In this study, we used 2 databases: the MIMIC-III database [13] and eICU-CRD [14]. The MIMIC-III database, developed by the Massachusetts Institute of Technology Laboratory for Computational Physiology, contains data from 53,432 adult patients and 8100 neonatal patients admitted between 2001 and 2012. The database includes demographic data, vital signs, medications, laboratory measurements, physician orders, procedure codes, diagnosis codes, imaging reports, hospitalization duration, and survival data, among others.

The eICU-CRD is a multicenter intensive care unit (ICU) database that covers data from more than 200,000 ICU admissions across 208 hospitals in the United States between 2014 and 2015. The data were collected through Philips eICU program, which includes vital sign measurements, nursing documentation, severity of illness scores, diagnostic information, treatment details, and more.

We randomly allocated 70% (2844/4063) of the dataset for the training set to develop the model, with10% (406/4063) as the validation set to adjust and determine the hyperparameters of the model, and the remaining 20% (813/4063) as the test set to evaluate the model performance; this portion was not used in the model training.

This study used publicly available critical care databases, including the MIMIC-III and the eICU-CRD. Both databases received institutional review board approval from the hospitals that originally contributed the data, and informed consent was obtained at the time of data collection. All patient records were deidentified in accordance with the Health Insurance Portability and Accountability Act (HIPAA). Access to the databases was granted through the PhysioNet credentialing process. Therefore, according to institutional policies regarding secondary research using publicly deidentified data, additional ethics approval from our institution was not required.

As shown in Figure 1, let the given set of patients be P = {p₁, p₂, ..., p_n} feature vector V_t = [hr, rr, sbp, dbp, map, spo2]. V_t represents the vital sign monitoring data of patient p at time point t, including “heart rate” and “systolic blood pressure,” among others. These data were continuously collected during the hospitalization period T, T = {t₁, t₂, …, t_j} and t_i–_ti–j = 5 minutes. Then, represents the multivariate time-series feature vectors of the patient at all time points during hospitalization. Therefore, the CA prediction task can be formally described as follows: given the multivariate time-series feature vector V for patient p in the hospitalization interval T, the goal of the prediction task is, for any time point t_i∈T, to use the 24 multivariate time-series feature vectors from the time window (t–24 and t) to predict the probability of a CA event occurring within the time window (t and t+12) using the TrGRU model.

This section presents the evaluation results of the model prediction performance, which were evaluated using accuracy, sensitivity, specificity, AUROC, AUPRC, and false alarm rate. Additionally, the performance of the proposed model was compared with that of the baseline models.

In the MIMIC-III dataset, several models used in existing studies were compared with the proposed model to evaluate its performance in predicting CA events. As shown in Figure 5, the proposed model achieved an AUROC of 0.957 and an AUPRC of 0.949, significantly outperforming the baseline models. In addition, other performance metrics were also compared to fully assess the validity of the model in relevant clinical contexts. As shown in Table 1, the proposed model achieved an accuracy of 0.904, a sensitivity of 0.859, a specificity of 0.933, and a false alarm rate of 0.067, all of which outperform the baseline models, yielding the best performance among these.

We used a meta-learning and rapid adaptation–based approach to perform the cross-dataset CA prediction task. On the MIMIC-III dataset, a model was pretrained through multitask meta-learning. After pretraining, the lower-layer parameters of the model were frozen on the eICU-CRD, and only the last 2 fully connected layers were fine-tuned. Domain adaptation was achieved through adaptive training. The experimental results showed that the model achieved a sensitivity of 0.813, an AUROC of 0.920, and an AUPRC of 0.848 on the independent eICU-CRD, demonstrating significantly better cross-dataset adaptation ability compared to current models. These results indicated that the framework effectively used knowledge from the source domain to enhance the few-shot learning performance in the target domain.

We also evaluated the ability of the model to predict CA 30, 20, and 10 minutes in advance, using sensitivity, specificity, AUROC, AUPRC, and the false alarm rate as evaluation metrics, as shown in Table 2.

