Immune-related adverse events of neoadjuvant immunotherapy in patients with perioperative cancer: a machine-learning-driven, decade-long informatics investigation

Abstract

Research on neoadjuvant immunotherapy (NAI) is increasingly focusing on immunotherapy-related adverse events (AEs). However, many unknowns remain in this field. Hence, through the machine learning (ML)-driven informatics analysis, this study aimed to profile the global decade-long scientific landscape of AEs of NAI and further reveal its critical issues and directions that deserve deeper exploration. During the past decade, the amount of research in the field of NAI safety has displayed a positive trend (annual growth rate: 30.2%), and it has achieved good global collaboration (international coauthorship: 17.43%). Using an unsupervised clustering algorithm, we identified six dominant research clusters, among which Cluster 1 (standardizing response assessment criteria for NAI to minimize its adverse reactions; average citation=34.86±95.48) had the highest impact and Cluster 6 (efficacy and safety of multiple therapy patterns combination) was an emerging research cluster (temporal central tendency=2022.43, research effort dispersion=0.52), with “irAEs” (s=0.4242 (95% CI: 0.01142 to 0.8371), R²=0.4125, p=0.0453), “ICIs” (immune checkpoint inhibitors) (s=1.127 (95% CI: 0.5403 to 1.714), R²=0.7103, p=0.0022), and “efficacy and safety” (s=0.5455 (95% CI: 0.1145 to 0.9764), R²=0.5157, p=0.0193) showing significant overall growth. More importantly, further hotspot burst analysis indicated “ICI” and “efficacy and safety” as the emerging research focuses, demonstrating that scholars in the field are increasingly aware of the importance of balancing NAI efficacy and safety. In conclusion, this study presents ML-derived evidence that outlines the safety challenges of NAI and highlights the importance of balancing its efficacy and safety for its application in patients with perioperative cancer.

Introduction

Neoadjuvant immunotherapy (NAI) primarily employs the immune checkpoint inhibitors (ICIs) programmed cell death protein-1 (PD-1)/programmed death-ligand 1 (PD-L1) and CTLA-4, alone or in combination, to prime the immune system before surgical intervention in patients with malignancies, which improves surgical outcomes and long-term survival.^{1 2} Nonetheless, this therapeutic modality is associated with high-grade adverse events (AEs), including acute and chronic immune-related AEs (irAEs), which can affect overall survival and treatment progress; thus, irAEs deserve careful clinical attention.^{3 4}

Researchers in this field are paying increasing attention to irAEs. A preliminary study revealed the importance of balancing the efficacy and safety of NAIs in patients with perioperative cancer, which was also emphasized in a recent meta-analysis.^{5 6} However, the current global research hotspots for NAI-related AEs in patients with perioperative cancer and the critical issues and future directions worthy of in-depth exploration remain unknown. Such issues require urgent attention and need to be addressed in current practice.

Over the past decade, exponential accumulation of unstructured scientific knowledge and data has occurred in this field, which continues to expand. This poses significant challenges for clinicians and researchers to comprehensively understand the research advancements in the field within a limited time frame; elucidate the structural framework, intrinsic relationships, and evolutionary patterns of scientific knowledge; and subsequently identify critical directions that require further exploration. Machine learning (ML)-based informatics approaches provide an important avenue to address these challenges.

Therefore, through ML-driven informatics analyses, this study aimed to profile the scientific knowledge obtained worldwide over the last decade regarding NAI clinical safety management and research, reveal global research hotspots and development trends, and further identify critical issues and directions worthy of in-depth exploration.

Data source and collection

Using the medical information retrieval and storage database Web of Science, in this study, we systematically collected literature related to the AEs of NAI in perioperative oncology by defining search terms and time frames. The search terms used are listed in online supplemental material 1. The time frame was set from January 1, 2014 to May 10, 2024. For quality control, non-English articles were excluded. The data were exported on May 10, 2024. The data export format is named “Plain text file.”

