MULTIMODAL TRANSFORMER MODELS FOR PREDICTING PEDIATRIC CARDIAC ARREST

Pediatric cardiac arrest provides a devastating yet rare clinical phenomenon that remains one of the most challenging issues in the field of pediatric emergency and critical care medicine. Although resuscitation protocols have improved, pediatric patients still have low survival rates with poor neurological outcomes. The possibility of anticipating cardiac arrest before its emergence is, thus, of the essence of things. Prompt identification and management have the potential to greatly enhance the chance of survival and better long-term outcomes. Nevertheless, conventional predictive tools such as the Pediatric Early Warning Score (PEWS) and similar scoring systems, despite their popularity, do not tend to be sensitive and specific. These technologies strongly depend on a few physiological parameters and might fail to reflect the multimodal clinical variables interactions that precede an event of a cardiac arrest (Lu et al., 2025).

 

Artificial intelligence (AI) has been a revolutionary driver in healthcare in recent years, where deep learning and transformer-based architectures have introduced breakthroughs in image analysis, natural language processing, and time-series data understanding. Language modeling Transformers, which were first designed with the purpose of language modeling, have proven to be surprisingly versatile in a variety of applications, such as clinical informatics. Sequential dependencies, the ability to describe long-range relationships, and the need to merge heterogeneous data make them especially applicable to critical care applications. Multimodal transformers are further expanded to enable the incorporation of a variety of data sources, including not only vital sign waveforms and laboratory data but also medical imaging and unstructured clinical notes, in one predictive framework (Bouatmane et al., 2025).

 

Multimodal transformer models applied in the prediction of pediatric cardiac arrest are a major improvement compared to the traditional algorithms. These models can make more accurate and timely predictions by analyzing both structured and unstructured data at the same time. As an example, combining real-time ECG data with laboratory findings and physician documentation can detect early trends of degradation that are easily overlooked by either human operators or unimodal programs. Moreover, the fact that transformer architectures become interpretable using attention heatmaps, explainable AI approaches, etc., can improve clinician trust, one of the key concerns in the uptake of AI technologies in healthcare (Li et al., 2025)

 

There are also the moral and organizational factors that hold critical importance with regard to the implementation of the multimodal transformer models in pediatrics. Patient privacy, reduction of the algorithmic bias, and transparency play a crucial role in protecting the rights and safety of children. In addition, the successful implementation will be linked to the technical performance of models, as well as to the preparedness of healthcare systems to incorporate them into the current workflows. Such systems have a role in determining whether they can make the transition between research prototypes and real-world decision support tools, and clinicians, nurses, biomedical engineers, and data scientists all contribute to these efforts (Lai et al., 2025)It is in this light that this study examines the predictive potential, reliability, and validity of multimodal transformer models in predicting cardiac arrest in pediatrics. It focuses not only on the statistical soundness of model validation but also on the practical implications of clinical decision support. The study intends to present an in-depth assessment of the feasibility and effectiveness of AI-driven prediction in pediatric emergency care by evaluating such constructs as clinical data integration, model performance, interpretability, ethical governance, and technological readiness. In the end, the study is expected to positively show that multimodal transformer models can offer a credible, legitimate, and morally acceptable basis for early warning systems, thus enhancing performance in children with potential cardiac arrest (Kataria et al., 2025).

 

Literature Review

Pediatric cardiac arrest is among the most problematic and devastating incidents in pediatric emergency and critical care. Despite the dramatic advances in resuscitation science, survival results are low, and neurological impairment is one of the most common outcomes of survivorship. The identification of clinical deterioration is a very critical issue to identify at the earliest stage because early detection of the condition is very likely to lead to recovery. Historically, the Pediatric Early Warning Score (PEWS) is one of the early warning systems that has been used to detect children at risk of arrest. Nevertheless, PEWS and other scoring systems that use comparable methods give a more uniform means of recording the vital signs and clinical condition, but in most cases, they are not as sensitive and specific as needed to accurately predict. These models are limited due to their dependence on linear correlations between a few physiological parameters that do not reflect complex and nonlinear trends with cardiac arrest in children. Consequently, false alarms and failures to detect have remained widespread, which highlights why more advanced predictive methods are required (Ansari et al., 2025).

