Background: Life-threatening ventricular arrhythmias (LTVAs) are main causes of sudden cardiac arrest and are highly associated with an increased risk of mortality. A prediction model that enables early identification of the high-risk individuals is still lacking.
Objective: We aimed to build machine learning (ML)–based models to predict in-hospital mortality in patients with LTVA.
Methods: A total of 3140 patients with LTVA were randomly divided into training (n=2512, 80%) and internal validation (n=628, 20%) sets. Moreover, data of 2851 patients from another database were collected as the external validation set. The primary output was the probability of in-hospital mortality. The discriminatory ability was evaluated by the area under the receiver operating characteristic curve (AUC). The prediction performances of 5 ML algorithms were compared with 2 conventional scoring systems, namely, the simplified acute physiology score (SAPS-II) and the logistic organ dysfunction system (LODS).
Results: The prediction performance of the 5 ML algorithms significantly outperformed the traditional models in predicting in-hospital mortality. CatBoost showed the highest AUC of 90.5% (95% CI 87.5%-93.5%), followed by LightGBM with an AUC of 90.1% (95% CI 86.8%-93.4%). Conversely, the predictive values of SAPS-II and LODS were unsatisfactory, with AUCs of 78.0% (95% CI 71.7%-84.3%) and 74.9% (95% CI 67.2%-82.6%), respectively. The superiority of ML-based models was also shown in the external validation set.
Conclusions: ML-based models could improve the predictive values of in-hospital mortality prediction for patients with LTVA compared with traditional scoring systems.
Sudden cardiac arrest (SCA) is associated with unacceptable high mortality rates worldwide [- ]. It is estimated that 275,000 people in Europe present with SCA treated by emergency medical services (EMSs) each year, only 29,000 of whom survive to hospital discharge [ ]. Moreover, the EMS-treated SCA incidences in the United States, Canada, and China were about 64.4, 54.7, and 71.2 per 100,000 person years, respectively [ , ]. Life-threatening ventricular arrhythmias (LTVAs), which occur with severely depressed ventricular function and an unstable hemodynamics state, are the main causes of SCA [ ]. The high morbidity and mortality of LTVA cause a substantial economic burden and a serious health burden for EMSs [ , ]. Risk stratification and mortality assessment for patients with LTVA provide not only essential strategies for clinical decision-making but also practical information for health policy and insurance services [ , ].
Because patients with LTVA have a poor prognosis, early recognition of high-risk individuals is critical for timely interventions as well as intensive care and monitoring. The simplified acute physiology score (SAPS-II)  and the logistic organ dysfunction system (LODS) [ ] are traditional severity assessment systems that can predict the risk of death in general patients who received EMS. The LODS and SAPS-II models are comprehensive scoring systems used to evaluate the general conditions of patients and are limited by weak specificity. Moreover, several specific risk scores have been developed to perform SCA prediction in patients with hypertrophic cardiomyopathy [ ] and arrhythmogenic cardiomyopathy [ ]. However, a prediction model assessing the prognosis in patients with LTVA is still lacking.
Recently, machine learning (ML) algorithms have been used to build prediction models in clinical medicine and have been demonstrated to have more favorable predictive performance than traditional models [- ]. In this study, we aimed to develop and validate ML-based models to predict in-hospital mortality in critically ill patients with LTVA. Additionally, the comparisons of the ML models and the traditional severity assessment systems SAPS-II and LODS were performed to further demonstrate the predictive value of ML algorithms.
Source of Data
Patients’ data collected from the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 2.0) database were used to perform model training. The MIMIC-IV is a large critical care database that contains data from more than 200,000 individuals who were admitted to various intensive care units (ICUs) at the Beth Israel Deaconess Medical Center between 2008 and 2019 . In addition, data from the eICU Collaborative Research Database (eICU, version 2.0), which is a multicenter critical care database with high granularity data for over 200,000 admissions to ICUs [ ], were used to conduct the external validation.