Unsupervised hierarchical clustering

Unsupervised clustering (UC) methods have a unique advantage in exploratory data analysis because they do not rely on data prelabeling, particularly for data analysis scenarios with unclear category labels. The literature often lacks a clear classification system in the field of research hotspot identification. UC can automatically identify potential structures and distribution patterns based on the intrinsic characteristics of the data, making it an ideal tool for bibliometric research. This method can effectively identify spontaneous research topic clusters in academic literature by discovering naturally formed aggregation features in the data. Under the research topic of NAI AEs, we constructed a dataset containing 834 keywords (each keyword represented a research topic) based on the literature in the field. The co-occurrence frequency was used in this study as a measure of the link strength (LS) between topics, where the total LS of a single topic is the cumulative value of its LS with all other topics in the dataset. Based on this, we constructed co-occurrence matrices to characterize the network relationships between topics, with the matrix element values corresponding to the LSs of the topic pairs. Finally, the community discovery algorithm Louvain was applied to cluster the co-occurrence matrix and aggregate research topics with similar co-occurrence characteristics into the same category.

Time-series analysis

Based on the unsupervised hierarchical clustering (UHC) results, an additional temporal dimension of data was overlaid on all research topics to demonstrate their temporal evolution patterns. Specifically, the average publication year for each research topic was obtained by averaging the publication time of all literature on each research topic. Then, the color mapping range was adjusted using the average publication year distribution of all research topics, with purple mapping of earlier research topics and yellow mapping of later research topics. Furthermore, the mean and SD of the average publication year for all the research topics in the same cluster were calculated to obtain the temporal central tendency (TCT) and degree of dispersion for each cluster. Lower TCTs represent early or classical research directions, whereas higher tendencies indicate emerging research directions. A lower degree of dispersion indicates more intensive research effort in the research direction.

Spatial network density analysis

Based on the UHC analysis, spatial network density visualization was subsequently used to present the total linkage strength and frequency of the research topics. The occurrence frequency is the total number of times a research topic appears in a field. Visual maps were obtained by mapping the corresponding values at different color depths.

Fitted curve analysis

Fitted curve analysis, a common statistical modeling technique, reveals the correlation characteristics between variables. This study explored the annual distribution patterns of specific research topics and their temporal evolutionary trends using mathematical modeling. Based on the bibliometric approach, we used the R package “bibliometrix” to perform text mining of all literature in the field and extracted a total of 391 academic keywords. When the number of time points with non-zero data was less than two, the curve analysis failed to fit; therefore, keywords with this condition were excluded. According to the results of the word frequency statistics, 79 core keywords were filtered out for statistical analysis. We constructed a linear regression model to explore the characteristics of the temporal distribution of keywords, the parameters of which were as follows: the significance level α=0.05, the coefficient of determination “R²” reflected the explanatory power of the model, and the regression coefficient “s” characterized the rate of temporal change. The effect of time on keyword frequency was considered statistically significant at p<0.05.

Hotspot burst analysis

Hotspot burst analysis was employed to identify research topics in the field that have received significant attention over a specific period and to quantify the intensity and duration of their bursts. This method is effective for revealing the temporal evolutionary trends of research topics within a field and helps scholars identify emerging topics with potential research value. Based on the analysis framework of the R language package “bibliometrix,” we analyzed hotspot bursts for 391 keywords across the domain, with the following parameter settings: a minimum word frequency of 2 and a minimum duration of 2 years. The R package “bibliometrix” extracts keywords from the literature and counts the occurrence frequency of each keyword during different years to create a “word frequency - year” matrix. In this process, the algorithm’s threshold automatically filters out low-frequency words and non-continuously occurring words. Furthermore, the algorithm quantifies the annual popularity changes of specific keywords to identify keywords with short-period frequency bursts and calculates their intensity and duration. Afterward, we visualized the results using the R package “ggplot2.”

Software and statistical analysis

These analyses were performed using the Java runtime environment and R. Scientific knowledge mapping was visualized using VOSviewer V.1.6.18 (0) and fitted curve plotting was performed using GraphPad Prism V.9.0.7,9 The R toolkits “bibliometrix” and “ggplot2” were used for data processing, statistics, and visualization.10,13 The measured data are presented as means, SD, and 95% CIs. The statistical test level was set at α=0.05, with a p value of less than 0.05 indicating statistical significance.

