Explainable machine learning for neoplasms diagnosis via electrocardiograms: an externally validated study.

Our primary dataset for training and internal evaluation was derived from the MIMIC-IV-ECG database [ 26 , 27 ], a subset of a large-scale critical care dataset collected at the Beth Israel Deaconess Medical Center in Boston, Massachusetts. This dataset encompasses patients admitted to the emergency department (ED) and intensive care unit (ICU). Target variables are based on discharge diagnoses encoded using the International Classification of Diseases Clinical Modification (ICD-10-CM). While a wide range of neoplasm-related codes is explored, this study focuses on those achieving internal and external validation AUROC scores above 0.7, covering neoplasms across diverse physiological systems (Fig.  1 ). Fig. 1 Schematic representation of our proposed approach. We use as internal dataset the MIMIC-IV-ECG dataset from which we use as input features demographics and ECG features to train a tree-based model and diagnose diverse neoplasms. For external validation we take a second cohort of patients from the ECG-View II dataet from which we collect the same set of features and neoplasms targets. The definition of neoplasms are represented by ICD10-CM codes Schematic representation of our proposed approach. We use as internal dataset the MIMIC-IV-ECG dataset from which we use as input features demographics and ECG features to train a tree-based model and diagnose diverse neoplasms. For external validation we take a second cohort of patients from the ECG-View II dataet from which we collect the same set of features and neoplasms targets. The definition of neoplasms are represented by ICD10-CM codes To construct a comprehensive and harmonized feature set, ECG features from MIMIC-IV were aligned with those from the ECG-VIEW-II database [ 28 ], which serves as our secondary dataset for external validation. ECG-VIEW-II includes data collected from patients at a South Korean tertiary teaching hospital. The standardized feature set comprises ECG-derived measurements (RR-interval, PR-interval, QRS-duration, QT-interval, QTc-interval in milliseconds; P-wave-axis, QRS-axis, and T-wave-axis in degrees) alongside demographic attributes (binary sex and age as a continuous variable) (Table  1 ). Table 1 A summary of variable characteristics across samples, including demographic details such as gender counts (with ratios) and the median age in years (with interquartile range, IQR), along with age distribution represented by quantile ratios Variable MIMIC-IV-ECG ECG-View II Gender (%)  Female 226,892 (48.50) 375,733 (48.44)  Male 240,837 (51.49) 399,802 (51.55) Age (%)  Median years (IQR) 66 (25) 52 (25)  Quantile 1 18-53 (23.83) 18-40 (24.03)  Quantile 2 53-66 (25.16) 40-52 (25.75)  Quantile 3 66-78 (25.60) 52-65 (24.94)  Quantile 4 78-101 (25.40) 65-109 (25.28) ECG features (IQR)  RR-interval 769 (264) 857 (227)  PR-interval 158 (38) 158 (28)  QRS-duration 94 (23) 90 (14)  QT-interval 394 (68) 392 (48)  QTc-interval 447 (47) 421 (37)  P-wave-axis 51 (32) 53 (28)  QRS-axis 13 (61) 48 (49)  T-wave-axis 42 (58) 44 (33) Similarly, it covers the median (IQR) values for ECG features like the RR-interval, PR-interval, QRS-duration, QT-interval, and QTc-interval in milliseconds, as well as the P-wave axis, QRS axis, and T-wave axis in degrees A summary of variable characteristics across samples, including demographic details such as gender counts (with ratios) and the median age in years (with interquartile range, IQR), along with age distribution represented by quantile ratios Similarly, it covers the median (IQR) values for ECG features like the RR-interval, PR-interval, QRS-duration, QT-interval, and QTc-interval in milliseconds, as well as the P-wave axis, QRS axis, and T-wave axis in degrees For the internal dataset, stratified folds are created based on diagnoses, age, and gender distributions, utilizing an 18:1:1 split as described in prior work [ 12 ]. A comparable stratification procedure is applied to the external dataset to maintain consistency. The training process prioritizes MIMIC-IV-ECG due to its broader ethnic diversity compared to ECG-VIEW-II, thereby enhancing the model’s generalization across diverse populations,as demonstrated in previous research [ 29 ], which employs a similar approach mostly for cardiac conditions and [ 30 ] for diverse liver disease conditions. This approach ensures robust internal training and reliable external validation across ethnically and geographically distinct