ICIs demonstrated notable signals (Fig. 4B), with durvalumab leading at ROR 15.8 and PRR 15.2, followed by pembrolizumab (ROR 11.2, PRR 10.9), nivolumab (ROR 10.8), and ipilimumab (ROR 7.4). Targeted therapies showed the highest signals (Fig. 4C), particularly trastuzumab deruxtecan (ROR 41.3, PRR 37.2, IC025 4.3), followed by gefitinib (ROR 16.0, PRR 15.4), osimertinib (ROR 10.6), lenvatinib (ROR 7.3), and nintedanib (ROR 6.8). Atypical cancer-related agents, such as azacitidine (ROR 2.4, IC025 0.8) and elotuzumab (ROR 2.1, IC025 −0.1), exhibited weaker signals.

Table 2.

Adverse event signal detection by five pharmacovigilance methods in antineoplastic drugs.

Drug name	ROR	ROR lower 95% CI	ROR upper 95% CI	PRR	PRR lower 95% CI	PRR upper 95% CI	MHRA (DE, PRR, χ2)	MGPS (EBGM05)	BCPNN (IC025)	Positive signal (≥1 method)	Positive signal (≥3 methods)
Cytotoxic chemotherapy
Alkylating agents
Busulfan	3.1	2.2	4.5	3.1	2.2	2.2	(29, 3.1, 39.4)	2.3	0.9	Yes	Yes
Cyclophosphamide	5.1	4.7	5.5	5	4.7	4.7	(843, 5.0, 2593.6)	4.9	2.1	Yes	Yes
Dacarbazine	8.5	6.5	11.1	8.3	6.4	6.4	(56, 8.3, 352.5)	3.7	2.2	Yes	Yes
Bendamustine	3.7	3	4.6	3.7	3	3	(91, 3.7, 177.1)	3.2	1.4	Yes	Yes
Antimetabolites
Fluorouracil	6.7	6.1	7.2	6.6	6	6	(585, 6.6, 2665.3)	6.2	2.4	Yes	Yes
Gemcitabine	6.8	6.2	7.5	6.7	6.1	6.1	(421, 6.7, 1993.1)	6.1	2.4	Yes	Yes
Tegafur	19.1	15.5	23.4	18.1	14.9	14.9	(97, 18.1, 1550.2)	3.8	3.2	Yes	Yes
Topoisomerase inhibitors
Etoposide	3.5	3.1	4	3.5	3.1	3.1	(220, 3.5, 390.2)	3.3	1.5	Yes	Yes
Topotecan	5.5	3.7	8.2	5.4	3.7	3.7	(25, 5.4, 86.0)	2.5	1.4	Yes	Yes
Irinotecan	6.5	5.8	7.4	6.4	5.7	5.7	(270, 6.4, 1215.5)	5.6	2.3	Yes	Yes
Doxorubicin	6	5.6	6.5	6	5.5	5.5	(631, 6.0, 2514.5)	5.7	2.3	Yes	Yes
Amrubicin	47.5	23.7	95.3	42	21.8	21.8	(9, 42.0, 321.3)	0.2	1.3	Yes	Yes
Microtubule inhibitors
Vincristine	6	5.5	6.7	6	5.4	5.4	(431, 6.0, 1735.6)	5.6	2.3	Yes	Yes
Vinblastine	8.7	6.6	11.4	8.5	6.5	6.5	(55, 8.5, 356.5)	3.6	2.2	Yes	Yes
Vinorelbine	9.7	8	11.7	9.5	7.9	7.9	(111, 9.5, 831.1)	5.2	2.6	Yes	Yes
Vindesine	3.4	1.6	7.1	3.3	1.6	1.6	(7, 3.3, 9.2)	1.3	0.1	Yes	Yes
Paclitaxel	7.5	7	8	7.4	6.9	6.9	(848, 7.4, 4437.1)	7	2.6	Yes	Yes
Docetaxel	6.8	6.2	7.5	6.7	6.2	6.2	(500, 6.7, 2364.0)	6.2	2.5	Yes	Yes
Cabazitaxel	3.2	1.7	6	3.2	1.7	1.7	(10, 3.2, 13.1)	1.6	0.3	Yes	Yes
Platinum agents
Cisplatin	3.9	3.5	4.4	3.9	3.5	3.5	(299, 3.9, 628.8)	3.7	1.7	Yes	Yes
Carboplatin	6.2	5.8	6.7	6.1	5.7	5.7	(748, 6.1, 3081.3)	5.9	2.4	Yes	Yes
Oxaliplatin	6.1	5.6	6.7	6.1	5.6	5.6	(515, 6.1, 2113.1)	5.7	2.3	Yes	Yes
Other cytotoxic
Dactinomycin	3	1.2	7.3	3	1.2	1.2	(5, 3.0, 4.8)	1.1	-0.2	Yes	No
Mitomycin	6.6	4.5	9.7	6.5	4.4	4.4	(26, 6.5, 115.9)	2.4	1.6	Yes	Yes
Bleomycin	10.9	9.1	13.2	10.6	8.8	8.8	(111, 10.6, 956.7)	5.2	2.7	Yes	Yes
Trabectedin	5.9	3	11.9	5.9	2.9	2.9	(8, 5.9, 27.6)	1.1	0.6	Yes	Yes
