Exploratory Analysis for Development Predictive Models of Immune Checkpoint Inhibitor-Induced Myocarditis Using a Nationwide Claims Database

Abstract

Immune checkpoint inhibitors (ICIs), essential in cancer therapy, can cause severe immune-related adverse events (irAEs), including myocarditis with a high fatality rate. Currently, the pathogenesis, biomarkers, and risk factors of ICI-induced myocarditis (ICIM) are not fully understood. This exploratory study aimed to develop machine learning-based models to predict the onset of ICIM within 3 months of starting ICI therapy, using a large health insurance database. The models were constructed using the Light Gradient Boosting Machine (LightGBM) and Random Forest algorithms, incorporating clinical variables such as comorbidities and prior medication classifications. In this study, a strategy combining undersampling and bagging was used to minimize the impact of highly imbalanced datasets. The Random Forest model demonstrated superior performance compared with the LightGBM model, and the SHapley Additive exPlanations (SHAP) analysis for the Random Forest model revealed that the concurrent use of ICIs was the most important variable for predictions. Although predictive performance remains limited (AUROC ≈ 0.63), this exploratory framework demonstrates the feasibility of developing data-driven risk prediction models for ICIM. Future studies with expanded datasets and integration of laboratory parameters are warranted to improve predictive accuracy and potential clinical applicability.

INTRODUCTION

Immune checkpoint inhibitors (ICIs) are a pivotal breakthrough in cancer treatment. These agents exert their antitumor effects by inhibiting immune checkpoint signaling and T cell activation. However, treatment with ICIs can lead to a spectrum of inflammatory adverse events, also known as immune-related adverse events (irAEs).¹⁾ Approximately 10% of patients with cancer treated with ICIs develop irAEs that are severe enough to require treatment discontinuation.²⁾ The fatality rate of ICI-induced myocarditis (ICIM) ranges from 25 to 50%, which is the highest reported rate among irAEs,³⁾ thus warranting particular attention.

Currently, the pathogenesis of ICIM remains unclear. However, it has been suggested that excessive activation of immune mechanisms involving T and other immune cells attacking the myocardial tissue may play a role. This process is thought to arise from the disruption of normal immune regulation caused by ICIs, which inhibit immune checkpoints, leading to autoimmune responses. As a curative treatment for ICIM has not yet been established, the primary therapeutic approach consists of palliative care against inflammatory responses, such as administration of corticosteroids and immunosuppressants.⁴⁾

Several factors have been identified as key biomarkers for the diagnosis of ICIM. Previous studies reported that troponin I, a cardiac-specific antigen, is released into the bloodstream, resulting in elevated serum troponin I levels in patients with ICIM.⁵⁾ Troponin I levels are increased in 94% of patients with ICIM, making it the most sensitive diagnostic indicator.⁵⁾ However, increases in troponin I levels can also occur in non-cardiac diseases such as chronic kidney disease.⁶⁾ Therefore, using troponin I as a predictive marker for ICIM is challenging because of the high rate of false positives.

Previously suggested risk factors for ICIM include ICI combination therapy, diabetes, obesity, radiation therapy, concurrent angiogenesis inhibitors, and anthracyclines.^4,5,7) However, the mechanisms by which these risk factors contribute remain unclear, and the effects of other factors, or the synergistic or antagonistic effects of multiple factors, remain unknown.

Owing to this, a definitive curative treatment for ICIM has not yet been established and no comprehensive predictive markers are available. Nonetheless, it is clinically important to identify at-risk patients before initiating treatment with ICIs, to allow appropriate monitoring and early intervention. However, the development of predictive models for ICIM has been hindered by a limited number of available clinical cases in previous studies.

This study was conducted to construct a model for predicting the onset of ICIM using a claims database that includes an extensive number of patient records. Clinical information available prior to the initiation of ICI therapy was used to calculate the comprehensive risk probability using multiple variables. The predictive process of the model was further elucidated with SHapley Additive exPlanations (SHAP)⁸⁾ to estimate risk-related factors.