cohorts. In this study, we develop individual tree-based models using Extreme Gradient Boosting (XGBoost) to address binary classification tasks, with a separate model for each selected ICD-10-CM code. To prevent overfitting, we implement early stopping with a patience of 10 iterations on the validation fold during training. To this end, model performance is evaluated using the area under the receiver operating characteristic curve (AUROC) on the test fold internally, and the complete external dataset as external evaluation. In addition to XGBoost, we include logistic regression (LR), and a multi-layer perceptron (MLP) as baseline models to contextualize performance as well as their computational complexity. Based on the results of this model benchmark, XGBoost was selected as the primary model throughout the manuscript. Detailed benchmarking results and hyperparameter settings for all models are provided in the Appendix . To improve calibration, we apply model-agnostic calibration and fit isotonic regression models on the validation set and report calibrated test set results. A recent review on evaluation criteria for prediction algorithms [ 31 ] identified three evaluation categories for predictive medical AI models: discrimination, calibration, and clinical utility. We address discriminative performance in terms of AUROC scores evaluated both on the internal test set and on an external dataset along with 95% confidence intervals derived through empirical bootstrapping with 1000 iterations. To address calibration, we show calibration curves for the internal test set. Finally, we demonstrate clinical utility through a net benefit analysis in comparison to common baselines (“refer all” and “refer none”) via decision curve analysis [ 32 ]. Our goal extends beyond simply evaluating model performance. In order to gain deeper insights into the trained models, we incorporate Shapley values into our workflow [ 33 ]. These values offer a way to assess feature importance by quantifying the individual contribution of each feature to the model’s predictions. The computational complexity and hyperparemeters setting for this approach are given in the Appendix .

Table 2 shows the predictive performance of our model across multiple neoplasms, assessed through AUROC scores on the internal and external test sets. The 95% prediction intervals offer an understanding of the reliability of these metrics. Similarly, within each figure we report the class prevalance of each neoplasm within their respective datasets, which provides context in regards the representative distribution of the populations. The MIMIC cohort shows prevalence between 0.11% to 12.38%, whereas the Korean cohort shows significantly lower prevalences between 0.04% to 3.26%. Table 2 Predictive performance results for the investigated neoplasms of diverse physiological systems Code: Description Internal AUROC (95% CI) [Prev.] External AUROC (95% CI) [Prev.] Respiratory neoplasms  C34: Lung cancer 0.800 (0.784, 0.815) [1.61%] 0.767 (0.765, 0.771) [1.83%]  C341: Upper lung cancer 0.723 (0.706, 0.750) [0.48%] 0.738 (0.723, 0.751) [0.04%]  C343: Lower lung cancer 0.855 (0.788, 0.887) [0.27%] 0.752 (0.747, 0.770) [0.04%]  C349: Unspecified lung cancer 0.792 (0.755, 0.822) [0.57%] 0.759 (0.757, 0.760) [1.72%] Urological neoplasms  C61: Prostate cancer 0.756 (0.746, 0.781) [1.27%] 0.795 (0.792, 0.797) [0.8%]  N40: Benign prostatic hyperplasia (BPH) 0.749 (0.739, 0.760) [12.38%] 0.820 (0.817, 0.821) [1.0%]  N400: BPH without symptoms 0.739 (0.727, 0.751) [9.55%] 0.828 (0.823, 0.831) [0.87%]  C679: Bladder cancer, unspecified 0.833 (0.803, 0.893) [0.18%] 0.757 (0.748, 0.762) [0.33%] Digestive neoplasms  C15: Esophageal cancer 0.818 (0.780, 0.876) [0.18%] 0.810 (0.807, 0.815) [0.26%]  C22: Liver cancer 0.808 (0.782, 0.825) [0.59%] 0.719 (0.715, 0.722) [1.43%]  C24: Biliary tract cancer 0.837 (0.756, 0.904) [0.11%] 0.706 (0.702, 0.712) [0.26%] Gynecological neoplasms  D25: Leiomyoma of uterus 0.808 (0.735, 0.854) [0.52%] 0.730 (0.727, 0.732) [3.26%]  N80: Endometriosis 0.879 (0.845, 0.907) [0.16%] 0.753 (0.750, 0.757) [1.47%] Cerebral neoplasms  C793: Brain metastases 0.738 (0.712, 0.762) [0.71%] 0.699 (0.693, 0.707) [0.13%] We list internal (MIMIC-IV) and external (ECG-View) AUROC performances with 95% confidence intervals as well as the class prevalance of the conditon in each dataset Predictive performance results for the investigated neoplasms of diverse physiological systems We list internal (MIMIC-IV) and external (ECG-View) AUROC performances with 