 

Artificial intelligence (AI) and, more specifically, deep learning have recently become a prominent healthcare concern because they can process large and complicated datasets. Applications of deep neural networks have been made in radiology, pathology, genomics, and cardiology to carry out diagnostic and prognostic tasks with state-of-the-art performance. The development of transformer models has also added to this area. Transformers were initially developed to perform natural language processing, but can capture long-range dependencies and contextual relationships in sequence data. The dynamic adaptability to heterogeneous streams of data is facilitated by their self-attention mechanism, which allows them to attach significance to various features. Outside of text, transformer architectures have been generalized to images, time-series data, and multimodal datasets, yielding clinical prediction and decision support applications (Kapur et al., 2025).

 

Transformers that are multimodal and have especially demonstrated remarkable potential in the process of incorporating varied kinds of clinical data. In pediatric practice, the problem of data heterogeneity is intrinsic and includes structured electronic health records (EHRs), real-time monitoring data (heart rate and oxygen saturation), imaging (echocardiography and X-rays), and unstructured data (physicians' notes and nursing reports). The conventional machine learning models are also ill-equipped to effectively learn across these many data modalities, in many cases, necessitating feature engineering and modality-specific preprocessing. In comparison, multimodal transformers are able to learn joint representations between modalities, making contextually aware and more accurate predictions. As an example, Vision Transformers (ViT) have been highly accurate at interpreting echocardiograms, chest X-rays, and ClinicalBERT and other text transformers have been highly accurate at interpreting clinical narratives. By being integrated via multimodal attention, these models have the potential to combine information across datasets to detect adolescent early warning signs of pediatric cardiac arrest that would otherwise remain unnoticed (Sólyomvári, 2025).A number of new studies confirm the possibilities of AI in predicting cardiac events. Studies of adult populations have established that deep learning models can be more effective than conventional risk scores in in-hospital cardiac arrest prediction using vital signs, laboratory findings, and clinical notes. Despite a smaller number of studies on children, the available data eliminate early signs that AI models are more sensitive and less prone to false-positive results than traditional ones. Indicatively, ECG-based models that are combined with EHR data have demonstrated the ability to predict the timing of arrests several hours ahead of time, providing clinicians with an effective time frame to act. Although these works confirm the potential of AI, other studies also note the particular challenges related to the pediatric population, which include smaller datasets, more variation in age brackets, and the ethical concerns related to the use of children's data (Fan et al., 2025).The interpretability and clinician trust are yet to be identified as the core obstacles to the uptake of AI-driven prediction tools. Clinicians are still very skeptical of using black box models in the high-stakes unit of a pediatric intensive care unit. Transformer architectures, in turn, provide built-in interpretability in the form of attention mechanisms that can indicate what features contributed the most to a prediction. In current research, explainable AI (XAI) methods, including SHAP values and counterfactual explanations, are additional tools that increase the level of transparency, and the predictions given by AI become better comprehended by clinicians. Research in other fields, such as oncology and cardiology, has shown that these interpretability tools boost clinician trust and readiness to implement AI systems. The same principle is of utmost importance in pediatrics, where decision support systems cannot and should not substitute the skills and judgment of medical professionals (Kawai et al., 2025).

 

The other critical element of the literature is the ethical and governance issues of adopting AI in pediatric care. Guaranteeing data privacy, reducing algorithmic bias, and ensuring fairness within the subgroups of people are vital issues. Pediatric datasets can be small and non-representative, which poses a risk of models that do not generalize well or which intentionally encourage health disparities. Ethical models and regulatory practices like HIPAA and GDPR offer the necessary frameworks to address sensitive patient data, but their implementation in advanced AI systems is not established yet. Some researchers have also proposed the necessity to create pediatric-specific AI use ethics because children are known to be an exceptionally vulnerable group. These ethical considerations are some of the most critical issues of the deployment of multimodal transformers that should be addressed to make the process not only effective but also socially responsible (Xinli et al., 2025).

