Skip to main content
Browse Subject Areas

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Deep-learning-based prognostic modeling for incident heart failure in patients with diabetes using electronic health records: A retrospective cohort study

  • Ilaria Gandin ,

    Roles Conceptualization, Formal analysis, Funding acquisition, Methodology, Writing – original draft, Writing – review & editing

    Affiliation Department of Medical Sciences, Biostatistics Unit, University of Trieste, Trieste, Italy

  • Sebastiano Saccani,

    Roles Formal analysis, Methodology, Writing – review & editing

    Affiliation Aindo, Trieste, Italy

  • Andrea Coser,

    Roles Formal analysis

    Affiliation Aindo, Trieste, Italy

  • Arjuna Scagnetto,

    Roles Data curation

    Affiliation Cardiovascular Center, University Hospital and Health Services of Trieste, Trieste, Italy

  • Chiara Cappelletto,

    Roles Conceptualization, Writing – review & editing

    Affiliation Cardiovascular Center, University Hospital and Health Services of Trieste, Trieste, Italy

  • Riccardo Candido,

    Roles Conceptualization, Writing – review & editing

    Affiliation Diabetes Center, University Hospital and Health Services of Trieste, Trieste, Italy

  • Giulia Barbati,

    Roles Conceptualization, Methodology, Supervision, Writing – review & editing

    Affiliation Department of Medical Sciences, Biostatistics Unit, University of Trieste, Trieste, Italy

  • Andrea Di Lenarda

    Roles Conceptualization, Funding acquisition, Investigation, Resources, Supervision, Writing – review & editing

    Affiliation Cardiovascular Center, University Hospital and Health Services of Trieste, Trieste, Italy


Patients with type 2 diabetes mellitus (T2DM) have more than twice the risk of developing heart failure (HF) compared to patients without diabetes. The present study is aimed to build an artificial intelligence (AI) prognostic model that takes in account a large and heterogeneous set of clinical factors and investigates the risk of developing HF in diabetic patients. We carried out an electronic health records- (EHR-) based retrospective cohort study that included patients with cardiological clinical evaluation and no previous diagnosis of HF. Information consists of features extracted from clinical and administrative data obtained as part of routine medical care. The primary endpoint was diagnosis of HF (during out-of-hospital clinical examination or hospitalization). We developed two prognostic models using (1) elastic net regularization for Cox proportional hazard model (COX) and (2) a deep neural network survival method (PHNN), in which a neural network was used to represent a non-linear hazard function and explainability strategies are applied to estimate the influence of predictors on the risk function. Over a median follow-up of 65 months, 17.3% of the 10,614 patients developed HF. The PHNN model outperformed COX both in terms of discrimination (c-index 0.768 vs 0.734) and calibration (2-year integrated calibration index 0.008 vs 0.018). The AI approach led to the identification of 20 predictors of different domains (age, body mass index, echocardiographic and electrocardiographic features, laboratory measurements, comorbidities, therapies) whose relationship with the predicted risk correspond to known trends in the clinical practice. Our results suggest that prognostic models for HF in diabetic patients may improve using EHRs in combination with AI techniques for survival analysis, which provide high flexibility and better performance with respect to standard approaches.


In the last decades Type 2 diabetes mellitus (T2DM) has become a global epidemic which is expected to affect over 592 million people worldwide by 2035 [1, 2]. Diabetes is associated with a decrease in life expectancy, in large part attributable to cardiovascular diseases [3]. In particular, patients with T2DM have more than twice the risk of developing heart failure (HF) compared to patients without diabetes mellitus [4, 5]. While 10% to 15% of the general population have diabetes, 44% of patients hospitalized for HF have diabetes mellitus [6]. The increased incidence of HF in diabetic patients persists even after adjusting for well-known risk factors for HF in general populations such as age, hypertension, hypercholesterolemia, and coronary artery disease. Therefore, in order to target high-risk individuals and reduce the risk of HF development with pharmacological agents (like sodium-glucose contransporter 2 inhibitor [7, 8]), the identification of diabetes-specific characteristics involved in HF development remains an important clinical question.