95% confidence intervals as well as the class prevalance of the conditon in each dataset Notably, from the respiratory system, the most accurately predictable neoplasm is “C343: Lower lung cancer” with 0.855 AUROC, from the urological system “C679: Bladder cancer, unspecified” with 0.833 AUROC, from the digestive system “C24: Biliary tract cancer” with 0.837 AUROC, from the gynecological system “N80: Endometriosis” with 0.879, and from the cerebral system “C793: Brain metastases” with 0.738. For simplicity, we restrict ourselves to results achieved by the XGBoost model. In the Appendix , we present additional results for the LR and MLP baselines. All three models often perform comparably, which underlines the robustness of our findings. Across all tasks, the XGBoost model shows the overall best performance and was therefore selected for all further investigations. Extending beyond discriminative performance in terms of AUROC scores, we demonstrate three facets of model performance in Fig.  2 at the example of condition “C38: Lung cancer”. The ROC curves (left panel) align with the high predictive performance in both the internal and external test set. The calibration curve (middle panel) demonstrates good calibration, underlining the reliability of predicted probabilities. The decision curve (right panel) demonstrates clinical utility in comparison to the two baseline strategies considered. Given the low prevalences of all conditions in the dataset, both the calibration curves and the relevant part of the decision curves concentrate in the low probability threshold range. Fig. 2 Exemplary performance analysis for the condition “C34: Lung cancer” condition, showing the model’s performance across three key evaluation metrics: AUROC curves (discrimination), calibration curves (agreement between predicted and observed risks), decision curve analysis (net benefit compared to “refer all” and “refer none” strategies). Corresponding plots for all other considered conditions can be found in the Appendix Exemplary performance analysis for the condition “C34: Lung cancer” condition, showing the model’s performance across three key evaluation metrics: AUROC curves (discrimination), calibration curves (agreement between predicted and observed risks), decision curve analysis (net benefit compared to “refer all” and “refer none” strategies). Corresponding plots for all other considered conditions can be found in the Appendix Figure 3 presents the explainability results using Shapley values. Across all investigated neoplasms, age is the most important feature. Higher age values (older patients) contribute positively to the respiratory, urological, and digestive systems. In contrast, lower age values (younger patients) contribute positively to the gynecological system. The cerebral system exhibits a mix of contributions from both younger and older patients. Similarly, low QT-interval values (faster ventricular repolarization) contribute positively across all investigated neoplasms, except in cases of malignant neoplasm of the bladder, which show only a few cases with high values. Apart from gender-specific neoplasms, male sex contribute more positively overall than female sex. Fig. 3 Explainability results for the investigated neoplasms. The beeswarm plot visualizes through a single dot per feature and sample if the feature contributes positively (right hand side) or negatively (left hand side) to the model prediction. In addition, the color-coding allows to infer if a point is associated with high (red) or low (blue) feature values Explainability results for the investigated neoplasms. The beeswarm plot visualizes through a single dot per feature and sample if the feature contributes positively (right hand side) or negatively (left hand side) to the model prediction. In addition, the color-coding allows to infer if a point is associated with high (red) or low (blue) feature values For the investigated respiratory neoplasms, QT-interval and RR-interval represent the two most important ECG features. In terms of feature values, low values of the RR-interval (faster heart rates), PR-interval (shorter atrial conduction time), and QRS-duration (more efficient ventricular depolarization) generally contribute positively, with the exception of lower lung cancer, where high QRS-duration values (delayed ventricular conduction) are more significant. High values of the QRS axis (altered electrical orientation of the ventricles) also contribute positively across these neoplasms. For the investigated urological neoplasms, QRS-duration