MATERIALS AND METHODS

Data Source and Study Design

Data were collected from the Medical Data Vision (MDV) database (purchased from Medical Data Vision Co., Ltd., Tokyo, Japan) covering 485 acute care institutions from April 2008 to December 2022. At the time of data collection, this health insurance claims database included 1040184 patients who were prescribed medications coded as L01 under the Anatomical Therapeutic Chemical (ATC) classification system, representing antineoplastic agents.

This analysis included hospitalized patients with cancer aged >18 years who received ICIs. Patients diagnosed with myocarditis before the index date were excluded. The index date was defined as the first prescription of an ICI after cancer diagnosis, occurring more than 1 year after the start of the observation period. Supplementary Table S1 shows the list of ICIs. Patients with missing body mass index (BMI) data or a Barthel Index score were excluded.

Data Extraction

The follow-up period was set at 3 months, given that over 80% of patients who develop ICIM experience onset within 3 months of initiating ICI therapy.⁵⁾ The predictive model developed in this study aimed to forecast onset within 3 months of the index date. Patients with ICIM were identified based on the diagnosis of myocarditis, as indicated in Supplementary Table 2, and an interruption in ICI treatment exceeding 2 months after the ICIM diagnosis.

We defined a look-back period of 1 year before the index date (day −365 to day 0). Comorbidities were identified using ICD-10 diagnostic codes recorded within this window, and concomitant medications were defined based on prescription records covering the same period. The definitions of comorbidities followed those used in the Elixhauser comorbidity index.⁹⁾ The Charlson Comorbidity Index, which assigns weighted scores to conditions associated with mortality risk,¹⁰⁾ was calculated using a modified version compatible with ICD-10 coding.¹¹⁾ Demographic variables (age, sex, BMI), comorbidities, and prior medications were included. Medications were categorized according to the ATC classification. All predictor variables were restricted to information available before or at the index date to avoid potential data leakage. A complete list of variables before feature selection is provided in Supplementary Table 3.

Data were extracted from the MDV database using Navicat for SQLite (version 16.3.7).

Model Development and Validation

For variable selection and model development, 80% of the dataset was used, whereas the remaining 20% was used for validation. The training and test datasets were divided to match the proportions of positive cases.

Initially, the training dataset was used to select the variables. For continuous variables, a pairwise Pearson’s correlation matrix was used to assess collinearity and establish a pairwise correlation threshold of r > 0.8. For categorical variables, the chi-square statistic was calculated as an indicator of the association with ICIM onset, with the exclusion of statistics below p = 0.1.

Two algorithms, a light-gradient boosting machine (LightGBM) and a Random Forest, were employed for model development. LightGBM uses first- and second-order negative gradients of the loss function to improve the prediction performance, along with a histogram-based decision tree algorithm that enhances the execution efficiency.¹²⁾ Random Forest, proposed by Breiman in 2001, is a bagging ensemble algorithm that uses decision trees as base learners.¹³⁾ It incorporates random feature selection to generate multiple trees, with the final prediction determined by voting among all tree classification outcomes.

Undersampling and bagging were implemented to create models for dealing with highly imbalanced datasets.¹⁴⁾ Weak learners trained on small datasets designed to have a positive-to-negative case ratio of 1 : 5 were aggregated until all negative cases were used. The model hyperparameters were set to more than 100 trials using the Optuna software (Supplementary Table 4) which automates the iterative process of hyperparameter optimization.¹⁵⁾ The optimization function was set to the F5 score, which is a variant of the F-measure commonly used to evaluate model performance in machine learning. It balances precision and recall in a way that emphasizes recall over precision by a factor of five. The formula for the F5 score is as follows:

  

$$F_{5}score=(1+5^2)\cdot \frac{precision\cdot recall}{(5^2\cdot precision)+recall}$$

The F5 score specifically highlights the importance of recall, and was selected because missing high-risk cases is more detrimental than including false positives.

Subsequently, the model was evaluated through 100 iterations of a threefold cross-validation of the training dataset. Evaluation metrics, such as the F5 score, Accuracy, Recall, Precision, Average Precision (AP), and area under the receiver operating characteristic curve (AUROC) were calculated. Finally, the model was validated using test data by employing the evaluation metrics used in the development.