Patient and Data Collection
Patients with LTVA in the hospital were included in this study. LTVAs were defined as documented sustained ventricular tachycardia (duration >30 seconds) or ventricular fibrillation. Patients aged <18 years or without adequate data, including demographic variables (eg, age, sex) and hospitalization details (eg, time of stay in the ICU, in-hospital mortality), were excluded. The data were collected from the database using Structure Query Language (PostgreSQL version 13.0). The variables included in this study could be divided into several categories: (1) demographics, including age, sex, race, height, and weight; (2) vital signs, including temperature, heart rate, blood pressure, respiratory rate, Glasgow coma scale (GCS), and urine output; (3) blood gas test, including pH, partial pressure of oxygen, partial pressure of carbon dioxide, and other parameters; (4) laboratory results, including white blood cell count, red blood cell count (RBC), creatinine, blood urea nitrogen, and other parameters; (5) comorbidities, including hypertension, diabetes, acute heart failure, myocardial infarction, and others; and (6) important interventions, including mechanical ventilation, renal replacement therapy, inotropic agents, and antibiotics. For variables with several records (eg, vital signs and laboratory tests), the minimum and maximum values within the first 24 hours after ICU admission were calculated, which may reflect the truly general conditions of the patients admitted to the ICU.
Several methods were introduced to handle the issues of extreme and missing data. Variables with >40% missing values were excluded from this study. Mean imputation was performed to fill in missing information for variables with <5% missing data. The multiple imputation was performed using the R multivariate imputation by chained equation (MICE) package for the variables with 5%-40% missing data .
Outcome Variables and Predictors
The primary outcome of interest was all-cause in-hospital mortality. The predictor variables considered in this study were selected based on a feature selection process ().
Model Development and Validation
In this study, the ML-based prediction models were established using the following steps:
- Feature selection: we applied CatBoost and LightGBM to initially evaluate the importance of individual features. Moreover, least absolute shrinkage and selection operator (LASSO) analysis was used to select the key predictors. LASSO is a form of L1 regularization that adds the absolute magnitude of feature coefficients as the penalty term rather than the squared magnitude that is typically used [ ]. After LASSO analysis, features with a zero coefficient were considered redundant and eliminated from the model fitting. As a result, automatic variable selection while simultaneously fitting the model could be realized.
- Algorithm selection and optimization: a total of 5 commonly used ML algorithms, including CatBoost, LightGBM, back propagation–neural network, random forest, and logistic regression, were applied to build the prediction models. We used the grid search strategy to identify the optimal combination of hyperparameters of these models to improve the prediction performance. Ten-fold stratified cross-validation was performed inside each grid option, and the optimal hyperparameter set was chosen based on the model in the grid search with the highest F1 score.
- Model fitting: before model fitting, a combination of the synthetic minority oversampling technique and undersampling was used to overcome the class-imbalance issue in this study. Qualified data from the MIMIC-IV database were randomly split into a training set (80% of the sample) and an internal validation set (20% of the sample) using the “train_test_split” function provided by Scikit-learn. As a result, the ML-based prediction models were developed.
In this study, the prediction models were validated both internally and externally. The discriminatory ability of the prediction model was evaluated by using the area under the receiver operating characteristic curve (AUC). The calibrated ability was qualitatively and quantitively assessed by the calibration curve and the Brier score, respectively. Calibration reflects the extent to which the predicted probabilities and actual probabilities agree and is qualitatively and quantitatively evaluated through the calibration curve and Brier score, respectively. The Brier score is calculated based on the Euclidean distance between the actual outcome and the predicted probability assigned to the outcome for each observation, with low values being desirable. In addition, decision curve analysis was used to demonstrate the decision benefit based on the models.
Several commonly used indices, including the AUC, sensitivity, specificity, and F1 score, were introduced to quantitively evaluate the predictive values of the models. To further demonstrate the favorable prediction performances of the ML-based models, the ML models and traditional severity assessment systems SAPS-II and LODS were compared in this study.