Several studies have developed clinical prognostic models for HF in diabetics patients, however they were not able to provide a comprehensive risk stratification and thus no score has yet been included in guideline care. In a recent work [9], Razaghizad et al. carried out a systematic evaluation on 15 models developed for hospitalization for HF in type 2 diabetes and showed that RECODe risk equation (together with TRS-HF_DM, another well performing model, although with a higher potential risk for bias) can be considered the most promising score for clinical use. Developed using data from the Action to Control Cardiovascular Risk in Diabetes study (ACCORD), the RECODe risk equations include age, sex, ethnicity, smoke, systolic blood pressure, history of cardiovascular disease, blood pressure-lowering drugs, statins, anticoagulants, HbA1c, total cholesterol, HDL, serum creatinine and urine albumin:creatinine ratio as predictors [10]. The model showed moderate-to-good discrimination and calibration in internal (c-index = 0.75, calibration slope = 1.01, intercept = -0.0004) and external validation (c-index = 0.76, calibration slope = 1.13, intercept = -0.011).

We hypothesized that an innovative approach based on the use of features available in electronic health records (EHRs) and artificial intelligence methods may improve performance in prognostic models in clinical settings. EHRs are the whole set of digital data originated at single-patient level in health care institutions as part of the clinical routine. Even though EHRs represent a valuable source of information (massive volumes, longitudinal nature, up-to-date, multiple domains) [11, 12], such data have been rarely include in risk scores because of low-quality issues, such as high heterogeneity and noise. Advances in artificial intelligence, in particular in neural networks architectures for deep learning, are offering computational techniques able to leverage the richness of EHRs for personalized healthcare [1315]. Although deep learning approaches have been recently extended from prediction tasks to survival analysis, in which modeling right censored data is required [1618], their application on EHRs for prognostic models has been limited.

In this study we investigate two research questions: (1) whether EHRs can improve risk stratification for HF in patients with type 2 diabetes; (2) the advantages of deep learning survival models over more standard approaches to account for non-linear effects and interactions between variables.

Materials and methods


The present study is a cohort observational, retrospective study on patients enrolled in the Cardiovascular Observatory of Trieste (Italy) [19] affected by diabetes mellitus that had a cardiological evaluation from November 1, 2009 until December 31, 2018. All information on patients was extracted from clinical records and administrative data obtained as part of routine medical care. Data include medical information collected by cardiologists during routine clinical practice, diagnostic codes, laboratory tests, procedures, and cardiovascular drugs prescriptions sorted out using electronic indexes, comorbidities. Diagnosis of diabetes was based on multiple criteria: recorded diagnosis of diabetes, evidence of exemption for healthcare expenses, evidence of glycated hemoglobin levels>6.5%, purchase of at least two antidiabetic medications within one year. The first cardiological examination was considered the index visit, from which the absence of HF was ascertained (patients already diagnosed with HF when entering the Cardiovascular Observatory were excluded from the study).


The primary endpoint was the onset of HF identified as the first between the following events: diagnosis of HF during hospitalization (ICD-9 codes: 39891, 40201, 40211, 40291, 40401, 40403, 40411, 40413, 40491, 40493, 4280–4284, 4289) and diagnosis of HF based at out-of-hospital clinical examination. Diagnosis of HF was performed according to ESC criteria: typical symptoms (breathlessness, ankle swelling and fatigue) and/or signs (elevated jugular venous pressure, pulmonary crackles and peripheral oedema) in presence of a structural and/or functional cardiac abnormality. Follow-up period for HF onset started at the index visit and ended on the administrative censoring date December 31, 2019. Death as a competing risk was not taken into account after an analysis of the Kaplan-Meier curve that showed a negligible bias within the first 60 months (S1 Fig). Baseline characteristics were compared between HF and HF-free individuals using chi-square test for categorical variables and t-test for continuous variables (or Mann-Whitney test, when appropriate).

Derivation of the models

The cohort was randomly divided in training, validation and test set (70%, 15% and 15% respectively) maintaining approximately the same ratio of patients that experienced HF and censored patients. The test set was fixed beforehand and has been held out from the training of the different models tested and has been solely used to evaluate the models. We developed two models: first, a linear proportional hazards regression model (COX); second, a non-linear proportional hazards deep neural network model (PHNN).