Results

Overview of global research on adverse events associated with neoadjuvant immunotherapy in patients with perioperative cancer

During the past decade, the field of NAI safety has displayed a positive development trend (annual growth rate: 30.2%) and has achieved good global collaboration (international coauthorship: 17.43%) (online supplemental material 2).

Spatiotemporal patterns of global research hotspots on adverse events related to neoadjuvant immunotherapy in patients with perioperative cancer

The hierarchical clustering analysis based on unsupervised ML revealed that the focuses of studies on NAI efficacy and safety can be divided into six dominant research clusters: (1) standardizing response assessment criteria for NAI to minimize its adverse reactions; (2) pathological response of neoadjuvant radiotherapy, chemotherapy, and immunotherapy; (3) immune checkpoint inhibitors, tumor-infiltrating lymphocytes, irAEs and patient outcomes; (4) biomarkers to predict efficacy and safety of preoperative immunotherapy; (5) management of irAEs in neoadjuvant radiotherapy; (6) efficacy and safety of multiple therapy pattern combination (figure 1A, online supplemental material 3). Time-series analysis further revealed the TCT and research effort dispersion (RED) of the different clusters: Cluster 1, TCT=2022.41 and RED=0.66; Cluster 2, TCT=2021.80 and RED=0.75; Cluster 3, TCT=2021.29 and RED=1.29; Cluster 4, TCT=2022.09 and RED=0.72; Cluster 5, TCT=2022.20 and RED=0.76; and Cluster 6, TCT=2022.43 and RED=0.52 (figure 1B). Among the clusters, Cluster 6 had the largest TCT and smallest RED, demonstrating that Cluster 6 was the latest research domain, and its research effort was the most intensive. This indicates that scholars in this field are gradually beginning to realize the importance of balancing the efficacy and safety of NAI in combination with other therapies and are devoting intensive research efforts to it accordingly. Spatial density networks based on total linkage strength and occurrence frequency provided an intuitive visualization overview of NAI safety research (figure 1C and D).

Figure 1. Spatiotemporal patterns of global research hotspots on adverse events related to neoadjuvant immunotherapy in patients with perioperative cancer. (A) Through an unsupervised hierarchical clustering algorithm, six dominant research clusters were identified. (B) Time series analysis revealed global temporal distribution patterns. The color mapping range was adjusted using the average publication year distribution of all research topics, with purple mapping of earlier research topics and yellow mapping of later research topics. (C) Spatial visualization presentation of total linkage intensity. (D) Spatial visualization presentation of occurrence frequency. irAEs, immune-related adverse events; PD-1, programmed cell death protein-1.

Development trend analysis and future direction prediction for research on adverse events related to neoadjuvant immunotherapy in patients with perioperative cancer

The following regression curve analyses showed that the numbers of publications associated with “ICI” (s=1.127 (95% CI: 0.5403 to 1.714), R²=0.7103, p=0.0022), “irAEs” (s=0.4242 (95% CI: 0.01142 to 0.8371), R²=0.4125, p=0.0453), “safety” (s=0.6121 (95% CI: 0.1951 to 1.029), R²=0.5888, p=0.0096), and “efficacy and safety” (s=0.5455 (95% CI: 0.1145 to 0.9764), R²=0.5157, p=0.0193) are significantly increasing (figure 2A). More importantly, based on the temporal distribution and burst intensity, the hotspot burst analysis identified “efficacy and safety” as an emerging burst hotspot, indicating that scholars in this field have gradually begun to realize the importance of balancing the efficacy and safety of NAI and are investing their research efforts to it (figure 2B).

Figure 2. Development trend analysis and future direction prediction for research on adverse events related to neoadjuvant immunotherapy in patients with perioperative cancer. (A) Regression curve analysis of global research hotspots on adverse events related to neoadjuvant immunotherapy. The coefficient of determination “R²” reflected the explanatory power of the model, and the regression coefficient “s” characterized the rate of temporal change. The effect of time on keyword frequency was considered statistically significant at p<0.05. (B) Hotspot burst analysis on adverse events related to neoadjuvant immunotherapy. PD-1, programmed cell death protein-1; PD-L1, programmed death-ligand 1.