and QRS axis are the two most important ECG features. In terms of feature importance, low values of the QRS axis (altered electrical orientation of the ventricles) and P wave axis (abnormal atrial electrical orientation) contribute positively. For the investigated digestive neoplasms, male sex is the most important feature for esophageal and liver cancer. Male sex also contribute more than female sex for the biliary tract, albeit in a less pronounced fashion. There is no consistent ECG feature of high importance across all the investigated neoplasms of the system. In terms of feature value, low values of the PR-interval (indicating faster atrial conduction), QRS-duration (shorter ventricular depolarization time), T-wave-axis (altered repolarization pattern), and QT-interval (faster ventricular repolarization) contribute positively. For the investigated gynecological neoplasms, the QRS axis is the most important ECG feature. In terms of feature value importance, low values of the QT-interval (faster ventricular repolarization) contribute positively, suggesting a quicker recovery of the ventricles after each heartbeat. For the only investigated cerebral neoplasm, the most important ECG features are the QT-interval, PR-interval, QRS-duration, T-wave-axis, and RR-interval, in that order. In terms of feature value importance, low values of all of these contribute positively. Low QT-interval values (faster ventricular repolarization), low PR-interval values (shorter atrioventricular conduction), low QRS-duration (faster ventricular depolarization), low T-wave-axis values (potentially indicating quicker repolarization of the ventricles), and low RR-interval values (indicating faster heart rate), which suggest stress response and systematic inflammation associated with neoplasms. Finally, we include in the Appendix a comparison of ECG features summarized using the median and interquartile range across binary outcomes (diagnosed vs. not diagnosed). This analysis supports the validity of our approach and highlights clinically meaningful distinctions in ECG characteristics between the two groups.

Neoplasms are among the leading causes of death globally with a 2024 projection of over 2 millon new neoplasms cases and more than 600.000 neoplasms deaths in the United States alone [ 1 ]. Despite progress in medical diagnostics and treatments, timely diagnosis continues to pose a significant challenge, as many neoplasms are identified only at advanced stages. Such delays adversely affect survival rates, highlighting the pressing need for accessible, non-invasive, and cost-effective diagnostic methods [ 2 ]. Current diagnostic methods, including imaging, biopsies, and tumor biomarkers, are often invasive, resource-intensive, or inaccessible in low-resource settings [ 3 ]. These limitations highlight the necessity for innovative approaches to improve neoplasms detection and outcomes. Electrocardiograms (ECGs), long regarded as a cornerstone for diagnosing cardiovascular conditions, have shown promise beyond their traditional applications. By capturing the heart’s electrical activity, ECGs provide critical insights into cardiac rhythm and function. Recent advances have expanded their utility into non-cardiac domains, such as predicting laboratory value abnormalities [ 4 ], patient deterioration in emergency settings [ 5 ], and other systemic health indicators, as reviewed in [ 6 ]. These studies suggest that the ECG, in combination with machine learning methods, could play an important role in identifying broader physiological disruptions. The relationship between neoplasms and the cardiovascular system is well-documented, particularly in the emerging field of cardio-oncology. Neoplasms can induce subtle cardiac alterations detectable through the ECG, whether by direct invasion, paraneoplastic syndromes, or systemic effects such as inflammation and hypercoagulability [ 7 ]. Additionally, neoplasms therapies, including chemotherapy and targeted treatments, are associated with cardiotoxicity, which may lead to arrhythmias, ischemia, or myocardial dysfunction [ 8 ]. Despite these known associations, the ECG remains underutilized as a diagnostic tool for neoplasms. Nevertheless, the ability to detect malignancy-related patterns in ECG signals offers a compelling opportunity to enhance neoplasms diagnosis. This study investigates the integration of ECG features with demographic data to improve neoplasm diagnoses using tree-based machine learning models. The objective is to develop an accessible, non-invasive, and