The proposed model could identify 90.6% (737/813) of the patients experiencing CA 30 minutes in advance, 92.6% (753/813) of the patients 20 minutes in advance, and 94.8% (771/813) of the patients 10 minutes in advance, with a false alarm rate of 8% or lower within 30 minutes. Compared to other studies, the model proposed in this study exhibited higher sensitivity and a lower false alarm rate.

We constructed 3 types of feature sets: raw features based on vital signs, statistical features, and a combination of raw features and statistical features. The 3 types of feature sets were input into the model for training to determine the impact of different feature sets on the model prediction performance. The results are shown in Table 3.

Regarding accuracy, sensitivity, AUROC, AUPRC, and F₁-score, the model performed better when statistical features were used than when raw features were used. When a combination of statistical and raw features was fed into the model, its predictive performance was also superior to that obtained using only raw features. However, its performance showed almost no difference compared to using only statistical features. Consequently, it can be inferred that statistical features enhanced the prediction performance of the model.

Vital sign patterns over time for patients in the CA group and the non-CA group are shown in Figure 6. Compared to the non-CA group, the CA group exhibited lower HR, SpO₂, and blood pressure values before the event, while respiratory rate was slightly higher. Vital signs began to decline 25 minutes before CA. The decline became more noticeable within the 10 minutes preceding the event. Among all the vital signs, blood pressure and SpO₂ showed an earlier decline. Compared to the CA group, the vital signs of the non-CA group remained relatively stable.

A deep learning model with real-time capability, high sensitivity, and a low false alarm rate was developed using only 6 vital signs based on the MIMIC-III database. The model relied on clinically accessible indicators to determine whether a patient would experience CA within the next hour at 5-minute intervals. It enabled real-time and continuous monitoring by overcoming the limitations of traditional prediction models that rely on a large number of features or features not commonly used in hospitals. The outcomes demonstrated that the model achieved excellent prediction performance. The TrGRU model developed in this study outperformed logistic regression, LGBM, and random forest across all metrics. Notably, compared to other models, TrGRU achieved both high sensitivity and a low false alarm rate. Table 4 presents a comparison between this study and other studies. Layeghian et al [22] developed a prediction model using ensemble learning methods based on multivariate features, vital signs, and clinical latent features, achieving an AUROC of 0.82 on the MIMIC-III dataset. However, their model was not designed for real-time monitoring.

The experiments investigating the impact of different feature sets on model prediction performance demonstrated that statistical features outperformed raw features in accuracy, sensitivity, AUROC, and AUPRC. When a combination of raw and statistical features was used, performance was also superior to that obtained using raw features alone but comparable to that obtained using only statistical features. Therefore, statistical features played a crucial role in predicting CA events.

Class imbalance is a critical issue that affects model training performance. When the data are severely imbalanced, models often perform poorly in predicting minority classes (ie, they tend to favor predicting negative samples). This situation is particularly common in medical datasets, as in the context of this study, the occurrence of CA events was much rarer than in normal conditions. To address this problem, we used upsampling to replicate positive samples and downsampling to reduce negative samples, thereby balancing the ratio of positive and negative class samples. This solution effectively improved the sensitivity of the model (sensitivity increased from 80.6% before sampling to 85.9% after sampling).

The TrGRU can identify 85.9% (689/813) of the patients 1 hour before a CA event, 90.6% (737/813) of the patients 30 minutes before, and 92.6% (753/813) and 94.8% (771/813) of the patients 20 and 10 minutes before, respectively. These results imply that even when it is too late to prevent CA, medical staff still have a certain amount of time to intervene, thereby improving patient survival rates. This is because the faster cardiopulmonary resuscitation is performed after CA, the greater the patient’s chance of survival, with the survival rate decreasing by 10% per minute before cardiopulmonary resuscitation is initiated. The model proposed by Yijing et al [8] identified 80% of the patients 25 minutes before CA and 93% of the patients 10 minutes before the event. The model proposed by Kwon et al [9] could also only identify 78% of the patients 30 minutes before CA. Therefore, the model proposed in this study demonstrated excellent performance in predicting CA events in advance, outperforming some existing studies.