Using elastic net regularization, a machine learning approach, we developed a Cox proportional hazard model that kept, among all possible covariates, only those identified as relevant predictors. According to the Cox hypothesis, the hazard function is assumed to be the product of two components h(t|X) = h0(t)r(X), where h0(t) is the baseline hazard function and r(X) is the risk function and approximated by the exponential of a linear function r(X) = exp(βX). Elastic net regularization algorithm implements a Cox model via penalized maximum likelihood and it is able to select among candidate predictors the best model in the context of collinearity [20]. In our study L2 regularization parameter was set to 1 and 10-fold cross validation was performed to select the L1 regularization parameter λ following the “one-standard-error rule” [21] (given the context of application of the model and the need for a parsimonious model, we selected the largest value λ for which the CV error is within 1 standard error of the minimizing rule). Missing values were imputed using MICE, a multiple imputation technique based on chained equations [22].

In line with the work of Katzman et al. [16], we developed a second model that lifts the linear hazard hypothesis and model the risk function as r(X) = exp(f(X)), with f(X) a very generic function of the covariates. In particular, f(X) = fϑ(X) has been approximated with a fully connected deep neural network (four layers with hidden size 128, 64, 32, 15 hidden units, rectifier non linearity, activation dropout 0.5 in all layers) parametrized by the weights ϑ. The standard package of auto-differentiation PyTorch was then used to minimize the partial likelihood over the train set to fit the network weights ϑ. Training was performed under early stopping using the validation set. To understand the dependence between variables and predicted risk, we reported the partial dependence plots (PDPs) [23] representing the marginal effect of covariates on predicted log-hazard. For a covariate of interest Xk, and for any value xk assumed by Xk, we estimate the partial dependence function as the average where n is the number of observations and xji is the value of covariate Xj for the i-th individual. Before training the model, a forward selection procedure was applied on the large number of variables included in the dataset. In the first step, a model was estimated using only one covariate at a time and the relative c-index was computed. The covariate with the highest c-index turned out to be “age”. In the second step, another round was performed using”age” and other variables one at a time. The feature corresponding to the model with highest c-index was then retained as the second feature. The process was repeated adding one feature at a time until adding new features did not bring a substantial improvement in the c-index. Notice that, to perform this feature selection, a validation set has been extracted from the training set, with the same size of the test set. This set was used to evaluate the many models considered, and not to train them. For the development of PHDL, missing values were imputed with the sample mean and a flag column was added to retain the information on values originally missing. Numerical columns were normalized to a normal distribution with a quantile transformer. In order to retain as much as possible information and to let the model learning from any hidden trend within the data, no additional feature engineering was applied. Validity of the proportional hazards assumption was checked using graphical diagnostics based on the scaled Schoenfeld residuals for all predictors selected by the models.

Validation of the models

The discrimination of the models was evaluated by c-index (Harrell’s estimator) which is a generalization of the area under the receiver operating characteristic curve for time-to-event data and can be interpreted as the ability of a model to rank patients from high to low risk. Mean and standard error for the index was obtained through 10-fold cross-validation. In addition, time-dependent ROC curve was generated at 2- and 5-year of follow-up and the relative area under the curves (AUCs) were compared between models using inferential techniques [24]. We carried out graphical assessment of calibration by dividing subjects into 10 groups using deciles of the predicted probabilities and comparing predicted/observed risk across strata. Moreover, for each model we reported the Integrated Calibration Index (ICI) [25], which is the weighted average of the absolute difference between observed and predicted risks, in which the absolute differences are weighted by the empirical density function of the predicted risks.

We also compared predictions from our model with those from the RECODe study. In the calculation of RECODe equation, information on urine albumin:creatinine ratio was absent (as stated by the authors, individuals without a known covariate can have the relative term omitted from the equations). Moreover, instead of using the RECODe baseline HF-free survival reported in the article (0.96) to recalibrate the model we calculated an updated baseline HF-free survival by calculating the 2- and 5-year HF-free survival in the test set (0.927 and 0.856 respectively).

Analyses were done using R (version 4.2.1; R Foundation for Statistical Computing, Vienna) and Python 3.8.10 and PyTorch 1.10.2. The study involved the use of clinical records and administrative data produced as part of routine medical care, in compliance with the local regulatory and privacy policies. Data were collected in an anonymous form. A written informed consent was obtained under the institutional review board policies of hospitals administration. The current study was approved by Comitato Etico Unico Regionale FVG (Protocol ID: 114_2020T). All information was linked and anonymized before the analysis.