interpretable diagnostic tool to aid in detection and monitoring of neoplasms. By complementing existing diagnostic methods and addressing their shortcomings, this approach aims to enhance neoplasm outcomes and expand access to diagnostic solutions. Traditional diagnostic approaches for neoplasms rely heavily on serum biomarkers, imaging techniques, and tissue biopsies. Serum biomarkers, while offering a less invasive alternative through blood sampling, often suffer from limited sensitivity and specificity, especially for initial stages of neoplasms or tumors located in hard-to-reach anatomical areas [ 9 ]. Imaging modalities such as CT scans, MRIs, and PET scans are essential for detecting and staging neoplasms but apart of being unaccesible for large population groups [ 10 ], are resource-intensive and may not always distinguish between benign and malignant lesions with high accuracy. Tissue biopsies, considered as the gold standard for diagnosing neoplasms, are invasive procedures that carry risks such as bleeding, infection, and sampling errors, which can lead to misdiagnoses or delays in treatment. These challenges highlight the need for advanced diagnostic tools that are truly non-invasive, improving timely detection, reducing procedural risks, and supporting personalized treatment strategies. Electrocardiograms (ECG) play an important role in diagnosing and monitoring cardiovascular diseases, providing a non-invasive means to evaluate the heart’s electrical activity. Traditionally, ECG analysis has focused on detecting arrhythmias, myocardial infarctions, and other cardiac disorders through electrical signal patterns. However, recent advances have broadened its applications beyond cardiology, as highlighted in reviews such as [ 6 , 11 ], with studies demonstrating its potential for systemic health monitoring. For example, Strodthoff et al. [ 12 ] recently showcased the ability to predict a wide range of cardiac and non-cardiac neoplasms from a single ECG from a unified model. Given its non-invasive nature, affordability, and accessibility, ECG emerges as a promising tool for developing novel diagnostic models, including those targeting neoplasm-related conditions. The interactions between the cardiovascular system and neoplasms are complex and multifaceted, with neoplasms influencing cardiovascular health and vice versa. Certain neoplasms, such as those of the lung and breast, are associated with increased risks of cardiovascular complications due to tumor-induced hypercoagulability, leading to thromboembolic events like deep vein thrombosis and pulmonary embolism [ 7 ]. Neoplasms survivors, including those treated for childhood neoplasms, also face an increased risk of cardiovascular issues later in life [ 13 ]. Additionally, neoplasms treatments, including chemotherapy, radiotherapy, and targeted therapies, frequently induce cardiotoxicity, manifesting as neoplasms like heart failure, arrhythmias, and myocardial ischemia [ 8 , 14 ]. Conversely, cardiovascular conditions can affect neoplasms progression and outcomes. Chronic heart diseases, through mechanisms like reduced systemic perfusion and hypoxia, may create a microenvironment conducive to tumor growth and metastasis. Furthermore, shared risk factors, including obesity, smoking, and systemic inflammation, exacerbate both cardiovascular and oncological neoplasms, underscoring their interconnected pathophysiology [ 15 ]. These bidirectional relationships highlight the importance of integrated multidimensional approaches for diagnosing, managing, and preventing cardiovascular complications in oncology and vice versa. Cardio-oncology is an emerging multidisciplinary field that addresses the cardiovascular health of patients with cancer [ 16 ]. With improved cancer survival rates and the increasing use of cardiotoxic therapies, there is a growing need to understand, detect, and manage cardiac complications in this population. Electrocardiography (ECG) plays a critical role in this setting, offering a readily accessible tool for early detection of arrhythmias, myocardial injury, and conduction disturbances. Cancer patients are at increased risk for arrhythmias and other ECG abnormalities due to a range of factors, including direct tumor effects (e.g., cardiac infiltration or compression), therapy-induced cardiotoxicity (e.g., chemotherapy, radiotherapy, immunotherapy), paraneoplastic syndromes and immune-mediated inflammation, and a possible inherent pro-arrhythmic state, even before treatment initiation [ 17 ]. Recent literature