The black box characteristic of ML models makes it difficult to interpret how an ML algorithm could perform accurate prediction in clinical medicine settings. Accordingly, the Shapley additive explanations (SHAP) value was introduced in this study. SHAP was used to interpret the results from a predictive model. The interpretation was based on the SHAP value for each feature, representing the contribution of a feature to the predicted risk of the event. A positive SHAP value indicated that the corresponding feature contributed to a higher risk of the event, whereas a negative SHAP value indicated that the corresponding feature led to a lower risk of the event. The magnitude of SHAP values represented the contribution of that feature toward prediction performance . The SHAP summary plot was constructed to interpret and rank the significance of input features based on the mean absolute SHAP values of each feature. The SHAP dependency plot was used to understand how a single feature could affect the output of the prediction models.
The Kolmogorov-Smirnov test was used to evaluate the normal distribution of the data. Continuous variables were expressed as the mean (SD) or median (IQR) depending on the distribution of the data. Categorical data were summarized as frequencies and percentages. Baseline clinical characteristics were compared between the survival and nonsurvival groups using a t test or Welch t test for continuous variables and the chi-square test for dichotomous variables. The prediction performances of the different models were compared using a Delong test. All statistical tests were 2-sided, and P<.05 was considered statistically significant. Statistical analyses were performed in R (version 4.0.4, R Foundation for Statistical Computing). Python (version 3.9.0, Python Software Foundation) was used to conduct the ML-relevant processes.
This study was an analysis of a third-party, anonymized, publicly available database with pre-existing institutional review board approval, and informed consent from our institution was exempted. Data usage was approved by the review board of PhysioNet (authorization code: 35965741). The study was reported according to the recommendations of the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) statement .
We included 3140 critically ill patients with LTVA from the MIMIC-IV database, of whom 632 (20.1%) individuals died during hospitalization (). Compared with the survival group, nonsurvival patients were older (nonsurvival: mean 71.0, SD 14.2 years; survival: mean 67.9, SD 14.9 years; P<.001) and had lower GCS scores (nonsurvival: mean 10.4, SD 4.8; survival: mean 13.0, SD 3.2: P<.001). Moreover, nonsurvival patients were prone to having unstable vital signs. Furthermore, the data of 2851 patients from another database were used to perform the external validation. The baseline characteristics of the 2 cohorts are shown in .
Model Development and Validation
A total of 94 features were initially selected as the potential predictors (). Feature importance, evaluated using the CatBoost and LightGBM models, is shown in . We found that the GCS score, maximum RBC (RBC_max), and length of stay (LOS) in hospital prior to LTVA were the most predictive features in both the CatBoost and LightGBM models, indicating the certain stability. In the LASSO analysis, the coefficients of 54 variables were shrunk to zero ( ). Consequently, we retained 40 key features for the remaining analyses ( ).
As we mentioned above, 5 widely used ML algorithms were used to build the prediction models with the grid search strategy to optimize the ML-based models.summarizes the characteristics of the 5 ML algorithms. The ML-based models showed favorable discriminatory ability, with AUCs of 90.5% (95% CI 87.5%-93.5%), 89.7% (95% CI 86.0%-93.4%), 90.1 (95% CI 86.8%-93.4%), 89.1% (95% CI 85.7%-92.5%), and 89.5% (95% CI 86.2%-92.8%) for the CatBoost, back propagation–neural network, LightGBM, random forest, and logistic regression models, respectively. Conversely, the traditional models presented with ordinary prediction performance, with AUCs of 74.9% (95% CI 67.2%-82.6%) and 78.0% (95% CI 71.7%-84.3%) for LODS and SAPS-II, respectively ( A). The detailed prediction results for each model are shown in . We found that the AUCs of the 5 ML algorithms were all significantly higher than those of the SAPS-II and LODS models (all P<.001; ).