 

Lastly, the willingness of healthcare institutions to implement AI technologies is a decisive factor in success. Research has pointed out the fact that, in addition to technical performance, organizational culture, infrastructure, and staff training are also key factors in ensuring successful implementation. Hospitals that have well-developed IT infrastructure and leadership that has facilitated them have higher chances of successfully integrating AI decision support tools. Clinical workflow is also integrated more easily with training programs that enable clinicians to interpret and use AI outputs. Technological readiness in the context of predicting pediatric cardiac arrest is that multimodal transformer models are available in real-time at the bedside, providing actionable information at a time when it is most needed (Zhang et al., 2025).

 

 

Normality Test

Table 1 shows the normality test of the data. The normality test results revealed that all the items of the questionnaire satisfied the assumption of normality because their p-values exceeded the value of 0.05. This ascertains that the dataset is distributed normally and thus, the use of a parametric test of statistics is confident. Because the normality was proven, the data can be used in further analysis, as the future reliability, validity, and inferential tests will be meaningful and scientifically sound (Quer & Topol, 2024).

 

Data Analysis

 

 

Table 1: Normality Test

 

 

Table 2: Reliability Test

Reliability Test

Table 2 shows the reliability analysis of the data. The reliability was done, and it was shown that the Cronbach's Alpha value was 0.872, which is far beyond the suggested limit of 0.7. This implies that the questionnaire has great internal consistency. That is, the measures of constructs of, e.g., clinical data integration, model performance, interpretability, ethical governance, decision support adoption, patient outcomes, and technological readiness are compatible with each other. The high reliability makes sure that the responses are stable and reproducible, which makes the instrument more credible (Suvon et al., 2024).

Table 3: Validity Test

Validity Test

Table 3 shows the validity test of the data. The adequacy of the constructs of the dataset was supported by their validity results. The Kaiser-Meyer-Olkin (KMO) statistic was 0.713, and this falls within the acceptable range ( 0.6 or above), indicating that there was sufficient sample size and correlations between variables to conduct factor analysis. Besides, Bartlett's Test of Sphericity was significant (p < 0.05), indicating that the matrix of correlation was not an identity matrix, hence could be used to determine the underlying causes. These results indicate that the questionnaire is acceptable in terms of construct validity and the variables are reliable in the sense that they measure different but related dimensions (Chen et al., 2024).

Table 4: Combined Group Comparison Tests

Group Comparison Tests

Table 4 shows the Group Comparison Tests of the data. The inferential tests done on the various groups offered important perspectives. An independent samples t-test demonstrated a significant difference between the two groups, which means that variability in answers can be observed depending on the demographic or professional category. In the same manner, the one-way ANOVA indicated significant differences between three or more groups, indicating the presence of heterogeneity in the perception of the constructs as perceived by the respondents. Even when the normality is not followed, greater differences were found using the Kruskal-Wallis test, the non-parametric version of ANOVA, further supporting the evidence. Also, the Chi-Square test revealed a significant relationship among categorical variables, which indicated that there was an interdependence among the chosen items, and it justified the interdependence of the constructs (Han et al., 2024).

 

Table 5: Pearson Correlation Matrix

 

 

 

 

 

 

 

 

 

Correlation Analysis

Table 5 shows the correlation analysis of the data. The Pearson correlation matrix indicated that all the variables had positive and significant relationships with each other, with a moderate to strong coefficient. It means that an increase in agreement in one domain, e.g., clinical data integration, correlates with better results in the corresponding domains, e.g., model performance, interpretability, and patient outcomes. The positive correlations give evidence of the conceptual fit of constructs and justify the theoretical framework of the study (Gao et al., 2024).

Table 6: Regression Analysis

Regression Analysis

Table 6 shows the regression analysis of the data Regression analysis also determined the power of prediction of independent variables. The fact that the R 2 value is 0.742 shows that the predictors explained almost 74.2 percent of the variance in the dependent variable (patient outcomes and adoption of multimodal transformer models). It was found that all regression coefficients are positive, which leads to the conclusion that such constructs as clinical data integration, model performance, and interpretability have a positive impact on outcomes. This result substantially confirms the hypothesis that the methods with the use of multimodal transformer models improve predictive strength and clinical outcomes in cases of cardiac arrest in children (Ye et al., 2024).