Discussion

NAI have garnered exponential worldwide attention since 2014, with numerous randomized clinical trials demonstrating their substantial clinical contributions, notably achieving significantly higher tumor reduction and pathological complete response (pCR) rates in patients with locally advanced malignancies before surgery. Despite a decade of clinical investigations, not all patients respond satisfactorily to NAI.^{14 15} Moreover, an increasing body of evidence-based systematic reviews, meta-analyses, and real-world observations has identified several safety challenges that could pose significant obstacles to the future application of NAI.^{5 6 16} Therefore, through the ML-driven informatics analysis, this study profiles the global decade-long scientific landscape of AEs of NAI and further reveals its critical issues and directions that deserve deeper exploration.

Although irAEs appeared in both Clusters 3 and 5, it is noteworthy that no specific type of irAEs (eg, rash, endocrine-associated irAEs, gastrointestinal irAEs, immune hepatitis, immune cardiomyopathy, immune pneumonitis, and immune nephritis) appeared in these two clusters or even in the other clusters. This suggests that, although scholars in this field are increasingly focusing on irAEs, few studies have focused on the specific types of irAEs associated with NAI. A potential reason for this phenomenon may be that clinical studies tend to use efficacy (eg, pCR and event-free survival (EFS)) as the primary endpoint and overall safety as the secondary endpoint, without a specific type of irAE as a clear research focus, which would lead to relatively limited research results on specific irAEs. Therefore, current clinical studies related to irAEs are still in the early stages of holistic exploration, mostly focusing on the overall incidence and general management strategies, and have not yet been refined to in-depth exploration of specific irAEs. This indicates that, while the field has received increased attention at the macro level, there is a significant lack of research on the differences in clinical manifestations, specific mechanisms, and precise management strategies.

In the future, the academic community is expected to take an active interest in and promote specialized research on specific types of irAEs. For example, biomarker discovery and prediction tool development for specific types of irAEs, mechanistic studies of specific types of irAEs to help clinics better understand and manage specific immunotherapy toxicities, and studies on individualized clinical management strategies based on specific types of irAEs to promote precision and standardization of the management of irAEs are needed. However, it is also worth noting that different types of irAEs exhibit complex clinical heterogeneity in different cancer types, different treatment modalities, and different ICIs, which poses a challenge for conducting in-depth specialized studies.

Regarding the analysis of keyword trends, it is worth discussing further that the identification of hotspot keywords is not directly equivalent to the ranking of the clinical importance of a drug but rather reflects research growth rates and emerging trends of interest to academics in a given period of research. For example, PD-1 inhibitors such as tislelizumab and sintilimab, developed in China, showed an upward trend in the development trend analysis, reflecting that, in recent years, the number of NAI-related clinical trials has increased dramatically, and their publication numbers in international journals have also risen significantly. Thus, they were identified as hotspot keywords in statistical modeling based on word-frequency growth trends. However, this does not mean that mainstream international PD-1 drugs, such as nivolumab or pembrolizumab, are less important. In fact, further analysis identified that the high-impact literature on these international mainstream drugs was published relatively early and showed a relatively smooth temporal trend, thus failing to achieve the criterion of time-growth significance of our analytical method (p<0.05) and ultimately being excluded from the development trend significance results.

These results provide potential insights for subsequent clinical practice, study design, and guideline development. Some PD-1 ICIs developed in China, such as tislelizumab and sintilimab, have shown increasing popularity, suggesting that the Asia-Pacific region, especially China, is gradually becoming a publication-intensive area in NAI clinical trials, which may influence the design focus, participant source, and drug accessibility strategy of international multicenter trials in the future. Practical differences resulting from regional imbalances in drug development may need to be considered in the development of future clinical guidelines or expert consensuses. In addition, the traditional PD-1/PD-L1 drugs, nivolumab and pembrolizumab, although not presented as significantly rising keywords in the results, their extensive and high-quality early literature remains an evidence-based foundation for current NAI treatments, and their clinical dominance remains irreplaceable. Therefore, during guideline development and standardization, unilateral interpretation of indicators such as research popularity should be avoided, and more emphasis should be placed on high-level evidence accumulated over time, which is of substantial significance in guaranteeing treatment safety and efficacy.