|Model||AUCa, % (95% CI)||SENSb, % (95% CI)||SPECc, % (95% CI)||F1 score, % (95% CI)|
|CatBoost||90.5 (87.5-93.5)||73.3 (65.4-81.3)||93.8 (92.2-95.4)||77.5 (70.5-84.5)|
|LightGBM||90.1 (86.8-93.4)||72.4 (68.1-76.7)||94.3 (92.4-96.2)||76.4 (69.2-83.6)|
|RFd||89.1 (85.7-92.5)||64.8 (56.7-73.0)||97.6 (96.0-99.2)||75.7 (68.5-82.9)|
|LRe||89.5 (86.2-92.8)||67.6 (60.0-75.3)||95.1 (93.5-96.8)||75.1 (66.5-82.0)|
|BP-NNf||89.7 (86.0-93.4)||69.3 (61.9-76.7)||96.5 (94.5-98.5)||77.7 (71.2-84.1)|
|LODSg||74.9 (72.3-80.5)||55.5 (45.5-65.5)||96.3 (93.6-99.0)||68.3 (58.9-77.7)|
|SAPS-IIh||78.0 (71.4-79.0)||58.8 (50.3-67.1)||93.9 (91.7-96.2)||69.0 (60.0-78.1)|
|CatBoost||87.2 (83.6-90.8)||60.7 (54.9-66.5)||96.6 (95.0-98.2)||71.8 (67.7-75.9)|
|LightGBM||84.9 (81.9-87.9)||67.1 (60.6-73.6)||94.2 (92.4-96.0)||73.9 (70.1-77.7)|
|RF||86.0 (83.4-88.6)||58.2 (52.5-63.9)||96.8 (95.6-98.0)||70.1 (65.8-74.4)|
|LR||81.1 (77.5-84.7)||57.2 (51.5-62.9)||94.3 (93.1-95.5)||76.4 (69.2-83.6)|
|BP-NN||85.6 (82.8-88.4)||64.3 (57.8-70.8)||94.8 (93.0-96.6)||65.8 (61.4-70.1)|
|LODS||73.7 (68.8-77.6)||51.0 (36.1-65.9)||96.6 (94.4-98.8)||64.8 (57.4-72.2)|
|SAPS-II||71.9 (63.8-80.0)||46.3 (37.2-55.4)||92.3 (88.9-95.7)||55.8 (46.2-65.4)|
aAUC: area under the receiver operating characteristic curve.
dRF: random forest.
eLR: logistic regression.
fBP-NN: back propagated–neural network.
gLODS: logistic organ dysfunction system.
hSAPS-II: simplified acute physiology score.
The 5 ML-based models showed excellent fit, with all Brier scores <0.1. However, the SAPS-II and LODS models largely deviated from the true events (B). In the decision curve analysis, the net benefits of all ML-based models outperformed the traditional models when the threshold probability was between 0.2 and 0.7, and the CatBoost model showed the best decision benefit compared with the others ( C).
In the external validation set, prediction performances of the 5 ML-based models were also significantly superior to the SAPS-II and LODS models. The AUCs of the SAPS-II and LODS models were only 73.7% (95% CI 68.8%-77.6%) and 71.9% (95% CI 63.8%-80.0%), respectively (and ). Accordingly, we suggest that the performance of the ML models is significantly superior to that of the traditional prognosis assessment systems for critically ill patients with LTVA.
To interpret the ML-based models, the SHAP method was used in this study. Specifically, we calculated the absolute mean SHAP values for each variable based on the CatBoost model applied to the internal validation set. For instance, with decreasing GCS scores and RBC_max values, the risk of in-hospital mortality of patients with LTVA increased. Moreover, the probability of in-hospital mortality increased with age and maximum central venous pressure values (A). Additionally, the SHAP dependency plot showed how different values of each feature affect the SHAP value and influence the output of the prediction models ( B and C). SHAP values for specific features exceeding zero represent an increased risk of in-hospital mortality.
The primary finding of this study demonstrates that ML algorithms significantly enhance the prognosis prediction performance for patients with LTVA. Compared with traditional scoring systems, including SAPS-II and LODS, the ML-based models enabled better prediction of in-hospital mortality in critically ill patients with LTVA.
LTVAs are the main cause of SCA and are highly related to an increased risk of in-hospital mortality. Several prediction models have been developed to predict SCA. O’Mahony et al  established a model that provided accurate individualized estimates for the probability of SCA in patients with hypertrophic myocardiopathy using 8 commonly used clinical parameters, with a C-statistic of 0.70. Adabag and Langsetmo [ ] built a SCA risk prediction model that could accurately predict SCA events in heart failure with preserved ejection fraction, with a C-statistic of 0.74. However, a model to predict outcomes in patients with LTVA is still lacking. The application value of ML algorithms in improving the mortality prediction performance for patients with LTVA remains unclear. Thus, we developed and validated ML-based models to precisely predict in-hospital mortality in patients with LTVA.