The cohort included 10,614 patients: 4,447 (42%) were females and mean age was 72 (SD = 11). Baseline characteristics according to heart failure diagnosis are presented in Table 1. During a median of 65 months, 1840 patients (17.3%) developed HF. Probability of HF-free survival at 2- and 5-year was 92.7% (95% CI [94.8,95.6]) and 85.6% (95% CI [84.9,86.3]) respectively.

Table 1. Characteristics of the cohort separately for individuals free from HF and individuals that have developed HF.

GFR was estimated using the EPI-CKD formula. History of cardiovascular disease includes stroke and myocardial infarction. SD = standard deviation; IQR = interquantile range; RASi = Renin–angiotensin system inhibitors; MRA = Aldosterone receptor antagonists.

Based on elastic net regularization, most relevant variables for predicting HF were age, diuretics, Charlson score, left atrium area, atrial fibrillation, organ damage, hypertension, adolsterone antagonist, glomerular filtration rate (protective factor) (Fig 1).

Fig 1. Predictors included in the penalized Cox model ordered by magnitude of effect.

GFR is the only variable associated with hazard ratio<1, meaning that higher values of GFR are associated with lower risk of HF.

Using the penalized Cox model we obtained a c-index of 0.734 (Table 2). Considering time points 2 years and 5 years, the area under the time-dependent ROC was 0.716 (95% CI [0.664,0.767]) and 0.770 (95% CI [0.732,0.807]) respectively. As depicted in Fig 2, calibration of prediction was acceptable at 2 years (ICI = 0.018) but considerably decreased for predictions at 5 years (ICI = 0.034).

Fig 2. Calibration for COX model.

Deviations from the diagonal line denote lack of calibration.

Table 2. Performance of the three models.

SE = standard error; AUC = Area under the ROC curve; ICI = integrated calibration index; CI = confidence interval.

As for the DL approach, the feature selection process identified 20 relevant variables (S2 Fig, S1 Table): age, BMI, four echocardiographic parameters (left ventricular wall motion score index, continuous wave aortic velocity, tissue doppler E wave velocity, tricuspid regurgitation), three ECG parameters (P axis absent, P axis, T axis), five comorbidities (renal disease, hypertension, lung disease, pericardium disease, peripheral artery disease), three laboratory measurements (hemoglobin, glycemia, triglyceride levels) and three categories of medication (diuretics, anticoagulants, RASi). In Fig 3 and S3 Fig it is possible to observe for each predictor the partial dependence plot, representing the relationship between the variable and the log-hazard.

Fig 3. Partial dependence plots for PHNN model.

The blue line (or bar for categorical values) represents the value of the log hazard for various values of the covariate (x axis). Higher values correspond to higher hazard.

Using the DL model we obtained a better discrimination, with a c-index of 0.768. In the time-dependent ROC analysis we observed better accuracy as well: AUC was 0.771 (95% CI [0.723,0.818]) at 2-year of follow-up and 0.780 (95% CI [0.743,0.817]) at 5-year of follow up, although differences were not statistically significant (2-year p-value = 0.059, 5-year p-value = 0.772). Moreover, the ICI for 2- and 5-year risk (0.008 and 0.015, respectively) indicated good calibration (Fig 4). Additional information on models performance is reported in S2 Table.

Compared with our models, RECODe risk score had worse discrimination (c-index of 0.670, see Table 2). In particular, predicted survival was less accurate compared to the penalized Cox model both at 2-year follow-up (AUC = 0.651, 95% CI [0.601,0.701], p-value = 0.042) and 5-year follow-up (AUC = 0.668, 95% CI [0.621,0.709], p-value<0.001). Significant decrease in discrimination was observed also with respect to the DL model (2-year p-value<0.001, 5-year p-value<0.001).


We developed two prognostic models for HF in diabetic patients using EHRs (one assuming linear hazard and the other assuming a non-linear hazard) and showed the benefit of implementing deep learning algorithms in terms of performance.

Among previous studies that developed prognostic models for HF risk in patients with diabetes, RECODe risk equations are considered the most promising for clinical use. However, such model performed poorly in our cohort. A possible explanation can be found in the characteristics of our cohort: we applied the model in a clinical setting that included patients with a wide range of comorbidities, unlike individuals part of the research cohorts in which the model was derived.