The superior model was further analyzed using SHAP to visualize the important variables. SHAP is a method derived from cooperative game theory that specifically uses the concept of Shapley values, which was originally proposed to distribute payoffs fairly among players in a coalition.⁸⁾ In machine learning, SHAP values provide a measure of the importance of each feature that contributes to the prediction outcome of a model.

Throughout the development and validation of the model, Python version 3.9.7 was used. The packages and their versions are listed in Supplementary Table 6.

Statistical Analysis

To assess the numerical data, normality was initially verified using the Shapiro–Wilk test. If normality was not established, the nonparametric Mann–Whitney U test was used. In cases where normality was confirmed, the homoscedasticity between groups was examined using Levene’s test. If homoscedasticity was established, the Student’s t-test was applied; otherwise, the Welch’s t-test was used. All statistical analyses were performed using R software version 4.1.2. and Python version 3.9.7.

Ethics Approval

The study complied with Chapter 1, Section 3, Part 1, Subsection (C), Item 3 of the ethical guidelines for Medical and Health Research Involving Human Subjects issued by the Ministry of Health, Labour and Welfare of Japan. According to this guideline, as it involved a secondary analysis of anonymized processed databases with no personally identifiable information included, approval from an institutional committee was not required. Consequently, informed consent from individual subjects or their legal guardians was not applicable.

RESULTS

Patient Characteristics

A total of 83424 patients were prescribed an ICI at least once in the MDV database between April 2008 and December 2022. A total of 33824 patients were excluded based on the exclusion criteria shown in Fig. 1. The resultant dataset included 48600 patients; however, 1154 patients were excluded due to missing BMI or Barthel Index data. Finally, 48446 patients were included in the dataset used to develop the predictive models, of which 117 (0.24%) had ICIM.

Fig. 1. Patient Selection Flowchart

Flowchart illustrating the process of patient selection from the Medical Data Vision database. ICIs: Immune checkpoint inhibitors.

The baseline patient characteristics are presented in Table 1. The overall average age was 69.4 years, with ICIM and non-ICIM groups averaging 69.1 and 69.4 years, respectively. The proportion of men was higher than that of females, accounting for 75.3% overall, 74.4% in the ICIM group, and 75.4% in the non-ICIM group.

Table 1. Baseline Characteristics of Patients Who Received ICI Prescriptions

	Overall	ICIM	Non-ICIM
n	48446	117	48329
Age, years (mean (S.D.))	69.4 (9.6)	69.1 (9.9)	69.4 (9.6)
Sex (%)
Female	11942 (24.7)	30 (25.6)	11912 (24.6)
Male	36504 (75.3)	87 (74.4)	36417 (75.4)
BMI, kg/m2 (mean (S.D.))	21.9 (3.7)	22.7 (4.0)	21.9 (3.7)
Barthel Index (mean (S.D.))	95.6 (15.2)	95.3 (15.2)	95.6 (15.2)
Charlson Index (mean (S.D.))	7.3 (3.3)	7.0 (3.2)	7.3 (3.3)
Type of ICI (%)
Nivolumab	16913 (34.9)	66 (56.4)	16847 (34.9)
Pembrolizumab	19376 (40.0)	39 (33.3)	19337 (40.0)
Durvalumab	3353 (6.9)	1 (0.9)	3352 (6.9)
Atezolizumab	8241 (17.0)	9 (7.7)	8232 (17.0)
Avelumab	544 (1.1)	2 (1.7)	542 (1.1)
Ipilimumab	2630 (5.4)	23 (19.7)	2607 (5.4)
ICI combined use (%)	2611 (5.4)	23 (19.7)	2588 (5.4)
Prescription count of prior medications (mean (S.D.))	16.6 (8.9)	16.4 (8.9)	16.6 (8.9)
Cancer type (%)
Head and neck squamous cell cancers	2816 (5.8)	6 (5.1)	2810 (5.8)
Esophageal and gastric cancers	8123 (16.8)	34 (29.1)	8089 (16.7)
Colorectal cancer	1725 (3.6)	7 (6.0)	1718 (3.6)
Hepatobiliary and pancreatic cancers	3219 (6.6)	7 (6.0)	3212 (6.6)
Lung cancer	28753 (59.4)	44 (37.6)	28709 (59.4)
Skin cancer	958 (2.0)	4 (3.4)	954 (2.0)
Breast cancer	683 (1.4)	5 (4.3)	678 (1.4)
Gynecologic cancer	507 (1.0)	3 (2.6)	504 (1.0)
Prostate cancer	1207 (2.5)	1 (0.9)	1206 (2.5)
Urologic cancer	7169 (14.8)	33 (28.2)	7136 (14.8)
Central nervous system cancers	75 (0.2)	0 (0.0)	75 (0.2)
Thyroid cancer	216 (0.4)	0 (0.0)	216 (0.4)
Hematologic cancers	1097 (2.3)	3 (2.6)	1094 (2.3)
Other cancers	750 (1.5)	3 (2.6)	747 (1.5)