There are several severity assessment systems that could be used to predict outcomes of patients who are critically ill. The application values of SAPS-II and LODS in predicting the prognosis in patients who received EMS have been demonstrated in previous studies [- ]. ML-based models could further improve the prediction performance to predict the mortality in certain subsets [ , ]. The ML algorithms also exhibited superiority over the SAPS-II and LODS models in this study. SAPS-II incorporates 12 physiological variables, age, type of admission, and 3 underlying disease variables, while LODS identifies levels of organ dysfunction for 6 organ systems (ie, neurologic, cardiovascular, renal, pulmonary, hematologic, and hepatic). In contrast, our developed models included additional important features, such as GCS score, RBC, and central venous pressure, which may have contributed to our improved prediction performance. Therefore, the enhanced performance of the ML models in our study may be attributed to their ability to effectively capture intricate relationships and patterns within a more extensive set of variables.
Of the 40 selected variables, we found that GCS score, RBC_max, and LOS in hospital were the most important predictors. The GCS score includes assessments of motor, verbal, and eye responses, which could reflect the level of consciousness. A previous study reported that the GCS score could be used to predict mortality in patients who are critically ill , which is consistent with our findings. Anemia on admission is also associated with an increased mortality in critical care patients [ ]. We also found there was an inverse relationship between RBC and in-hospital mortality. Moreover, LOS was a key variable for predicting mortality, as expected. Other predictors included vital signs, hepatic and renal function tests, and relevant comorbidities. Although abnormal potassium levels and the use of macrolide antibiotics are known risk factors for LTVA and may increase the risk of mortality [ , ], we observed weak contributions of these features to our predictive models. This could be attributed to the complex interplay between multiple factors in the development of arrhythmias and the limitations imposed by the sample size of our study.
In addition, the ML-based models in this study had certain interpretability. Many of the features used to perform mortality risk prediction in this study were tangible, and some of them had been proven intimately correlated with mortality in patients who are critically ill. The SHAP method, which could provide a visual interpretation of the ML-based models at the global and local levels, was introduced in this study. Specifically, the SHAP summary plot, which interpreted and ranked the significance of input features, and the SHAP dependency plot, which explained how a single feature affected the output of the ML-based models, were presented in this study.
The prediction models in this study enable early and accurate selection of patients at high risk of in-hospital death, which may be conducive to risk stratification, clinical decision-making, and the improvement of prognosis in patients with LTVA. On the other hand, our prediction models were developed based on critical care cohorts. Essentially, sicker patients with more multisystem derangements are anticipated to have poorer prognosis. Considering their satisfactory specificities, our prediction models may help to identify subgroups of patients with LTVA treated with EMSs who are not anticipated to have an adverse prognosis. It is useful to avoid wasting medical services, such as through unnecessary intensive monitoring and aggressive therapies in the ICU.
Although we demonstrated the predictive value of ML algorithms in predicting mortality in critical patients with LTVA, some limitations should be acknowledged. First, all data involved in this study were retrospectively collected from public databases, and further prospective studies are needed to confirm the findings. Second, some important variables, such as cardiac troponin and ejection fraction, were excluded from this study due to unacceptably high rates of missing values, which may be a potential source of bias. Third, the prediction models were developed based on critical care cohorts, which may affect application to the general population.
In summary, the presented ML-based models exhibited better predictive values than did traditional severity assessment systems, such as SAPS-II and LODS, in predicting the mortality of critical patients with LTVA. Our findings indicate that ML algorithms could be used to improve model performance for predicting outcomes of patients with LTVA. Future prospective studies are needed to confirm the findings.
This study was supported by the High-level Hospital Project of Fuwai Hospital, Chinese Academy of Medical Sciences (grant 2022-GSP-GG-25).
The data sets used and analyzed during this study are available from the corresponding author upon reasonable request.
This study was designed by LL and LD. Zhuxin Zhang, LZ, YX, Zhenhao Zhang, and ZH were responsible for data collation and statistical analysis. LL wrote the first draft. YY reviewed and checked the manuscript. All authors read and approved the final manuscript.