Due to the “black box” nature of deep learning, it is challenging to establish relationships between real-time prediction results and features. In contrast, TrGRU provided a certain level of interpretability. We analyzed patterns of feature values over time by comparing the changes in vital signs between the CA group and the non-CA group before the occurrence of CA. This provided health care professionals with an objective basis for assessing physiological conditions. Clinical studies have shown that synergistic abnormal changes in vital signs (such as a sharp increase in respiratory rate combined with abnormal fluctuations in body temperature) often exhibit significant characteristic patterns before the occurrence of CA [23-25]. These patterns provide crucial evidence for clinical prediction and the determination of intervention windows.

To improve model generalizability, we used a meta-learning approach to address the issue of the model’s inability to adapt to different datasets. Using the MIMIC-III dataset as the pretraining dataset and the eICU-CRD dataset as the target for rapid adaptation, the pretrained model requires only a small number of samples to quickly adjust and adapt to the new dataset. The results demonstrated that the method proposed in this study significantly improved the generalization of the model’s capability and achieved excellent performance. Kim et al [6] trained their model on the MIMIC-IV dataset and subsequently conducted cross-dataset testing on the eICU-CRD dataset, achieving an AUROC of 0.80. Kwon et al [9] also performed cross-dataset validation on their proposed model, ultimately achieving an AUROC of 0.837 and an AUPRC of 0.239.

This study primarily made 4 contributions. First, compared with existing studies, we developed a highly sensitive CA prediction model. Sensitivity exceeded 90% in predicting CA events within 30 minutes, which could allow health care professionals to intervene earlier and improve patient survival rates. Second, we used a meta-learning approach to enhance the generalizability of the model, enabling it to adapt to different datasets rapidly. To the best of our knowledge, this was the first study to use meta-learning to address the adaptability of medical prediction models across different datasets. Third, the model in this study significantly reduced the false alarm rate, which could effectively help avoid unnecessary interventions or treatments, thereby improving the efficient use of medical resources and reducing health care costs. Additionally, reducing false alarms helps prevent “alarm fatigue” [26], ensuring that health care professionals remain highly alert and responsive to each alarm. Fourth, we developed a prediction model using fewer variables, overcoming limitations of existing models that rely on variables not commonly used in hospitals. This approach offers the advantage of being applicable to various hospital settings.

Our work mainly involves 3 limitations. First, due to the “black box” nature of deep learning, it is unable to establish relationships between prediction results and input data, and the interpretability of results is a critical factor for clinicians in making medical decisions; therefore, deep learning–based medical decision support still faces practical challenges. In recent years, the interpretability of deep learning has been explored in research [27-31], which is a key focus for our next steps. Second, CA may be caused by a variety of diseases, and in this study, we did not consider the heterogeneity between patients with different diseases [32]. Therefore, the generalizability of TrGRU cannot be guaranteed. Accordingly, it is necessary to expand the dataset to cover various types of diseases to further enhance the generalizability and robustness of the model. Third, we did not conduct feature selection. However, based on the research results, it appears that this omission did not significantly affect the performance of the model. In the future, feature selection could be used to further optimize the model presented in this study.

CA is a sudden and critical event with extremely low survival rates and a poor prognosis, posing a severe threat to patients. Existing CA prediction systems often suffer from low sensitivity and high false alarm rates. Consequently, there is an urgent clinical need for a reliable prediction system to assist health care professionals in real-time monitoring of CA events. The model proposed in this study can accurately predict CA events in real time, with high sensitivity and a low false alarm rate. Additionally, the use of meta-learning significantly enhances the adaptability of the model across different datasets. The TrGRU model shifts the golden resuscitation time window forward for patients who experience CA, which will play a significant role in improving patient survival rates. Moreover, its applicability to different clinical settings facilitates broader adoption.