BMI: body mass index; ICIs: immune checkpoint inhibitors; ICIM: immune checkpoint inhibitor-induced myocarditis; S.D.: standard deviation.

Feature Selection and Model Development

As shown in Supplementary Fig. 1, none of the pairwise Pearson correlation coefficients for the continuous variables exceeded 0.8, demonstrating a lack of collinearity among the variables. After assessing the association with ICIM onset, the number of categorical variables decreased from 585 to 87 (Supplementary Table 3).

In the next stage of model development, two algorithms, LightGBM and Random Forest, were used to construct two models. Combining undersampling with bagging is the best strategy for handling highly imbalanced datasets. Therefore, the approach of bagging an ensemble of weak learners, which was induced over balanced training datasets through undersampling, was adopted. After determining the hyperparameters in the training dataset for model optimization (Supplementary Table 5), the performance of the model was evaluated using the training dataset (Fig. 2).

Fig. 2. Evaluation Metrics for Models Assessed with 100 Iterations of Three Cross-Validation Repeated on the Training Dataset

Evaluation metrics for light-gradient boosting machine (LightGBM) and Random Forest models after 100 iterations of three-fold cross-validation on the training dataset. The metrics included the F5 score, Accuracy, Recall, Precision, Average Precision (AP), and Area Under the Receiver Operating Characteristic Curve (AUROC). ^*p < 0.05, statistically significant differences between the models.

Following 100 iterations of threefold cross-validation, the AUROC values for LightGBM and Random Forest were 0.609 ± 0.024 and 0.699 ± 0.009, respectively (Supplementary Figs. 2, 3). For LightGBM, the F5 score, Accuracy, Recall, Precision, and AP were 0.103 ± 0.011, 0.797 ± 0.080, 0.403 ± 0.101, 0.006 ± 0.002, and 0.005 ± 0.002, respectively. For Random Forest, the corresponding scores were 0.160 ± 0.013, 0.916 ± 0.028, 0.355 ± 0.050, 0.012 ± 0.002, and 0.009 ± 0.001. Overall, in the evaluation using the training dataset, the Random Forest model demonstrated superior performance compared with that using LightGBM.

Subsequently, the predictive performance was validated using a test dataset. As shown in Table 2 and Fig. 3, there was a minimal difference in the predictive accuracy between the LightGBM and Random Forest models. The AUROC values for LightGBM and Random Forest were 0.635 (95% confidence interval [CI]: 0.493–0.754) and 0.634 (95% CI: 0.507–0.751). For LightGBM, the Accuracy, F5 score, Recall, and Precision were 0.969 (95% CI: 0.966–0.972), 0.178 (95% CI: 0.176–0.180), 0.260 (95% CI: 0.056–0.754), and 0.020 (95% CI: 0.004–0.034), respectively. For Random Forest, the corresponding scores were 0.974 (95% CI: 0.971–0.977), 0.191 (95% CI: 0.189–0.193), 0.260 (95% CI: 0.055–0.407), and 0.025 (95% CI: 0.004–0.040). Compared with the confusion matrices in Figs. 3A and 3B, the allocation of predictions for actual positive ICIM cases was the same. However, regarding the allocation of negative non-ICIM cases, the Random Forest model had slightly lower false-positive rates (LightGBM: 2.921% and Random Forest: 2.415%).