Conflicts of Interest
Baseline characteristics.DOCX File , 24 KB
All variables used in the machine learning models.DOCX File , 22 KB
Variables excluded after feature selection by least absolute shrinkage and selection operator (LASSO).DOCX File , 15 KB
Feature selection by least absolute shrinkage and selection operator (LASSO) analysis. (A) Plots for LASSO analysis coefficients. (B) Cross-validation plot for the penalty term.DOCX File , 426 KB
Characteristics of the included machine learning algorithms.DOCX File , 15 KB
Comparisons of the area under the receiver operating characteristic curve of different models using the Delong test.DOCX File , 14 KB
External validation of different models.DOCX File , 415 KB
- Berdowski J, Berg RA, Tijssen JG, Koster RW. Global incidences of out-of-hospital cardiac arrest and survival rates: systematic review of 67 prospective studies. Resuscitation 2010 Nov;81(11):1479-1487 [CrossRef] [Medline]
- Sasson C, Rogers MA, Dahl J, Kellermann AL. Predictors of survival from out-of-hospital cardiac arrest. Circ Cardiovasc Qual Outcomes 2010 Jan;3(1):63-81 [CrossRef]
- Becker LB, Aufderheide TP, Graham R. Strategies to improve survival from cardiac arrest: a report from the Institute of Medicine. JAMA 2015 Jul 21;314(3):223-224 [CrossRef] [Medline]
- Atwood C, Eisenberg MS, Herlitz J, Rea TD. Incidence of EMS-treated out-of-hospital cardiac arrest in Europe. Resuscitation 2005 Oct;67(1):75-80 [CrossRef] [Medline]
- Kiguchi T, Okubo M, Nishiyama C, Maconochie I, Ong MEH, Kern KB, et al. Out-of-hospital cardiac arrest across the world: first report from the International Liaison Committee on Resuscitation (ILCOR). Resuscitation 2020 Jul;152:39-49 [CrossRef] [Medline]
- Shao F, Li CS, Liang LR, Li D, Ma SK. Outcome of out-of-hospital cardiac arrests in Beijing, China. Resuscitation 2014 Nov;85(11):1411-1417 [CrossRef] [Medline]
- Link MS, Berkow LC, Kudenchuk PJ, Halperin HR, Hess EP, Moitra VK, et al. Part 7: adult advanced cardiovascular life support: 2015 American Heart Association guidelines update for cardiopulmonary resuscitation and emergency cardiovascular care. Circulation 2015 Nov 03;132(18_suppl_2):S444-S464 [CrossRef]
- Rahimi K, Watzlawek S, Thiele H, Secknus MA, Hayerizadeh BF, Niebauer J, et al. Incidence, time course, and predictors of early malignant ventricular arrhythmias after non-ST-segment elevation myocardial infarction in patients with early invasive treatment. Eur Heart J 2006 Jul;27(14):1706-1711 [CrossRef] [Medline]
- Yannopoulos D, Bartos J, Raveendran G, Walser E, Connett J, Murray TA, et al. Advanced reperfusion strategies for patients with out-of-hospital cardiac arrest and refractory ventricular fibrillation (ARREST): a phase 2, single centre, open-label, randomised controlled trial. The Lancet 2020 Dec;396(10265):1807-1816 [CrossRef]
- Wahbi K, Ben Yaou R, Gandjbakhch E, Anselme F, Gossios T, Lakdawala NK, et al. Development and validation of a new risk prediction score for life-threatening ventricular tachyarrhythmias in laminopathies. Circulation 2019 Jul 23;140(4):293-302 [CrossRef] [Medline]
- Cadrin-Tourigny J, Bosman L, Nozza A, Wang W, Tadros R, Bhonsale A, et al. A new prediction model for ventricular arrhythmias in arrhythmogenic right ventricular cardiomyopathy. Eur Heart J 2022 Aug 21;43(32):e1-e9 [https://europepmc.org/abstract/MED/35441664] [CrossRef] [Medline]
- Le Gall J, Lemeshow S, Saulnier F. A New Simplified Acute Physiology Score (SAPS II) Based on a European/North American Multicenter Study. JAMA 1993 Dec 22;270(24):2957 [CrossRef]