Both our models were superior to RECODe risk equations. COX model obtained with elastic net regularization identified eight clinical predictors already associated with HF risk. As reported in [9], diabetes duration, diuretics, atrial fibrillation, arterial hypertension, are risk factors commonly included in prediction models for HF, along with GFR as protective factor. Charlson comorbity index [26] was found to be a risk factor for heart failure readmission in another large-scale study [27]. One of the most relevant predictors was loop diuretics (HR = 1.25), which are commonly used in the treatment of HF, thus possibly a sign of presence of masked HF cases not detectable from EHRs [28]. Model’s discrimination ability was acceptable but calibration showed poor results (in particular for 5-year risk). This could be due to the low flexibility in the model’s specification.

On the other side, the implementation of a deep neural network in the PHNN model made possible to reach moderate performance in terms of discrimination and well-calibrated predictions. As for variable selection, eight of the PHNN covariates were either predictors of RECODe equations (age, systolic blood pressure, blood pressure-lowering drugs, anticoagulants) or common risk predictors reported in [9] (BMI, hemoglobin, chronic kidney disease, peripheral artery disease). It is interesting to notice that one of the COX predictors was diagnosis of atrial fibrillation: such variable was not included in the PHNN model, however in the list of predictors we observe “absence of P axis” and “P axis” that is an ECG feature closely related with atrial fibrillation. Moreover, one of the predictors was continuous wave aortic velocity which high values could indicate aortic stenosis, a condition that often coexists with atrial fibrillation and predispose to HF [29]. Others relevant echocardiographic parameters were: 1) tissue doppler E wave velocity, for which we estimated an inverse proportional relationship with HF risk; 2) tricuspid valve regurgitation, showing an increase in HF risk in case of moderate-to-severe regurgitation; and 3) wall motion score index, proportionally related with HF risk. Regarding laboratory measurements, hemoglobin and glycemia showed a well-known U-shape effect; whereas triglycerides exhibit an unexpected trend, for which HF risk decreases for higher values. Concerning comorbidities, the model included chronic kidney disease, hypertension, pulmonary disease and peripheral artery disease as risk factors, while pericardium disease as protective factor. Moreover, three categories of therapies influenced the HF risk: use of diuretics (other than loop diuretics) and anticoagulants as risk factors, RASi use as protective factor. The T axis, an ECG feature also included in the model, has no straightforward interpretation.

Although our neural network model showed a limited gain in the identification of diabetic patients that are going to develop HF, an improvement in the performance that can not be considered clinically significant, we demonstrated the feasibility of using EHRs and AI to approach the prognostic problem and obtained consistent results with respect to the clinical knowledge in the field. Moreover, our deep learning model showed adequate calibration, an important aspect of a predictive model that not always couples with discriminative ability. In particular, recent advances in deep learning methods have demonstrated incredible gains in prediction accuracy, but producing well-calibrated probabilities remains a challenge for AI tools [13, 30, 31]. In fact, this is one of the major obstacles for the use of AI tools in clinical practice for personalized medicine, since using uncalibrated predictions to determine a patient’s individual risk could led to incorrect medical decisions [32]. Our results could be relevant for future developments of prognostic risks, with a view to integrated cardiovascular prediction tools. The utility of AI applied to massive raw datasets, like ECG and echocardiograms, is being demonstrated as a powerful tool for phenotyping of cardiac conditions that can be employed at the point of care [33, 34]. In the case of ECGs, recent studies have introduced tools combining deep representations of data obtained from convolutional neural networks (in substitution to manual feature engineering) with EHRs variables [35, 36]. In the same way, for survival analysis, employing deep learning models represents the most promising and feasible way to operate in ultrahigh dimensional settings (eg. signals and images), a task that standard modeling strategies (including regularization methods) simply can not undertake. A key concern with DL approaches is the lack of transparency, since the inner-workings of such models is intrinsically a “black-box”. However, we believe that the development and application of advanced explainability techniques can provide relevant information on models’ behavior and could contribute to build the trustworthiness required for their usage in the clinical practice, as we expressed in a recent work [15].

The present study has some limitations. First, the cohort in examination is formed by individuals that underwent a cardiological evaluation. Even if cardiological assessment is highly recommended for diabetic patients, in our cohort we can not exclude the presence of selection bias towards individuals with higher cardiovascular risk with respect to the general diabetic population. In addition to this, the PHNN model include ECG and echocardiographic parameters as predictors, that can be obtained during a cardiological evaluation and this limit the applicability of the model to patients that were visited by a cardiologist. Second, our model was not validated in independent external cohorts. Future studies should be directed to measure the performance in one or more independent cohorts. Third, using explainability techniques, we are able to study and describe the marginal effect of single predictors, however we have no information on the interaction effects that could have an important role in the determination of predicted risk. Forth, in the current setting of the study possible changes in predictors variables during follow-up are not taken in account. The proposed model should be intended only as a prognostic tool at the basal evaluation.