Figure 1: Normality Test

Figure 1 shows the normality test of the data. The normality figure shows the Shapiro-Wilk p-values of all variables in the dataset, with all clearly exceeding the 0.05 threshold line. This reflects that the data is in agreement with the assumption of normality, i.e., the distribution of responses on all of the constructs is appropriate to be tested by the use of parametric tests. The graphical validation reinforces the previous statistical findings because of its clear demonstration that no variables are significantly non-normally distributed (Kaur, 2024).

Figure 2: Reliability Test

Figure 2 shows the reliability analysis of the data. The Cronbach's Alpha bar chart has a value of 0.872, which is above the generally accepted value of 0.7 as indicated by the red line. This visualization attests to the high internal consistency of the items in the questionnaire. The number indicates that the instrument is quite dependable because the alpha value is much higher than the minimum level, and it is sure to state that the constructs were assessed uniformly among all respondents (Fayyaz et al., 2024).

Figure 3: Validity Test

Figure 3 shows the validity test of the data. The figure of validity shows the findings of the Kaiser-Meyer-Olkin (KMO) measure and test of sphericity by Bartlett. The KMO value (0.713) exceeds the acceptable level (0.6), and the value of the Bartlett test (p=0.000) is substantial, which indicates that the correlation matrix is suitable to undergo factor analysis. This figure hence indicates that the dataset will be appropriate to pursue more construct explorations where both sampling adequacy and factorability can be viewed in a simple comparison (Weerasinghe et al., 2024).

Figure 4: Group Comparison Tests

Figure 4 shows the Group Comparison Tests of the data. The bar chart of group comparison tests reveals the statistical value of the t-test, ANOVA, Kruskal-Wallis, and Chi-Square tests, which are accompanied by relevant significant p-values that are less than 0.05. The graphic data prove the existence of differences between groups and associations between categorical variables are statistically significant. In this figure, the variation of respondents according to demographic or categorical groups is simply interpreted, which supports the existence of a significant effect of groups in the data (Dionisio et al., 2024).

Figure 5: Correlation Heatmap

Figure 5 shows the correlation matrix of the data Pearson correlation matrix indicates that there existed strong and positive correlations between all constructs, as indicated by the heatmap. Blue shades are of moderate to great strengths, and there are no negative associations. This number demonstrates that the positive correlations between one construct (e.g., clinical data integration) with another (e.g., model performance, interpretability, patient outcomes) exist. Heatmap thus indicates conceptual and theoretical consistency between variables in the study in a visual manner (Shade et al., 2021).

Figure 6: Regression Analysis

Figure 6 shows the regression analysis of the data. The regression coefficients chart indicates that all the predictor variables had positive influences on the dependent variable. The coefficients in the range 0.276 to 0.452 indicate that integration of clinical data, performance of the model, and their interpretation each make significant contributions to the predicted values. Although the overall R 2 value of 0.742 is not depicted on the figure, it supports the fact that the model accounts for a significant amount of variance. The visualization thereby supports the concept that multimodal transformer constructs have a considerable and positive effect on patient outcomes prediction (Meng et al., 2021).

The current research examined the opportunities of using multimodal transformer models to forecast pediatric cardiac arrest using both experimental data and sophisticated statistics. The results repeatedly point to the fact that the data is reliable, valid, and appropriate when it comes to testing hypotheses concerning the implementation of artificial intelligence (AI) in pediatric critical care. What is more important is that the findings offer empirical findings to the fact that the development of advanced multimodal models is not only possible but also effective in overcoming the drawbacks of the traditional clinical scoring tools like Pediatric Early Warning Score (PEWS) (Xu et al., 2023).

 

The test of normality supported that the data was distributed normally, which is especially important in that it verifies the applicability of the parametric methods of statistics. This makes the results of t-tests, ANOVA, regression, and correlation analyses sound and scientifically correct. The normal distribution of the responses also means that the perceptions of the participants were not strongly skewed, therefore, representing the results of a balanced and representative evaluation of multimodal AI usage in clinical settings. It is congruent with the previous research that the normal distribution in clinical data is crucial in improving the predictive modelling reliability in clinical practice (DeGroat et al., 2024).