Pemetrexed	8.8	7.8	9.8	8.6	7.7	7.7	(298, 8.6, 1960.6)	6.9	2.7	Yes	Yes
Estramustine	37.9	25.8	55.6	34.3	23.8	23.8	(29, 34.3, 906.8)	0.7	2.7	Yes	Yes
Targeted therapy
Monoclonal antibodies (non-ICI)
Trastuzumab	4.9	4.4	5.5	4.9	4.3	4.3	(281, 4.9, 844.0)	4.5	2	Yes	Yes
Cetuximab	4.9	4.2	5.7	4.8	4.1	4.1	(159, 4.8, 474.0)	4.2	1.9	Yes	Yes
Bevacizumab	5.2	4.8	5.7	5.1	4.7	4.7	(539, 5.1, 1735.8)	4.9	2.1	Yes	Yes
Rituximab	5.4	5	5.8	5.3	5	5	(819, 5.3, 2740.5)	5.2	2.2	Yes	Yes
pertuzumab	5.9	4.9	7	5.8	4.9	4.9	(127, 5.8, 499.0)	4.6	2.1	Yes	Yes
Obinutuzumab	5.1	3.9	6.7	5.1	3.9	3.9	(55, 5.1, 176.0)	3.5	1.7	Yes	Yes
Panitumumab	11.3	9.5	13.5	11	9.3	9.3	(133, 11.0, 1193.2)	5.7	2.8	Yes	Yes
Small molecule kinase inhibitors
Imatinib	2.2	1.8	2.6	2.2	1.8	1.8	(126, 2.2, 76.6)	2.1	0.8	Yes	Yes
Erlotinib	5.7	4.8	6.7	5.6	4.8	4.8	(149, 5.6, 552.8)	4.6	2	Yes	Yes
Gefitinib	16	13.8	18.6	15.4	13.3	13.3	(176, 15.4, 2332.7)	6.3	3.3	Yes	Yes
Osimertinib	10.6	9.4	11.9	10.3	9.2	9.2	(290, 10.3, 2387.1)	7.5	2.9	Yes	Yes
Sorafenib	2.3	1.8	2.9	2.3	1.8	1.8	(64, 2.3, 45.5)	2.1	0.7	Yes	Yes
Axitinib	2.1	1.6	2.7	2.1	1.6	1.6	(61, 2.1, 32.1)	1.9	0.6	Yes	No
Nintedanib	6.8	5.8	7.8	6.6	5.7	5.7	(178, 6.6, 841.6)	5.3	2.3	Yes	Yes
Lenvatinib	7.3	6.4	8.2	7.1	6.3	6.3	(267, 7.1, 1383.7)	6	2.5	Yes	Yes
Crizotinib	4.8	3.9	6	4.8	3.9	3.9	(85, 4.8, 250.1)	3.8	1.7	Yes	Yes
Alectinib	6.8	5.3	8.8	6.7	5.3	5.3	(63, 6.7, 301.9)	3.9	2	Yes	Yes
Idelalisib	2.2	1.5	3.4	2.2	1.5	1.5	(21, 2.2, 13.1)	1.8	0.4	Yes	No
Abemaciclib	8.7	7.3	10.4	8.5	7.2	7.2	(131, 8.5, 860.2)	5.5	2.5	Yes	Yes
mTOR/ PI3K inhibitors
Everolimus	7	6.2	7.8	6.9	6.2	6.2	(320, 6.9, 1574.6)	6	2.4	Yes	Yes
Temsirolimus	10.5	7.9	13.9	10.2	7.7	7.7	(49, 10.2, 398.1)	3.2	2.3	Yes	Yes
Antibody-drug conjugates (ADCs)
Brentuximab vedotin	4.8	3.4	6.7	4.7	3.4	3.4	(34, 4.7, 96.9)	2.8	1.4	Yes	Yes
Trastuzumab emtansine	6.2	4	9.7	6.1	3.9	3.9	(20, 6.1, 80.6)	2.1	1.3	Yes	Yes
Polatuzumab vedotin	3.4	2.3	5	3.4	2.3	2.3	(27, 3.4, 43.0)	2.4	0.9	Yes	Yes
Trastuzumab deruxtecan	41.3	35.8	47.7	37.2	32.4	32.4	(210, 37.2, 7286.3)	4.1	4.3	Yes	Yes
Other targeted
Eribulin	17	14	20.6	16.3	13.5	13.5	(108, 16.3, 1526.1)	4.5	3.2	Yes	Yes
Elotuzumab	2.1	1.1	4	2.1	1.1	1.1	(9, 2.1, 4.0)	1.4	-0.1	Yes	No
Ramucirumab	19.8	16.3	24	18.8	15.6	15.6	(111, 18.8, 1845.4)	4.2	3.3	Yes	Yes
Immunotherapy
Immune checkpoint inhibitors (ICIs)
Ipilimumab	7.4	6.6	8.3	7.2	6.5	6.5	(303, 7.2, 1599.0)	6.2	2.5	Yes	Yes
Nivolumab	10.8	10.1	11.6	10.5	9.9	9.9	(890, 10.5, 7307.1)	9.5	3.1	Yes	Yes
Pembrolizumab	11.2	10.4	11.9	10.9	10.2	10.2	(979, 10.9, 8290.5)	9.8	3.1	Yes	Yes
Atezolizumab	11.1	10	12.3	10.8	9.7	9.7	(382, 10.8, 3314.6)	8.3	3	Yes	Yes
Durvalumab	15.8	13.6	18.3	15.2	13.1	13.1	(186, 15.2, 2428.4)	6.5	3.3	Yes	Yes
Hormonal therapy
Azacitidine	2.4	1.9	3	2.3	1.9	1.9	(70, 2.3, 52.4)	2.2	0.8	Yes	Yes
Triptorelin	2.5	1.5	4.4	2.5	1.5	1.5	(13, 2.5, 10.4)	1.7	0.3	Yes	No