The results of the UHC, time-series, and hotspot burst analyses point in unison to the fact that scholars in this field are gradually beginning to realize the importance of balancing the efficacy and safety of NAI and devoting intensive research efforts to it. This study provides potential insights into subsequent clinical practice, study designs, and guideline development.

Clinically, there has been a greater tendency to maximize efficacy in the past, and the high efficacy of NAI has been widely recognized to some extent.^{1 14 15} In this study, we found that the awareness and efforts to emphasize both efficacy and safety are increasing significantly in current academia, suggesting that current clinical practice is shifting from “efficacy first” to “efficacy-safety balance”. Moreover, this indicates that clinicians should pay attention to the need for personalization of treatment regimens when making subsequent clinical decisions, preferring low-toxicity immunotherapy regimens for patients at high risk of irAEs and considering combination regimens to maximize efficacy for patients with a low risk of irAEs. Meanwhile, clinicians should improve patient irAE monitoring during treatment, and hospitals should consider establishing a specialized irAE management team to guarantee patient safety.

In the past, studies in the field of NAI mainly focused on assessing efficacy metrics such as pCR, while systematic studies of irAEs are lacking. The results of this study suggest that future clinical trial designs should place equal importance on efficacy and safety metrics and promote studies with long-term follow-up durations to adequately explore long-term safety issues after NAI treatment. In addition, previous guidelines mostly focused on the efficacy and indications of NAI, with less attention on safety management. The findings of this study suggest the need for guideline developers to emphasize the concept of NAI efficacy-safety balance in a more systematic way and to develop standardized processes and specific recommendations for the risk assessment and management of irAEs in different patients.

Existing research in this field has focused on highly prevalent tumors such as lung cancer, breast cancer, colorectal cancer, gastric cancer, bladder cancer, esophageal cancer, and melanoma. With the increased focus on medical precision, future studies should prospectively explore the potential application of NAI in rare cancer types and tumors with specific molecular subtypes. This will not only reveal the universal rules and unique mechanisms of immunotherapy in a wider range of cancer types but also help to expand the applications of NAI, thus providing new therapeutic avenues for patients for whom existing treatment options have been ineffective.

Notably, although this study emphasized the importance of balancing the efficacy and safety of NAI, accurately differentiating between those with efficacy benefits and those with a high risk of AEs remains a challenge in current clinical practice. Artificial intelligence technologies (such as deep learning networks, transfer learning algorithms, ML algorithms) and multiomics data (such as radiomics, pathomics, genomics, proteomics, metabolomics, and microbiomics) offer significant opportunities to construct biomarkers or models that accurately predict the efficacy and safety of immunotherapy.17,20 However, in practice, the sequenced tissues often come from only a small region of the tumor, making it difficult to completely determine the entire landscape of the mass and the biological differences between different regions. In contrast, recent spatial omics models have demonstrated superior performance in characterizing tumor heterogeneity, thereby enhancing precise individualized guidance for clinical decision-making.^{17 21} It is also important to note that the construction of irAE prediction models is challenging because of the involvement of numerous complex factors and characteristics. Future research should focus on and further investigate additional factors and characteristics, such as patient genetic background heterogeneity, individual variations in autoimmune status, multiorgan involvement, dose-independent effects, and temporal heterogeneity in irAE onset.

Conclusion

This study used ML-driven bibliometric analysis to systematically characterize global scientific landscapes and emerging trends in irAEs of NAI among patients with perioperative cancer over the past decade. Our findings highlighted a significant global increase in research attention towards balancing the efficacy and safety of NAI, identifying critical research clusters and trends such as standardizing response assessment criteria, combination therapy strategies, and management of irAEs. Despite increasing awareness, there remains a notable gap in specialized studies on specific types of irAEs, their underlying mechanisms, and precise clinical management strategies. Clinically, our study underscores the necessity for personalized therapeutic regimens, refined AE monitoring, and the establishment of multidisciplinary teams dedicated to irAE management. Future research should leverage advanced artificial intelligence and multiomics technologies to develop robust predictive models for irAEs, facilitating the safe and effective clinical application of NAI.