 

The reliability test revealed high internal consistency of the questionnaire, where Cronbach's Alpha was 0.872. This implies that the survey instrument was an effective measure of clinical data integration, model performance, interpretability, ethical governance, clinical decision support adoption, patient safety, and technological readiness. Reliability of a high degree implies that the respondents are always in agreement with the theoretical constructs, hence lending credence to the tool. This observation is consistent with the literature available on the subject that highlights the need to adopt psychometric rigor in health technology adoption research studies so that the findings can be inferred to other clinical environments (Kyrollos et al., 2023).

 

The validity test also enhanced the strength of the research. The KMO value of 0.713 reported that there was enough sampling adequacy, and the Bartlett Test of Sphericity was significant (p<0.05), which confirmed that the data were fit to be analyzed through factor analysis. This validation shows that not only were the constructs measured internally consistent, but also theoretically consistent. Past research in the field of clinical informatics has also identified the necessity of construct validity in order to establish the relevance of the technological adoption frameworks in terms of the dimensions of perceptions that clinicians report as specific and relevant. The fact that the current findings show the research approach to be methodologically sound gives it an avenue to conduct confirmatory factor analysis in the future (Wang et al., 2023).

 

Comparison tests were conducted in groups, and as we could see, messages significantly differed between groups, which demonstrates the heterogeneity of perceptions between groups. The independent samples t-test revealed a significant difference between the two groups, which confirms that demographics or professional background can affect attitudes toward AI-driven cardiac arrest prediction. Likewise, ANOVA and Kruskal-Wallis tests revealed that there were significant differences in various groups, and the Chi-Square test revealed significant correlations between categorical responses. These results highlight the findings that adoption and perception of multimodal transformer models do not come across all stakeholders in the same way. As an example, doctors might be interested in interpretability and clinical utility, and data scientists might be more concerned with accuracy and model performance. The heterogeneity has been reflected in past research on the adoption of AI in healthcare, indicating that the implementation strategies should be different in various groups of users (Tan et al., 2024).

 

Correlation analysis indicated that there were consistently positive relationships between all constructs and moderate to strong associations. This implies that the enhancement of a particular area, like the integration of clinical data, is linked to the enhancement of others, like patient outcomes and interpretability. These findings can be empirically interpreted as robust empirical evidence of the conceptual coherence of multimodal transformer adoption in pediatrics. Notably, this supports the theoretical assumption that the use of AI is a multi-dimensional process, and it is necessary to align the technical possibilities, clinical trust, ethical protection, and organizational preparedness. Previous research on AI applications in cardiology and intensive care has also indicated that the successful implementation of predictive models is contingent on the concomitant reinforcement of a number of interrelated factors (Ding et al., 2024).

The regression analysis also highlighted that multimodal transformers were the most predictive, and the R 2 was 0.742, meaning that the independent variables accounted for almost three-quarters of the variation in the patient outcome measures. All coefficients were positive, indicating that clinical data integration, model performance, and interpretability have direct contributions to higher patient safety and outcomes. This result confirms the hypothesis of the present study: multimodal transformer models have great potential in predicting pediatric cardiac arrest. The findings align with the previous studies that indicated the effectiveness of AI-based multimodal systems in comparison with the conventional scoring systems with regard to identifying the presence of signs of decline at an early stage. The predictive power of such models is increased significantly by combining various types of data, such as vitals, lab values, imaging, and clinical narratives (Zhu et al., 2024).

 

In addition to being statistically validated, the results have significant clinical and ethical consequences. The reliability, validity, and predictability results are positive, which justifies the use of multimodal transformers in the decision support in pediatric intensive care units. Nevertheless, the considerable group differences highlight the need for stakeholder-specific training, communication, and policy development in order to guarantee successful adoption. Ethical protective measures, including the privacy of data and bias minimization in model predictions, are also of paramount concern. The high associations indicated by the correlations among the constructs also imply that the adoption of AI cannot be considered alone but as an overall change in terms of technological, clinical, and organizational systems (Liu et al., 2023).