BCPNN = Bayesian confidence propagation neural network, CI = confidence interval, DE = disproportionality estimate, EBGM05 = empirical Bayes geometric mean (5th percentile), IC025 = lower bound of the 95% credibility interval for the information component, MGPS = Multi-item Gamma Poisson Shrinker, MHRA = Medicines and Healthcare products Regulatory Agency, PRR = proportional reporting ratio, ROR = reporting odds ratio, χ² = Chi-squared statistic.

Figure 4.

Forest plots of reporting odds ratios (RORs) with 95% confidence intervals for drug-associated interstitial lung disease (ILD) in the FAERS database. (A) Top chemotherapy drugs with the highest RORs for ILD among all analyzed medications. (B) Immune checkpoint inhibitors (ICIs) with significant ILD signals, including durvalumab, pembrolizumab, atezolizumab, nivolumab, and ipilimumab. (C) Targeted agents associated with ILD signals, showing a broad range of disproportionality risks.

3.5.5. Analysis of reported treatment duration in FAERS cases

Analysis of reported drug treatment duration data from the FAERS database revealed distinct patterns across various agents (Fig. 5A–D, Table 3). Notably, the analysis revealed substantial variation in reported treatment durations, both within and between drug classes.

Figure 5.

Treatment duration analysis for drug-associated interstitial lung disease (ILD) in the FAERS database. (A) Comparison of median duration across drug types. (B) Drug duration distribution. (C) Quadrant analysis of drug duration and coefficient of variation. (D) Relationship between median duration, number of cases and coefficient of variation across drug types.

Table 3.

Drug treatment duration metrics calculated based on FAERS.