Table 2. Evaluation Metrics of Predicting for Test Dataset

Metrics	LightGBM	Random forest
Accuracy	0.969 (0.966–0.972)	0.974 (0.971–0.977)
F5 score	0.178 (0.176–0.180)	0.191 (0.189–0.193)
Recall	0.260 (0.056–0.754)	0.260 (0.055–0.407)
Precision	0.020 (0.004–0.034)	0.025 (0.004–0.040)
AUROC	0.635 (0.493–0.754)	0.634 (0.507–0.751)

95% confidence interval shown in parentheses. AUROC: Area under receiver operating characteristic curve; LightGBM: Light gradient boosting machine.

Fig. 3. Validation of the Developed Model Using the Test Dataset

Performance of (A) LightGBM and (B) Random Forest models applied to the test dataset for validation. (C) Receiver Operating Characteristic (ROC) curves for LightGBM and Random Forest models. The curves were plotted using the true-positive and false-positive rates derived from the test dataset. The area under the curve (AUC) values indicate the overall ability of the model to discriminate between ICIM and non-ICIM cases.

Model Explainability

Lastly, the prediction process for 23 patients with ICIM in the test dataset was visualized using SHAP with the Random Forest prediction model. The SHAP summary plot (Fig. 4A) and beeswarm plot (Fig. 4B) indicate the 10 most important variables for predictions using the Random Forest model. The variable that contributed most notably to the predictions was the combined use of ICIs, which was found to be associated with an increase in the output prediction probability. The following variables, ranked by their contribution to the model’s predictions, were obtained: prescriptions of R05CB agents (osmotically acting laxatives, such as magnesium oxide), prescriptions of N05AH agents (antipsychotics, such as diazepines, oxazepines, thiazepines, and oxepines), lung cancer diagnosis, urologic cancer diagnosis, age, esophageal and gastric cancer diagnosis, prescriptions of ipilimumab, number of medications prescribed, and A06AD (mucolytics, such as acetylcysteine and bromhexine) prescriptions. Among these, the variables that contributed to an increase in the prediction probability were diagnoses of urological, esophageal, and gastric cancers, and prescriptions of ipilimumab, with the prediction probability increasing with age. Conversely, the variables that contributed to a decrease in prediction probability were prescriptions of R05CB and N05AH agents as well as lung cancer diagnoses, with the prediction probability decreasing as the number of medications prescribed increased.

Fig. 4. SHAP Analysis of the Random Forest Model for ICIM Onset in the Test Dataset

(A) SHAP summary plot illustrating the impact of the top 10 most important features on the Random Forest model’s predictions for ICIM onset. (B) SHAP beeswarm plot depicting how each of the top 10 most important features influence the model’s predictions. Each dot represents a SHAP value for a feature, indicating its contribution to the predicted probability of developing ICIM.

DISCUSSION

In this study, we developed machine learning prediction models for ICIM using a health insurance claims database and estimated the key risk factors for ICIM by visualizing the prediction processes of the models. The Random Forest model demonstrated superior performance compared with the LightGBM model, and SHAP analysis revealed that the combined use of ICIs was the most significant variable for predictions. This study clarifies the clinical information that affects the risk of developing ICIM and suggests potential therapeutic strategies in clinical practice.

In previous studies focusing on irAE prediction, machine learning models were developed to identify the overall occurrence of irAEs, hypophysitis, and pneumonitis.^16–18) However, no predictive model for ICIM has been constructed to date. The reasons for this are the lack of clinical cases and difficulty in model development owing to data imbalance. Herein, we accumulated the necessary number of cases for model development using a health insurance claims database. Our study involved 83424 patients who were prescribed ICIs at least once, and 48446 patients were included in the final dataset after applying the exclusion criteria and accounting for missing data. Among these, 117 patients had ICIM, making this the largest dataset for ICIM prediction to date. However, due to the low incidence of ICIM, the datasets used were highly imbalanced between the positive and negative cases. Generally, machine learning prediction models exhibit decreased performance with imbalanced data compared with balanced data.¹⁹⁾ Herein, we combined undersampling and bagging to minimize the impact of data imbalance on model performance.¹⁴⁾ Using this strategy in combination with LightGBM and Random Forest algorithms, two models that could effectively handle imbalanced datasets were developed. The Random Forest model demonstrated superior performance during the evaluation of the training dataset, and both models showed minimal differences in predictive performance during test dataset validation. The AUROC values for both models were modest, indicating room for improvement. Nevertheless, SHAP analysis for model explainability highlighted important variables such as ICI combination use and age, which contributed to the prediction of ICIM, reinforcing model validity.^4,20)