- Le Gall J, Klar J, Lemeshow S, Saulnier F, Alberti C, Artigas A, et al. The logistic organ dysfunction system. A new way to assess organ dysfunction in the intensive care unit. ICU Scoring Group. JAMA 1996 Sep 11;276(10):802-810 [CrossRef] [Medline]
- Miron A, Lafreniere-Roula M, Steve Fan CP, Armstrong KR, Dragulescu A, Papaz T, et al. A validated model for sudden cardiac death risk prediction in pediatric hypertrophic cardiomyopathy. Circulation 2020 Jul 21;142(3):217-229 [https://www.ahajournals.org/doi/abs/10.1161/CIRCULATIONAHA.120.047235?url_ver=Z39.88-2003&rfr_id=ori:rid:crossref.org&rfr_dat=cr_pub 0pubmed] [CrossRef] [Medline]
- Cadrin-Tourigny J, Bosman L, Wang W, Tadros R, Bhonsale A, Bourfiss M, et al. Sudden cardiac death prediction in arrhythmogenic right ventricular cardiomyopathy: a multinational collaboration. Circ Arrhythm Electrophysiol 2021 Jan;14(1):e008509 [https://www.ahajournals.org/doi/abs/10.1161/CIRCEP.120.008509?url_ver=Z39.88-2003&rfr_id=ori:rid:crossref.org&rfr_dat=cr_pub 0pubmed] [CrossRef] [Medline]
- Li L, Zhang Z, Xiong Y, Hu Z, Liu S, Tu B, et al. Prediction of hospital mortality in mechanically ventilated patients with congestive heart failure using machine learning approaches. Int J Cardiol 2022 Jul 01;358:59-64 [CrossRef] [Medline]
- Layeghian Javan S, Sepehri MM, Layeghian Javan M, Khatibi T. An intelligent warning model for early prediction of cardiac arrest in sepsis patients. Comput Methods Programs Biomed 2019 Sep;178:47-58 [CrossRef] [Medline]
- Li J, Liu S, Hu Y, Zhu L, Mao Y, Liu J. Predicting mortality in intensive care unit patients with heart failure using an interpretable machine learning model: retrospective cohort study. J Med Internet Res 2022 Aug 09;24(8):e38082 [https://www.jmir.org/2022/8/e38082/] [CrossRef] [Medline]
- Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data 2023 Jan 03;10(1):1 [https://doi.org/10.1038/s41597-022-01899-x] [CrossRef] [Medline]
- Pollard TJ, Johnson AEW, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Sci Data 2018 Sep 11;5:180178 [https://doi.org/10.1038/sdata.2018.178] [CrossRef] [Medline]
- Zhang P. Multiple imputation with multivariate imputation by chained equation (MICE) package. Ann Transl Med 2016 Jan;4(2):30 [CrossRef] [Medline]
- Kok BC, Choi JS, Oh H, Choi JY. Sparse extended redundancy analysis: variable selection via the exclusive LASSO. Multivariate Behav Res 2021;56(3):426-446 [CrossRef] [Medline]
- Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, et al. From local explanations to global understanding with explainable AI for trees. Nat Mach Intell 2020 Jan;2(1):56-67 [https://europepmc.org/abstract/MED/32607472] [CrossRef] [Medline]
- Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ 2015 Jan 07;350:g7594 [https://core.ac.uk/reader/81583807?utm_source=linkout] [CrossRef] [Medline]
- O'Mahony C, Jichi F, Pavlou M, Monserrat L, Anastasakis A, Rapezzi C, et al. Hypertrophic Cardiomyopathy Outcomes Investigators. A novel clinical risk prediction model for sudden cardiac death in hypertrophic cardiomyopathy (HCM risk-SCD). Eur Heart J 2014 Aug 07;35(30):2010-2020 [CrossRef] [Medline]
- Adabag S, Langsetmo L. Sudden cardiac death risk prediction in heart failure with preserved ejection fraction. Heart Rhythm 2020 Mar;17(3):358-364 [CrossRef] [Medline]