has expanded our understanding of ECG changes in cancer patients: Case reports illustrate how cardiac metastases can imitate acute coronary syndromes (ACS). For instance, ST-segment elevation in a lung cancer patient was due to right ventricular metastasis, despite normal cardiac biomarkers [ 18 ]. A systematic review of 36 reports found that cardiac metastases often produce convex ST elevations in specific coronary territories without typical ischemic progression [ 19 ]. Mechanical effects also contribute. In a cohort of 264 lung cancer patients, the presence of J waves correlated strongly with direct tumor-heart contact [ 20 ]. Paraneoplastic and immunerelated mechanisms can also alter ECGs. One patient on immune checkpoint inhibitors developed ECG findings suggestive of myocarditis alongside immune-mediated myositis [ 21 ]. Even before treatment, cancer patients may show abnormal ECGs. A propensity-matched study comparing newly diagnosed cancer patients with surgical controls found significantly more baseline conduction delays and repolarization abnormalities in the cancer group [ 22 ], suggesting a possible inherent pro-arrhythmic state. Overall, the ECG remains a frontline diagnostic tool in cardio-oncology. Understanding its nuances in cancer patients—across various stages of disease and treatment is essential for risk stratification, monitoring, and timely intervention [ 23 – 25 ]. As the field evolves, integrating ECG findings with imaging, biomarkers, and genetic data will further enhance cardiovascular care in oncology.

This study demonstrates the potential of using ECG biomarkers for the early detection of neoplasms, offering a non-invasive, cost-effective diagnostic tool. By identifying specific ECG patterns linked to neoplastic conditions, we show how the integration of machine learning methods can bridge the gap between cardiology and oncology, uncovering novel pathways for diagnosis. The strong predictive performance and feature importance findings highlight the robustness of ECG features in distinguishing between various neoplasm types, even differentiating benign from malignant conditions. This research underscores the value of ECG in cardio-oncology, with the potential to aid in both neoplasm diagnosis and monitoring treatment-related cardiotoxicity. Despite the limitations, including potential confounding by therapy-related factors, this study paves the way for further investigation into ECG’s diagnostic capacity. Future studies can refine these findings, enhancing the accuracy and application of ECG-based neoplasm detection, ultimately improving patient care by integrating ECG monitoring into broader clinical management strategies.

Detecting neoplasms through ECG features may initially seem unconventional, as the ECG is traditionally associated with diagnosing cardiovascular conditions. However, the physiological interplay between the cardiovascular system and neoplastic processes offers a unique perspective for diagnostic innovation. Although the mechanisms linking neoplasms to ECG abnormalities are not yet fully understood, they present an intriguing avenue for further investigation. Our findings uncover specific ECG patterns that serve as distinctive markers for neoplastic conditions, suggesting underlying physiological connections that are detectable through machine learning methods. This interdisciplinary approach underscores the potential of bridging oncology and cardiology to uncover novel diagnostic pathways and improve non-invasive neoplasms diagnosis strategies. The remarkable predictive strength of a select group of ECG features emphasizes their capacity to accurately identify neoplasms from a single ECG. Consistently high AUROC values in both internal and external validations confirm the robustness of these features, even across varied cohorts. The unique patterns identified across different physiological systems highlight the interconnectedness between cardiac and oncological health. Remarkably, our approach is able to distinguish between benign and malignant neoplasms or diverse neoplasms with alike symptoms such as malignant neoplasm of prostate against benign prostatic hyperplasia, as well as leiomyoma of the uterus and endometriosis. The variation in predictive performance observed across different neoplasm types likely reflects underlying physiological and pathophysiological heterogeneity in how various cancers influence cardiac electrophysiology, as captured by the ECG. For example, neoplasms such as lower lung cancer due to their anatomical proximity