In this study we create a prognostic tool for the management of diabetic patients at risk of developing incident HF using an AI approach that leverages the potential of EHRs. This approach may also be extended to other sources of data, like signals (ECGs) and images (echocardiography, magnetic resonance imaging).

Supporting information

S1 Fig. In blue, cumulative incidence function (CIF).

In red, 1—Kaplan-Meier curve.


S2 Fig. Features selection for the PHNN model.

Blue bars correspond to C-index obtained using the single variable. The red line corresponds to the cumulative performance on the validation set adding one variable at the time.


S3 Fig. Partial dependence plots for PHNN model.


S1 Table. Missing rate of predictors involved in the models.


S2 Table. Sentivity, specificity, positive predicted value, negative predicted value for COX and PHNN model for 2- and 5-year predictions.

Values refer to the cut-off level that obtained the higher value in the Youden’s index (Youden WJ. Index for rating diagnostic tests. Cancer 1950;3(1):32–5).



Authors would like to acknowledge the anonymous reviewers for their valuable comments and suggestions.


  1. 1. Guariguata L, Whiting DR, Hambleton I, Beagley J, Linnenkamp U, Shaw JE. Global estimates of diabetes prevalence for 2013 and projections for 2035. Diabetes Res Clin Pract. 2014 Feb;103(2):137–49. pmid:24630390
  2. 2. Zimmet P, Alberti KG, Magliano DJ, Bennett PH. Diabetes mellitus statistics on prevalence and mortality: facts and fallacies. Nat Rev Endocrinol. 2016 Oct;12(10):616–22. pmid:27388988
  3. 3. Khan SS, Butler J, Gheorghiade M. Management of comorbid diabetes mellitus and worsening heart failure. JAMA. 2014 Jun 18;311(23):2379–80. pmid:24938559
  4. 4. Nichols GA, Hillier TA, Erbey JR, Brown JB. Congestive heart failure in type 2 diabetes: prevalence, incidence, and risk factors. Diabetes Care. 2001 Sep;24(9):1614–9. pmid:11522708
  5. 5. Dei Cas A, Khan SS, Butler J, Mentz RJ, Bonow RO, Avogaro A, et al. Impact of diabetes on epidemiology, treatment, and outcomes of patients with heart failure. JACC Heart Fail. 2015 Feb;3(2):136–45.
  6. 6. Echouffo-Tcheugui JB, Xu H, DeVore AD, Schulte PJ, Butler J, Yancy CW, et al. Temporal trends and factors associated with diabetes mellitus among patients hospitalized with heart failure: Findings from Get With The Guidelines-Heart Failure registry. Am Heart J. 2016 Dec;182:9–20. pmid:27914505
  7. 7. Zinman B, Wanner C, Lachin JM, Fitchett D, Bluhmki E, Hantel S, et al. Empagliflozin, Cardiovascular Outcomes, and Mortality in Type 2 Diabetes. N Engl J Med. 2015 Nov 26;373(22):2117–28. pmid:26378978
  8. 8. Wiviott SD, Raz I, Bonaca MP, Mosenzon O, Kato ET, Cahn A, et al. Dapagliflozin and Cardiovascular Outcomes in Type 2 Diabetes. N Engl J Med. 2019 Jan 24;380(4):347–57. pmid:30415602
  9. 9. Razaghizad A, Oulousian E, Randhawa VK, Ferreira JP, Brophy JM, Greene SJ, et al. Clinical Prediction Models for Heart Failure Hospitalization in Type 2 Diabetes: A Systematic Review and Meta-Analysis. J Am Heart Assoc. 2022 May 17;11(10):e024833. pmid:35574959
  10. 10. Basu S, Sussman JB, Berkowitz SA, Hayward RA, Yudkin JS. Development and validation of Risk Equations for Complications Of type 2 Diabetes (RECODe) using individual participant data from randomised trials. Lancet Diabetes Endocrinol. 2017 Oct;5(10):788–98. pmid:28803840
  11. 11. Himes BE, Dai Y, Kohane IS, Weiss ST, Ramoni MF. Prediction of chronic obstructive pulmonary disease (COPD) in asthma patients using electronic medical records. J Am Med Inform Assoc JAMIA. 2009 Jun;16(3):371–9. pmid:19261943