Recognizing the variables that increase the likelihood of developing ICIM is crucial for identifying high-risk patients. SHAP analysis resulted in 10 variables of particular importance, with risk-increasing variables including the combined use of ICIs, age, ipilimumab, and cancer type, such as urologic, esophageal, and gastric cancers. Among these, the previously reported risk factors were combined ICI use and age.^4,21) Ipilimumab, an anti-CTLA-4 antibody, does not significantly differ in ICIM onset rates compared with ICIs targeting other immune checkpoint molecules.²¹⁾ However, in cases of combined ICI use, the combination often includes anti-CTLA-4 antibodies, highlighting ipilimumab as a risk-increasing factor. The varying contributions of cancer types suggest that the incidence of ICIM may differ depending on tumor context. Possible explanations include differences in ICI regimen composition, prior therapies (e.g., chemotherapy, radiotherapy), and variations in the immune microenvironment across the various cancer types. Of note, signals observed in lung, urologic, and esophago-gastric cancers warrant further investigation in future studies. Interestingly, SHAP analysis also identified features that have not been noted previously. R05CB (mucolytics) and A06AD (osmotic laxatives) are typically prescribed for respiratory symptoms or constipation and may serve as proxy indicators of cancer type, concomitant therapies, or overall health status. Their inverse association with ICIM risk may reflect surrogate effects of treatment patterns—such as higher healthcare access frequency or earlier supportive interventions—rather than a direct pharmacological role. Prescription count was also identified as an influential variable. While higher counts may reflect a greater comorbidity burden and increased healthcare contact, they could simultaneously lead to closer monitoring and earlier treatment discontinuation, thereby appearing to reduce the observed ICIM risk. Similarly, prescriptions for N05AH (antipsychotics) may act as markers of comorbidity or disease severity (e.g., delirium, psychiatric conditions, sleep disorders). At the same time, they may introduce bias through drug–drug interactions or diagnostic/reporting practices. Overall, our findings reinforce previously known risk factors and also suggest that clinical context, supportive medications, and treatment patterns may contribute to the onset of ICIM. In particular, the roles of cancer type and supportive therapies, which have not been systematically examined in previous studies, merit further exploration.

Considering the high mortality rate associated with ICIM, it is critical to avoid its development in patients receiving ICI therapy. Early prediction of ICIM is highly important, as it enables proactive intervention for high-risk patients and prevents the occurrence of severe events. This not only protects patients’ lives but also contributes to reducing healthcare costs. Furthermore, the introduction of a predictive model is expected to advance personalized treatment for individual patients. Currently, serum troponin I levels are commonly used to monitor ICIM in clinical practice; however, the high rate of false positives is a concern.^5,22) Therefore, improving newly developed predictive models is likely to be extremely useful.

A major limitation of this study pertains to missing data. Because the dataset was compiled at the hospital level from acute care institutions, no information on comorbidities and prior medications before and after hospital transfers or visits to other facilities was captured. Additionally, patients with missing demographic information, such as BMI and Barthel Index, were excluded, which may have introduced selection bias. Furthermore, when incorporating laboratory biomarkers, extensive missingness reduced the analyzable cohort from approximately 48000 to only 6305 patients, which limited the statistical power and prevented the development of a robust model. Despite this reduction, a modest improvement in the F₅ score was observed, suggesting that laboratory markers such as troponin I can enhance predictive performance (data not shown). To overcome these limitations, future studies should leverage larger datasets, including international cohorts, and systematically integrate laboratory values to construct more accurate and generalizable prediction models for ICIM. Second, it is unclear whether the institutions included in the MDV database used in this study are representative of those in Japan. Although we used data from a relatively large number of institutions, we postulate that the model performance was unaffected by the selection of these institutions. Because we used only one dataset for model development and validation, our models should be validated using data from different institutions in Japan. Furthermore, to extend the applicability of the model globally, it will be necessary to validate it using data from institutions in other countries. In future global applications, it is essential to consider economic conditions and social status. In Japan, the universal health insurance and high-cost medical care systems ensure that almost everyone has equal access to medical treatment. Therefore, potential bias arising from economic conditions and social status were ignored in this study. Nevertheless, in constructing a global model, it will be necessary to pay attention to such biases. Improving the performance of the predictive model developed in this study is crucial for clinical application. The reasons for the insufficient performance may include a lack of sufficient cases and useful variables. Combining this dataset with other claims data could further increase the number of training cases. The relevant variables may include values from blood tests; however, the accumulation of test values is limited to a few medical institutions in the MDV database, making it difficult to include these factors.