Drugs	Number of cases	Mean	SD	Median	Lower quartile	Upper quartile	IQR	CV
Cytotoxic chemotherapy
Alkylating agents
Busulfan	34	17.34	33.83	4	2	12	10	1.95
Cyclophosphamide	463	82.84	235.83	28	5	70	65	2.85
Dacarbazine	52	139.32	429.61	76	57	120	63	3.08
Bendamustine	91	258.22	1283.95	30	5	106	101	4.97
Antimetabolites
Fluorouracil	508	135.04	350.57	59	8	162	154	2.60
Gemcitabine	505	132.14	448.95	43	11	92	81	3.40
Tegafur	66	63.1	99.1	28	5	82	77	1.57
Pemetrexed	351	149.38	688.25	35	6	86	80	4.61
Topoisomerase inhibitors
Etoposide	150	41.29	46.99	23	4	68	64	1.14
Topotecan	33	195.66	595.29	15	5	66	61	3.04
Irinotecan	266	119.46	381.01	33	8	113	105	3.19
Doxorubicin	336	83.77	245.57	38	5	77	72	2.93
Microtubule inhibitors
Vincristine	165	88.31	281.25	28	6	71	65	3.18
Vinblastine	50	87.55	90.62	76	21	120	99	1.04
Vinorelbine	69	155.34	909.01	37	8	85	77	5.85
Vindesine	8	61.07	49.01	73	6	87	81	0.80
Paclitaxel	809	86.74	230.38	36	8	80	72	2.66
Docetaxel	547	233.83	490.63	49	8	127	119	2.10
Cabazitaxel	34	188.23	392.2	38	1	176	175	2.08
Platinum agents
Cisplatin	195	83.76	314.19	41	5	84	79	3.75
Carboplatin	541	89.4	354.27	29	3	79	76	3.96
Oxaliplatin	472	140.39	278.18	71	14	167	153	1.98
Other cytotoxic
Dactinomycin	7	61	32.75	44	42	86	44	0.54
Mitomycin	19	223.81	551.21	56	9	126	117	2.46
Bleomycin	96	80.3	260.4	47	8	99	91	3.24
Trabectedin	17	58.25	29.35	57	29	86	57	0.50
Epirubicin	112	113	292.77	29	4	84	80	2.59
Amrubicin	5	120.69	181.69	47	14	137	123	1.51
Estramustine	16	297.22	481.78	65	16	319	303	1.62
Targeted therapy
Monoclonal antibodies (non-ICI)
Trastuzumab	281	338.66	667.03	64	11	200	189	1.97
Cetuximab	220	75.69	158.88	25	5	79	74	2.10
Bevacizumab	537	168.79	491.64	68	16	171	155	2.91
Rituximab	550	392.25	974.9	59	14	240	226	2.49
Pertuzumab	144	468.97	687.95	71	10	925	915	1.47
Obinutuzumab	67	236.07	487.69	60	10	239	229	2.07
Panitumumab	249	132.47	259.85	56	15	157	142	1.96
Ramucirumab	115	72.98	105.97	43	9	84	75	1.45
Small molecule kinase inhibitors
Imatinib	129	202.19	387.98	56	15	186	171	1.92
Erlotinib	215	137.19	392.3	29	9	93	84	2.86
Gefitinib	292	114.07	350.34	29	10	71	61	3.07
Osimertinib	336	131.49	270.33	40	11	104	93	2.06
Sorafenib	80	236.77	666.73	26	9	99	90	2.82
Axitinib	39	92.56	171.35	49	14	80	66	1.85
Nintedanib	70	259.93	500.13	53	11	162	151	1.92
Lenvatinib	300	106.4	301.56	48	17	110	93	2.83
Crizotinib	114	54.64	98.69	24	8	50	42	1.81
Alectinib	61	57.63	109.98	25	7	87	80	1.91
Idelalisib	25	195.79	541.44	61	29	145	116	2.77
Abemaciclib	133	249.71	432.53	76	28	313	285	1.73
mTOR/ PI3K inhibitors
Everolimus	397	224.42	878.67	58	20	190	170	3.92
Temsirolimus	81	81.99	219.36	25	8	63	55	2.68
ADCs
Brentuximab vedotin	34	74.13	108.49	28	1	99	98	1.46
Trastuzumab emtansine	56	761.12	744.17	379	64	1661	1597	0.98
Polatuzumab vedotin	24	33.51	33.78	21	5	62	57	1.01
Trastuzumab deruxtecan	204	430.85	751.24	60	1	442	441	1.74
Other targeted
Eribulin	137	140.91	273.04	43	12	112	100	1.94
Elotuzumab	13	428.02	549.49	153	35	630	595	1.28
Immunotherapy
ICIs
Ipilimumab	269	80.14	194.33	29	7	71	64	2.42
Nivolumab	893	106.12	286.09	36	7	98	91	2.70
Pembrolizumab	1126	103.41	332.91	35	7	90	83	3.22
Atezolizumab	349	121.46	317.2	43	9	118	109	2.61
Durvalumab	251	60.91	135.23	32	8	71	63	2.22