- Le Gall J, Neumann A, Hemery F, Bleriot J, Fulgencio J, Garrigues B, et al. Mortality prediction using SAPS II: an update for French intensive care units. Crit Care 2005;9(6):R645-R652 [https://ccforum.biomedcentral.com/articles/10.1186/cc3821] [CrossRef] [Medline]
- Poncet A, Perneger TV, Merlani P, Capuzzo M, Combescure C. Determinants of the calibration of SAPS II and SAPS 3 mortality scores in intensive care: a European multicenter study. Crit Care 2017 Apr 04;21(1):85 [https://ccforum.biomedcentral.com/articles/10.1186/s13054-017-1673-6] [CrossRef] [Medline]
- Jones AE, Fitch MT, Kline JA. Operational performance of validated physiologic scoring systems for predicting in-hospital mortality among critically ill emergency department patients. Crit Care Med 2005 May;33(5):974-978 [CrossRef] [Medline]
- Motwani M, Dey D, Berman DS, Germano G, Achenbach S, Al-Mallah MH, et al. Machine learning for prediction of all-cause mortality in patients with suspected coronary artery disease: a 5-year multicentre prospective registry analysis. Eur Heart J 2017 Feb 14;38(7):500-507 [https://europepmc.org/abstract/MED/27252451] [CrossRef] [Medline]
- Hou N, Li M, He L, Xie B, Wang L, Zhang R, et al. Predicting 30-days mortality for MIMIC-III patients with sepsis-3: a machine learning approach using XGboost. J Transl Med 2020 Dec 07;18(1):462 [https://translational-medicine.biomedcentral.com/articles/10.1186/s12967-020-02620-5] [CrossRef] [Medline]
- Wijdicks EF, Kramer AA, Rohs T, Hanna S, Sadaka F, O'Brien J, et al. Comparison of the full outline of unresponsiveness score and the Glasgow coma scale in predicting mortality in critically ill patients*. Crit Care Med 2015 Feb;43(2):439-444 [CrossRef] [Medline]
- Lasocki S, Pène F, Ait-Oufella H, Aubron C, Ausset S, Buffet P, et al. Management and prevention of anemia (acute bleeding excluded) in adult critical care patients. Ann Intensive Care 2020 Jul 22;10(1):97 [https://europepmc.org/abstract/MED/32700082] [CrossRef] [Medline]
- Patel RB, Tannenbaum S, Viana-Tejedor A, Guo J, Im K, Morrow DA, et al. Serum potassium levels, cardiac arrhythmias, and mortality following non-ST-elevation myocardial infarction or unstable angina: insights from MERLIN-TIMI 36. Eur Heart J Acute Cardiovasc Care 2017 Feb;6(1):18-25 [https://europepmc.org/abstract/MED/26714972] [CrossRef] [Medline]
- Kundu S, Williams SR, Nordt SP, Clark RF. Clarithromycin-induced ventricular tachycardia. Ann Emerg Med 1997 Oct;30(4):542-544 [CrossRef] [Medline]
|AUC: area under the receiver operating characteristic curve|
|EMS: emergency medical services|
|GCS: Glasgow coma scale|
|ICU: intensive care unit|
|LASSO: least absolute shrinkage and selection operator|
|LODS: logistic organ dysfunction system|
|LOS: length of stay|
|LTVA: life-threatening ventricular arrhythmia|
|MICE: multivariate imputation by chained equation|
|MIMIC-IV: Medical Information Mart for Intensive Care IV|
|ML: machine learning|
|RBC: red blood cell count|
|RBC_max: maximum red blood cell count|
|SAPS-II: simplified acute physiology score|
|SCA: sudden cardiac arrest|
|SHAP: Shapley additive explanations|
|TRIPOD: Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis|
Edited by T de Azevedo Cardoso; submitted 28.03.23; peer-reviewed by L He, E Kawamoto; comments to author 17.07.23; revised version received 20.07.23; accepted 18.09.23; published 15.11.23Copyright
©Le Li, Ligang Ding, Zhuxin Zhang, Likun Zhou, Zhenhao Zhang, Yulong Xiong, Zhao Hu, Yan Yao. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 15.11.2023.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.