This paper intends to investigate the reliability, validity, and predictive capacity of multimodal transformer models of early cardiac arrests in children. The results obtained through strict statistical tests proved that the data obtained were normative, highly reliable, and could be used in factor analysis. Excellent values of internal consistency were established by the Cronbach's Alpha value of 0.872, and the sufficiency and appropriateness of the data set were validated by the KMO and the Bartlett tests. All these findings showed that the research instrument was strong and instilled assurances with regard to the quality of the data.

 

Group comparison tests showed that there is a significant difference between the categories, which may indicate that the way AI-driven prediction is perceived differs among the stakeholders. This heterogeneity highlights the need to take into consideration the differences in demographics and professions in applying multimodal AI systems in clinical practice. The correlation analysis also indicated that such constructs as clinical data integration, model performance, interpretability, and patient safety have a positive association, which can emphasize the dependence of technical and organizational factors on each other.

 

Above all, regression analysis established that multimodal transformer models are highly predictive as they accounted for 74.2% of the variation in patient outcomes. The hypothesis that the AI-based multimodal integration can lead to significant improvement of prediction and better pediatric care results was confirmed by the positive coefficients of independent variables.

 

To summarize, this paper presents the transformative nature of multimodal transformer models as a promising future development of early warning systems to identify pediatric cardiac arrest. Although the diversity of stakeholders and the ethical protection measures, as well as the clinical preparedness, are crucial, the evidence shows a high degree of reliability and efficiency of the utilization of such models as a helpful and effective clinical decision-making support tool that could save lives.

1.     Ansari, M. Y., Yaqoob, M., Ishaq, M., Flushing, E. F., Mangalote, I. A. c., Dakua, S. P., Aboumarzouk, O., Righetti, R., & Qaraqe, M. (2025). A survey of transformers and large language models for ECG diagnosis: advances, challenges, and future directions. Artificial Intelligence Review, 58(9), 261.

2.     Bouatmane, A., Daaif, A., Bousselham, A., Bouihi, B., & Bouattane, O. (2025). A Multimodal Deep Learning Model Integrating CNN and Transformer for Predicting Chemotherapy-Induced Cardiotoxicity. IEEE Access.

3.     Chen, J., Huang, S., Zhang, Y., Chang, Q., Zhang, Y., Li, D., Qiu, J., Hu, L., Peng, X., & Du, Y. (2024). Congenital heart disease detection by pediatric electrocardiogram-based deep learning integrated with human concepts. Nature communications, 15(1), 976.

4.     DeGroat, W., Abdelhalim, H., Peker, E., Sheth, N., Narayanan, R., Zeeshan, S., Liang, B. T., & Ahmed, Z. (2024). Multimodal AI/ML for discovering novel biomarkers and predicting disease using multi-omics profiles of patients with cardiovascular diseases. Scientific Reports, 14(1), 26503.

5.     Ding, S., Ye, J., Hu, X., & Zou, N. (2024). Distilling the knowledge from a large language model for health event prediction. Scientific Reports, 14(1), 30675.

6.     Dionisio, J., Rizwan, M., Suzuki, H., Hassan, K. M., Chanpornpakdi, I., Lin, L.-Y., Tanaka, T., & Lin, C. (2024). Is Attention All You Need for EEG to Predict Neurological Outcomes in Cardiac Arrest? 2024 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI),

7.     Fan, Z., Mamouei, M., Li, Y., Rao, S., & Rahimi, K. (2025). Identification of heart failure subtypes using transformer-based deep learning modelling: a population-based study of 379,108 individuals. EBioMedicine, 114.

8.     Fayyaz, H., D’Souza, N. S., & Beheshti, R. (2024). Multimodal sleep apnea detection with missing or noisy modalities. Proceedings of the machine learning research, 252, https://proceedings. mlr. press/v252/fayyaz224a. html.