ADCs = Antibody-drug conjugates, CV = coefficient of variation, FAERS = FDA Adverse Event Reporting System, ICI = immune checkpoint inhibitor, IQR = interquartile range, SD = standard deviation.

The calculated median treatment durations exhibited a wide overall range, from 4 days (busulfan) to 379 days (trastuzumab emtansine) (Fig. 5A). The longest median duration was reported for trastuzumab emtansine (379 days; 56 cases) (Fig. 5B), consistent with its role as a long-term maintenance therapy for HER2-positive breast cancer. The shortest median duration was reported for busulfan (4 days; 34 cases). This broad spectrum may reflect the diverse clinical applications of these agents, from short-term conditioning regimens to prolonged maintenance therapies. Variability in reported duration, as measured by the coefficient of variation (CV), was also widespread. Among the agents with the highest variability (CV > 3.0) were vinorelbine (CV = 5.85), everolimus (CV = 3.92), carboplatin (CV = 3.96), and pembrolizumab (CV = 3.22), suggesting inconsistent reporting or highly heterogeneous dosing schedules across different treatment contexts. In contrast, a subset of agents exhibited relatively consistent reported durations (CV < 1.5), including polatuzumab vedotin (CV = 1.01), trastuzumab emtansine (CV = 0.98), and trabectedin (CV = 0.50) (Fig. 5C).

When examining median duration by drug class, immunotherapy agents (ipilimumab, nivolumab, atezolizumab, pembrolizumab, durvalumab) generally clustered at the lower end of the duration spectrum (median range: 29–43 days). The patterns for targeted therapies and chemotherapeutic agents were more heterogeneous. For example, among targeted therapies, short-term use was suggested for crizotinib (median 24 days) and osimertinib (median 40 days), while long-term administration was observed for trastuzumab emtansine (median 379 days) and pertuzumab (median 71 days). Similarly, among chemotherapy agents, short-term use was commonly reported for busulfan (median 4 days) and etoposide (median 23 days), while longer durations were noted for oxaliplatin (median 71 days) and dacarbazine (median 76 days) (Fig. 5D).

4. Discussion

This study used the FAERS database to evaluate the signal strength of ILD associated with antineoplastic drugs through a multi-method signal detection approach and drug class-specific stratification. The findings reveal significant variations in ILD risk across therapeutic classes, with targeted therapies and ICIs exhibiting the most pronounced signals. Notably, the antibody-drug conjugate trastuzumab deruxtecan showed the highest ILD signal (ROR 41.3), followed by chemotherapy agents such as amrubicin. These observations suggest a potential need for heightened vigilance and enhanced monitoring, particularly among vulnerable populations such as patients with NSCLC.

Among the 38,950 patients for overall demographic analysis, the proportion of male to female patients was approximately 1.18 times, which did not fully align with literature reports suggesting a higher incidence of ILD in males.^[18] The existence of a gender difference in the incidence of ILD remains debated, with the underlying mechanisms still under investigation.^[18] Of the patients included in this study, the age for 8600 patients (22.08%) was unknown. However, an analysis of the reports with known ages revealed that elderly patients remained the primary source of ILD cases (46.41%), consistent with previous research findings.^[19] Studies have indicated that age is one of the primary risk factors for drug-associated ILD.^[19] After adjusting for cohort effects, the age-standardized prevalence (ASPR) and mortality (ASMR) of ILDs showed a rapid increase in relative risk with age, especially in the >55 age group. The RR for ASPR and ASMR in the 80 to 84 age group was 8.2 times and 33.64 times higher, respectively, compared to the control group.^[19] Thus, it was hypothesized that age may be a potential risk factor for ILD, warranting more attention among physicians and patients, especially for the elderly.