to the heart or their potential to trigger paraneoplastic syndromes, systemic inflammation, or changes in autonomic regulation, may induce more pronounced alterations in ECG signals. These changes make such neoplasms more readily detectable by ECG-based models. Conversely, cancers that are located further from the thoracic cavity or that exert limited systemic effects may not manifest discernible ECG signatures, resulting in reduced model performance for those categories. In this study, age was identified as a key factor, with older patients contributing more to the most of the neoplasms except patients associated with gynecological neoplasms. This aligns with previous findings that report an increased incidence of ventricular arrhythmias linked with a worse prognosis in older neoplasms patients [ 34 , 35 ]. Additionally, our findings show that males contribute more than females across many neoplasms, which is consistent with studies showing a higher occurrence of premature ventricular contractions in male neoplasms patients [ 34 ]. Lastly, the association of lower QT-interval values across several neoplasms types mirrors findings that higher heart rates, as seen in tachycardia, are independent predictors of poor survival in neoplasms patients [ 36 ]. ECG is a valuable tool for detecting electrical abnormalities; however, it cannot directly diagnose or localize neoplasms. Accurate detection and localization require additional imaging modalities, such as echocardiography or MRI. Therefore, at this stage, we consider ECG a preliminary screening tool that can help identify abnormalities but must be complemented by imaging techniques for definitive neoplasm assessment. Changes in the ECG may serve as indicators for the presence of heart damage or abnormal heart activity by the prescence of diverse neoplasms in patient’s body, thus supporting neoplasms diagnostic and risk stratification once counfounding addressed. Nevertheless, for cardiac monitoring in oncology patients ECGs can be integrated into comprehensive cardio-oncology management strategies, where they are used for monitoring the cardiotoxicity of neoplasms treatments. This includes regular ECG checks alongside imaging modalities and cardiac biomarkers such as troponins and NT-proBNP, which help assess treatment-related cardiovascular risks. By monitoring ECG patterns during therapy, especially for high-risk drugs, clinicians can early detect signs of cardiotoxicity. These findings are invaluable in guiding clinical decisions, such as adjusting drug dosages, initiating cardioprotective strategies, or providing early interventions to mitigate further heart damage [ 16 , 37 ]. Ultimately, this integrated approach helps balance the efficacy of neoplasms therapies with the safety of the heart, improving the overall quality of life for patients while maintaining treatment effectiveness. First, regarding patient stratification, we acknowledge that external variables may introduce confounding effects, such as newly identified diagnoses and preexiting conditions. Since the ICD-10 codes in the dataset reflex a mix of these, the model predictions may partially capture therapy-induced cardiac changes, such as cardiotoxic effects of treatment, rather than signals solely related to the neoplasm itself. Resolving this ambiguity is an important next step for follow-up studies. Second, it is worth noting that prior work [ 12 ] has investigated label correlations for the MIMIC-IV dataset and found no significant label correlations. This defutes the potential claim that models detect other conditions commonly co-occurring with neoplasms. This aligns with very well with [ 22 ], which clearly demonstrates the feasibility of finding cardiac abnormalities in newly diagnosed cancer patients. Many ECG changes are non-specific and may arise from non-neoplasmsous conditions, such as electrolyte imbalances or ischemic heart disease, making it difficult to attribute abnormal ECG patterns to neoplasms alone. Future research should investigate how ECG abnormalities vary across age groups and distinguish these from typical age-related ECG changes [ 38 ]. Moreover, exploring the causal relationships between ECG patterns and neoplasms will be crucial [ 39 ]. Studies focusing on raw ECG waveforms, including external validation, could further enhance diagnostic accuracy [ 5 , 12 ]. The potential of raw ECG waveforms to outperform traditional ECG features in diagnostic tasks underscores the importance of continuing to refine this method for better diagnostic precision.