In this study, we explored the feasibility of developing a machine learning model to predict the onset of ICIM using a large-scale health insurance claims database. Although the current model’s performance remains modest because of the rarity of ICIM and the limitations of available variables, this study establishes an initial framework for data-driven risk prediction. The results highlight the potential value of integrating broader datasets, including international cohorts and laboratory biomarkers, to improve predictive accuracy. Rather than providing a definitive predictive tool, our analysis offers a foundation upon which more refined and clinically applicable models can be built in the future.

Acknowledgments

One of the authors, R.Y., would like to express their sincere gratitude to the Public Interest Incorporated Foundation “Ohmoto Ikueikai” for their generous financial assistance through the scholarship over the past 2 years.

DECLARATIONS

Funding

This work was partly supported by JSPS KAKENHI Grant Numbers: JP 24K09913 (to H.H.), 24KK0181 (to T.U.), JSPS Fellows 25KJ1851 (to R.Y.), and Ryobi Teien Memorial Foundation Research Grant.

Author Contributions

Conceptualization: T.U.; Methodology: R.Y., H.H.; Data curation: R.Y.; Formal analysis and investigation: R.Y., K.N.; Software and visualization: R.Y., K.N.; Validation: H.H., K.N., M.U.; Writing—original draft preparation: R.Y.; Writing—review and editing: H.H., K.N., M.U, A.M., A.F.O., P.P., M.T., Y.Z., T.U.; Project administration: H.H.; Funding acquisition: H.H., T.U.; Resources: H.H., M.T.; Supervision: Y.Z., T.U.

Conflict of Interest

The authors declare no conflict of interest.

Data Availability

The data that support this study’s findings are available from commercial providers (Medical Data Vision Co., Ltd.), but restrictions apply to their availability. These data were used under license for the current study and are not publicly available. However, data are available from the authors upon reasonable request and with permission of Medical Data Vision Co., Ltd. “https://www.mdv.co.jp/contactus/form.php?inquiry_package=5.”

Supplementary Materials

This article contains supplementary materials.