In this study, we evaluated the risk of ILD associated with common antineoplastic drugs. Beyond the traditional cytotoxic chemotherapy agents, it is observed that a higher incidence and association strength of ILD with the widespread use of immunotherapeutic and targeted therapy drugs. ICIs, which are increasingly used in cancer treatments, act by modulating the immune system’s intrinsic regulatory mechanisms, specifically targeting CTLA-4, PD-1, and PD-L1 to enhance antitumor activities.^[20] However, this modulation can also lead to an overactivation of the immune system, resulting in immune-related adverse events (irAEs).^[21] The pathogenesis of ICI-related ILD is hypothesized to involve excessive activation of the immune system and an increase in preexisting autoantibodies.^[21] Recent researches have highlighted its prevalence and associated risk factors, with the prevalence of nivolumab-induced ILD in NSCLC patients being around 4.0% in clinical trials,^[22] a condition that stands in contrast to the higher incidences observed in real-world data, which range from 14.5 to 19%.^[23] Meta-analyses indicate a higher incidence of ILD with PD-1 inhibitors (such as nivolumab and pembrolizumab) compared to PD-L1 inhibitors, with no significant differences identified between nivolumab and pembrolizumab.^[23] In comparison to anti-PD-L1 and anti-CTLA-4 medications, PD-1 inhibitors (including nivolumab, and pembrolizumab) exhibited lower ILD signals, suggesting that the strength of the association with ILD might vary among different types of ICIs. Studies have suggested potential risk factors for ICI-related ILD include preexisting conditions, tumor characteristics, and treatment methodologies.^[23] The interplay between molecular targeted agents and ILD has also been examined,^[2,10] with these drugs also showing signals of association with ILD, a finding in line with our observations. Our analysis revealed that targeted drug-paired ILD cases were the most frequently reported and had strong safety signals, indicating stronger disproportionality signals for ILD associated with these medications. Throughout the data analysis, potential associations were observed between multiple antineoplastic drugs and ILD. Certain commonly used antineoplastic drugs, such as fluorouracil and taxane drugs, were linked with relatively higher frequencies of ILD reports, suggesting a potential association.

While the initial classification of drugs was pragmatic and aligned with common practice, we acknowledge its inherent limitation. Categorizing complex agents into mutually exclusive groups such as chemotherapy, targeted therapy, and ICIs is a simplification that may not fully capture the pharmacological diversity and overlapping mechanisms of modern oncologic therapies. This could potentially introduce a risk of misclassification, particularly for agents with multimodal actions (e.g., antibody-drug conjugates, which possess both targeted and cytotoxic properties). To address this, we conducted a supplemental analysis using a more nuanced, MOA-based classification scheme, which offers a mechanistic lens for interpreting the safety signals. Despite this effort, any categorical system remains an imperfect representation of a drug’s complete biological profile, a factor that should be considered when interpreting the association between drug categories and adverse event profiles.

From the perspective of reporting countries, the top countries with the highest number of ILD reports were Japan, the United States, France, and the United Kingdom, all of which are developed nations. These countries accounted for most of the total number of reports, suggesting a potential correlation between the level of economic development and the degree of attention paid to drug safety. It was also recommended that the severity of ILD reports from other countries should also be considered. However, it should be noticed that the substantial geographic disparity in the source of reports from Japan. This overrepresentation is a well-recognized feature of the FAERS database and other spontaneous reporting systems for certain drug-event pairs, and it is largely attributable to Japan’s robust and mandatory pharmacovigilance framework. While this heightened vigilance is commendable and yields valuable safety data, it introduces a potential for reporting bias that must be acknowledged. Consequently, the absolute frequencies and PRRs observed in our overall analysis should be interpreted with caution, as they may not be fully generalizable to other regions with different medical practices, genetic backgrounds, and reporting cultures. Also, it is important to note that for the majority of significant drugs identified in our study, the disproportionality signals were also consistently present and significant within subsets of data from other major regions, such as the United States and Europe. This consistency suggests that the core drug-ILD associations we identified are not artifacts exclusive to the Japanese reporting landscape but are likely to represent genuine safety signals. Future studies utilizing balanced multinational cohorts or dedicated pharmacoepidemiologic designs are warranted to confirm these findings.