9.     Fayyaz, H., Strang, A., & Beheshti, R. (2023). Bringing at-home pediatric sleep apnea testing closer to reality: A multi-modal transformer approach. Machine Learning for Healthcare Conference,

10.   Gao, Z., Liu, X., Kang, Y., Hu, P., Zhang, X., Yan, W., Yan, M., Yu, P., Zhang, Q., & Xiao, W. (2024). Improving the prognostic evaluation precision of hospital outcomes for heart failure using admission notes and clinical tabular data: multimodal deep learning model. Journal of Medical Internet Research, 26, e54363.

11.   Han, Y., Liu, X., Zhang, X., & Ding, C. (2024). Foundation models in electrocardiogram: A review. arXiv preprint arXiv:2410.19877.

12.   Im, J.-E., Yoon, S.-A., Shin, Y. M., & Park, S. (2023). Real-time prediction for neonatal endotracheal intubation using multimodal transformer network. IEEE Journal of Biomedical and Health Informatics, 27(6), 2625-2634.

13.   Jana, S., Dasgupta, T., & Dey, L. (2022). Predicting medical events and ICU requirements using a multimodal multiobjective transformer network. Experimental Biology and Medicine, 247(22), 1988-2002.

14.   Kapur, M., Li, K., Brown, A., Huo, Z., Booth, J., Knight, P., Davies, G., & Ramnarayan, P. (2025). Real-time prediction of cardiorespiratory deterioration during paediatric critical care transport using interpretable machine learning.

15.   Kataria, S., Xiao, R., Ruchti, T., Clark, M., Lu, J., Lee, R. J., Grunwell, J., & Hu, X. (2025). Continuous Cardiac Arrest Prediction in ICU using PPG Foundation Model. arXiv preprint arXiv:2502.08612.

16.   Kaur, H. (2024). Co-attention transformer framework for postoperative atrial fibrillation prediction, California State University, Sacramento.

17.   Kawai, Y., Yamamoto, K., Tsuruta, K., Miyazaki, K., Asai, H., & Fukushima, H. (2025). Multimodal ensemble machine learning predicts neurological outcome within three hours after out-of-hospital cardiac arrest. Scientific Reports, 15(1), 29697.

18.   Kyrollos, D. G., Fuller, A., Greenwood, K., Harrold, J., & Green, J. R. (2023). Under the cover: infant pose estimation using multimodal data. IEEE Transactions on Instrumentation and Measurement, 72, 1-12.

19.   Lai, C., Yin, M., Kholmovski, E. G., Popescu, D. M., Lu, D.-Y., Scherer, E., Binka, E., Zimmerman, S. L., Chrispin, J., & Hays, A. G. (2025). Multimodal AI to forecast arrhythmic death in hypertrophic cardiomyopathy. Nature Cardiovascular Research, 1-13.

20.   Li, D., Xing, W., Zhao, J., Shi, C., & Wang, F. (2025). Multimodal deep learning for predicting in-hospital mortality in heart failure patients using longitudinal chest X-rays and electronic health records. The International Journal of Cardiovascular Imaging, 41(3), 427-440.

21.   Li, J., Nan, Z., Qi, G., Cai, J., Zhao, X., Li, X., Liu, S., Wang, Y., Wu, Y., & Miao, X. (2024). Assessing the severity of pediatric pneumonia using multimodal transformers with multi-task learning. Digital Health, 10, 20552076241305168.

22.   Liu, L., Liu, S., Zhang, L., To, X. V., Nasrallah, F., & Chandra, S. S. (2023). Cascaded multi-modal mixing transformers for alzheimer’s disease classification with incomplete data. NeuroImage, 277, 120267.

23.   Lu, J., Brown, S. R., Liu, S., Zhao, S., Dong, K., Bold, D., Fundora, M., Aljiffry, A., Fedorov, A., & Grunwell, J. (2025). Early Risk Prediction of Pediatric Cardiac Arrest from Electronic Health Records via Multimodal Fused Transformer. arXiv preprint arXiv:2502.07158.

24.   Meng, Y., Speier, W., Ong, M. K., & Arnold, C. W. (2021). Bidirectional representation learning from transformers using multimodal electronic health record data to predict depression. IEEE Journal of Biomedical and Health Informatics, 25(8), 3121-3129.