REFERENCES

• 1)  Kennedy LB, Salama AKS. A review of cancer immunotherapy toxicity. CA Cancer J. Clin., 70, 86–104 (2020).
• 2)  Haratani K, Hayashi H, Chiba Y, Kudo K, Yonesaka K, Kato R, Kaneda H, Hasegawa Y, Tanaka K, Takeda M, Nakagawa K. Association of immune-related adverse events with Nivolumab efficacy in non-small-cell lung cancer. JAMA Oncol., 4, 374–378 (2018).
• 3)  Salem J-E, Manouchehri A, Moey M, Lebrun-Vignes B, Bastarache L, Pariente A, Gobert A, Spano J-P, Balko JM, Bonaca MP, Roden DM, Johnson DB, Moslehi JJ. Cardiovascular toxicities associated with immune checkpoint inhibitors: an observational, retrospective, pharmacovigilance study. Lancet Oncol., 19, 1579–1589 (2018).
• 4)  Waliany S, Lee D, Witteles RM, Neal JW, Nguyen P, Davis MM, Salem J-E, Wu SM, Moslehi JJ, Zhu H. Immune checkpoint inhibitor cardiotoxicity: understanding basic mechanisms and clinical characteristics and finding a cure. Annu. Rev. Pharmacol. Toxicol., 61, 113–134 (2021).
• 5)  Mahmood SS, Fradley MG, Cohen JV, Nohria A, Reynolds KL, Heinzerling LM, Sullivan RJ, Damrongwatanasuk R, Chen CL, Gupta D, Kirchberger MC, Awadalla M, Hassan MZO, Moslehi J, Shah SP, Ganatra S, Thavendiranathan P, Lawrence DP, Groarke JD, Neilan TG. Myocarditis in patients treated with immune checkpoint inhibitors. J. Am. Coll. Cardiol., 71, 1755–1764 (2018).
• 6)  Spallarossa P, Tini G, Sarocchi M, Arboscello E, Grossi F, Queirolo P, Zoppoli G, Ameri P. Identification and management of immune checkpoint inhibitor-related myocarditis: use troponin wisely. J. Clin. Oncol., 37, 2201–2205 (2019).
• 7)  Hu J-R, Florido R, Lipson EJ, Naidoo J, Ardehali R, Tocchetti CG, Lyon AR, Padera RF, Johnson DB, Moslehi J. Cardiovascular toxicities associated with immune checkpoint inhibitors. Cardiovasc. Res., 115, 854–868 (2019).
• 8)  Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst., 30, 4768–4777 (2017).
• 9)  Elixhauser A, Steiner C, Harris DR, Coffey RM. Comorbidity measures for use with administrative data. Med. Care, 36, 8–27 (1998).
• 10)  Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. J. Chronic Dis., 40, 373–383 (1987).
• 11)  Quan H, Sundararajan V, Halfon P, Fong A, Burnand B, Luthi J-C, Saunders LD, Beck CA, Feasby TE, Ghali WA. Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Med. Care, 43, 1130–1139 (2005).
• 12)  Ke G, Meng Q, Finley T, et al. LightGBM: a highly efficient gradient boosting decision tree. Adv. Neural Inf. Process. Syst., 30, 3149–3157 (2017).
• 13)  Breiman L. Random forests. Mach. Learn., 45, 5–32 (2001).
• 14)  Wallace BC, Small K, Brodley CE, Trikalinos TA. Class imbalance, redux. 11th International Conference on Data Mining. IEEE Publications, Los Alamitos, CA, USA, pp. 754–763 (2011).
• 15)  Akiba T, Sano S, Yanase T, Ohta T, Koyama M. Optuna: A next-generation hyperparameter optimization framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD 2019), pp. 2623–2631 (2019).
• 16)  Sung C, An J, Lee S, et al. Integrative analysis of risk factors for immune-related adverse events of checkpoint blockade therapy in cancer. Nat. Cancer, 4, 844–859 (2023).
• 17)  Kim W, Cho YA, Min KH, Kim D-C, Lee K-E. Machine learning approaches for assessing risk factors of adrenal insufficiency in patients undergoing immune checkpoint inhibitor therapy. Pharmaceuticals, 16, 1097 (2023).
• 18)  Gong L, Gong J, Sun X, Yu L, Liao B, Chen X, Li Y-S. Identification and prediction of immune checkpoint inhibitors-related pneumonitis by machine learning. Front. Immunol., 14, 1138489 (2023).
• 19)  Ashraf S, Saleem S, Ahmed T, Aslam Z, Muhammad D. Conversion of adverse data corpus to shrewd output using sampling metrics. Vis. Comput. Ind. Biomed. Art, 3, 19 (2020).
• 20)  Zamami Y, Niimura T, Okada N, Koyama T, Fukushima K, Izawa-Ishizawa Y, Ishizawa K. Factors associated with immune checkpoint inhibitor-related myocarditis. JAMA Oncol., 5, 1635–1637 (2019).
• 21)  Liu M, Cheng X, Ni R, Zheng B, Huang S, Yang J. Cardiotoxicity of immune checkpoint inhibitors: a frequency network meta-analysis. Front. Immunol., 13, 1006860 (2022).
• 22)  Lyon AR, López-Fernández T, Couch LS, et al. 2022 ESC Guidelines on cardio-oncology developed in collaboration with the European Hematology Association (EHA), the European Society for Therapeutic Radiology and Oncology (ESTRO) and the International Cardio-Oncology Society (IC-OS). Eur. Heart J., 43, 4229–4361 (2022).