The multi-method approach, encompassing ROR, PRR, MHRA, MGPS, and BCPNN, provided a robust framework for identifying and validating ILD signals,^[24,25] with concordance across 605 drug-ILD associations enhancing signal reliability. Word cloud analyses (data not shown) offered valuable insights by contrasting absolute report frequencies with relative risk profiles. For instance, nivolumab and pembrolizumab were frequently reported due to their widespread use, whereas lenvatinib and afatinib stood out for their elevated relative risks.^[26] This dichotomy suggests that clinicians should evaluate drug risk-benefit profiles by integrating usage patterns with intrinsic toxicity, prioritizing monitoring for less commonly used agents.^[27,28]

Our stratified analysis, incorporating both a detailed cross-tabulation (Table S2, Supplemental Digital Content, https://links.lww.com/MD/Q917) and a Sankey diagram (Fig. S1, Supplemental Digital Content, https://links.lww.com/MD/Q917), provides a nuanced understanding of ILD risk that extends beyond a drug-centric view. This approach observed that the risk of drug-associated ILD was intrinsically linked to the specific treatment context, rather than being a uniform class-effect across all indications. The concentration of ILD reports in patients treated for NSCLC, observed across diverse drug classes, suggests a pivotal role for the underlying pulmonary environment and disease-specific biology in the pathogenesis of this adverse event. Patients with NSCLC may possess a preexisting susceptibility to lung injury due to their primary disease, which could be exacerbated by various antitumor therapies. Furthermore, the variation in O/E ratios for the same drug across different indications emphasizes the importance of indication-specific pharmacovigilance.

The findings from treatment duration analysis suggest tailored clinical approaches based on drug category and variability. For immunotherapy and targeted therapy drugs with lower CVs, standardized treatment protocols may be feasible. However, chemotherapy drugs, particularly those with high CVs like Vinorelbine, necessitate enhanced monitoring to ensure safety. The outlier case of Trastuzumab emtansine supports its consideration for extended maintenance therapy, though further validation with larger datasets is recommended to confirm these trends across diverse patient populations.

This study has several limitations. First, FAERS is a spontaneous reporting system, which relies on voluntary reports from doctors, pharmacists, patients, and others. Therefore, the reported incidence of some adverse events might not truly reflect their actual frequency. Second, the varying strategies for using antineoplastic drugs, along with the heterogeneity of patient populations and diseases, could influence the outcomes. Additionally, the lack of detailed demographic and clinical data, such as smoking history or preexisting lung conditions, constrained adjustments for confounding factors. Third, there exists a potential bias related to the timing of reports. The reporting moment might not align perfectly with the actual occurrence of the event. The observational nature of the study precludes causal inference, positioning these findings as hypotheses for further investigation rather than definitive evidence. Fourth, our analysis of treatment duration is constrained by the completeness and accuracy of the start date and end date fields in FAERS. These data are spontaneously reported and often missing or imprecise. Consequently, the calculated durations should be interpreted as indicative of relative patterns rather than precise estimates of clinical exposure. Furthermore, the absence of total drug utilization data (e.g., prescription volumes or patient-years) prevents the calculation of incidence rates and limits our ability to fully contextualize the volume of ILD reports.

5. Conclusion

In summary, this study identifies significant disparities in ILD signals across antineoplastic drug classes, with targeted therapies and ICIs posing the greatest concern. The integration of multi-method signal detection and class-specific analysis offers a reliable approach to identifying agents, while demographic analysis highlights vulnerable populations such as NSCLC patients. These findings highlight the need for further research into tailored monitoring and risk management strategies in clinical practice. Future research should focus on elucidating the molecular mechanisms of ILD, validating signals in prospective cohorts, and investigating pharmacogenetic factors that may predispose specific patients to this adverse event.

Author contributions

Conceptualization: Zilong Gong, Ying Huang, Huilin Xu.

Data curation: Zilong Gong, Qiong Wang, Huilin Xu.

Formal analysis: Zilong Gong, Qiong Wang, Ying Huang, Huilin Xu.

Investigation: Zilong Gong.

Methodology: Zilong Gong.

Validation: Zilong Gong, Qiong Wang, Ying Huang, Huilin Xu.

Writing – original draft: Zilong Gong, Qiong Wang, Ying Huang, Huilin Xu.

Writing – review & editing: Zilong Gong, Qiong Wang, Ying Huang, Huilin Xu.

Supplementary Material

medi-104-e46325-s001.docx ^{(231.6KB, docx)}