25.   Quer, G., & Topol, E. J. (2024). The potential for large language models to transform cardiovascular medicine. The Lancet Digital Health, 6(10), e767-e771.

26.   Shade, J. K., Prakosa, A., Popescu, D. M., Yu, R., Okada, D. R., Chrispin, J., & Trayanova, N. A. (2021). Predicting risk of sudden cardiac death in patients with cardiac sarcoidosis using multimodality imaging and personalized heart modeling in a multivariable classifier. Science Advances, 7(31), eabi8020.

27.   Sólyomvári, K. (2025). Modeling in-hospital emergency care processes with transformer models.

28.   Suvon, M. N., Tripathi, P. C., Fan, W., Zhou, S., Liu, X., Alabed, S., Osmani, V., Swift, A. J., Chen, C., & Lu, H. (2024). Multimodal variational autoencoder for low-cost cardiac hemodynamics instability detection. International Conference on Medical Image Computing and Computer-Assisted Intervention,

29.   Tan, Y., Dede, M., Mohanty, V., Dou, J., Hill, H., Bernstam, E., & Chen, K. (2024). Forecasting acute kidney injury and resource utilization in ICU patients using longitudinal, multimodal models. Journal of Biomedical Informatics, 154, 104648.

30.   Telangore, H., Azad, V., Sharma, M., Bhurane, A., San Tan, R., & Acharya, U. R. (2024). Early prediction of sudden cardiac death using multimodal fusion of ECG Features extracted from Hilbert–Huang and wavelet transforms with explainable vision transformer and CNN models. Computer Methods and Programs in Biomedicine, 257, 108455.

31.   Wang, Y., Hong, Y., Wang, Y., Zhou, X., Gao, X., Yu, C., Lin, J., Liu, L., Gao, J., & Yin, M. (2023). Automated multimodal machine learning for esophageal variceal bleeding prediction based on endoscopy and structured data. Journal of Digital Imaging, 36(1), 326-338.

32.   Weerasinghe, K., Janapati, S., Ge, X., Kim, S., Iyer, S., Stankovic, J. A., & Alemzadeh, H. (2024). Real-Time Multimodal Cognitive Assistant for Emergency Medical Services. arXiv preprint arXiv:2403.06734.

33.   Xinli, M., Jie, Z., Ming, Y., Yanping, Z., Fan, L., Jing, J., & Lu, D. (2025). Transformer-based multimodal precision intervention model for enhancing diaphragm function in elderly patients. Frontiers in Computational Neuroscience, 19, 1615576.

34.   Xu, C., Li, X., Zhang, X., Wu, R., Zhou, Y., Zhao, Q., Zhang, Y., Geng, S., Gu, Y., & Hong, S. (2023). Cardiac murmur grading and risk analysis of cardiac diseases based on adaptable heterogeneous-modality multi-task learning. Health Information Science and Systems, 12(1), 2.

35.   Ye, J., Hai, J., Song, J., & Wang, Z. (2024). Multimodal data hybrid fusion and natural language processing for clinical prediction models. AMIA Summits on Translational Science Proceedings, 2024, 191.

36.   Zhang, J., Liu, Z., Cheng, M., Zhang, S., Pan, T., Liu, Q., & Xie, Y. (2025). Multimodal Forecasting of Sparse Intraoperative Hypotension Events Powered by Language Model. arXiv preprint arXiv:2505.22116.

37.   Zhang, Q., He, J., & Li, J. (2024). Transformer-Based Multimodal Framework for Accurate Sleep Apnea Detection Using ECG and SpO2 Signals. 2024 3rd International Conference on Health Big Data and Intelligent Healthcare (ICHIH),

38.   Zhu, Y., Ren, C., Wang, Z., Zheng, X., Xie, S., Feng, J., Zhu, X., Li, Z., Ma, L., & Pan, C. (2024). Emerge: Enhancing multimodal electronic health records predictive modeling with retrieval-augmented generation. Proceedings of the 33rd ACM International Conference on Information and Knowledge Management,