  12. 12. Hulme OL, Khurshid S, Weng LC, Anderson CD, Wang EY, Ashburner JM, et al. Development and Validation of a Prediction Model for Atrial Fibrillation Using Electronic Health Records. JACC Clin Electrophysiol. 2019 Nov;5(11):1331–41. pmid:31753441
  13. 13. Guo A, Smith S, Khan YM, Langabeer Ii JR, Foraker RE. Application of a time-series deep learning model to predict cardiac dysrhythmias in electronic health records. PloS One. 2021;16(9):e0239007. pmid:34516567
  14. 14. Kaji DA, Zech JR, Kim JS, Cho SK, Dangayach NS, Costa AB, et al. An attention based deep learning model of clinical events in the intensive care unit. PloS One. 2019;14(2):e0211057. pmid:30759094
  15. 15. Gandin I, Scagnetto A, Romani S, Barbati G. Interpretability of time-series deep learning models: A study in cardiovascular patients admitted to Intensive care unit. J Biomed Inform. 2021 Sep;121:103876. pmid:34325021
  16. 16. Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Med Res Methodol. 2018 Feb 26;18(1):24. pmid:29482517
  17. 17. Lee C, Yoon J, Schaar M van der. Dynamic-DeepHit: A Deep Learning Approach for Dynamic Survival Analysis With Competing Risks Based on Longitudinal Data. IEEE Trans Biomed Eng. 2020 Jan;67(1):122–33. pmid:30951460
  18. 18. Nagpal C, Li X, Dubrawski A. Deep Survival Machines: Fully Parametric Survival Regression and Representation Learning for Censored Data With Competing Risks. IEEE J Biomed Health Inform. 2021 Aug;25(8):3163–75. pmid:33460387
  19. 19. Iorio A, Sinagra G, Di Lenarda A. Administrative database, observational research and the Tower of Babel. Int J Cardiol. 2019 Jun 1;284:118–9. pmid:30558991
  20. 20. Simon N, Friedman J, Hastie T, Tibshirani R. Regularization Paths for Cox’s Proportional Hazards Model via Coordinate Descent. J Stat Softw. 2011 Mar;39(5):1–13. pmid:27065756
  21. 21. Hastie T, Tibshirani R, Wainwright M. Statistical learning with Sparsity: the lasso and generalizations. Boca Raton: CRC Press LLC; 2015.
  22. 22. Buuren S van, Groothuis-Oudshoorn K mice: Multivariate Imputation by Chained Equations in R. J Stat Softw [Internet]. 2011 [cited 2022 Oct 4];45(3). Available from:
  23. 23. Friedman JH. Greedy function approximation: A gradient boosting machine. Ann Stat [Internet]. 2001 Oct 1 [cited 2022 Oct 4];29(5). Available from:
  24. 24. Blanche P, Dartigues JF, Jacqmin-Gadda H. Estimating and comparing time-dependent areas under receiver operating characteristic curves for censored event times with competing risks. Stat Med. 2013 Dec 30;32(30):5381–97. pmid:24027076
  25. 25. Austin PC, Harrell FE, van Klaveren D. Graphical calibration curves and the integrated calibration index (ICI) for survival models. Stat Med. 2020 Sep 20;39(21):2714–42. pmid:32548928
  26. 26. Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis. 1987;40(5):373–83. pmid:3558716
  27. 27. Robertson J, McElduff P, Pearson SA, Henry DA, Inder KJ, Attia JR. The health services burden of heart failure: an analysis using linked population health data-sets. BMC Health Serv Res. 2012 Apr 25;12:103. pmid:22533631
  28. 28. Pellicori P, Fitchett D, Kosiborod MN, Ofstad AP, Seman L, Zinman B, et al. Use of diuretics and outcomes in patients with type 2 diabetes: findings from the EMPA‐REG OUTCOME trial. Eur J Heart Fail. 2021 Jul;23(7):1085–93. pmid:34031968
  29. 29. Eleid MF, Sorajja P, Michelena HI, Malouf JF, Scott CG, Pellikka PA. Flow-gradient patterns in severe aortic stenosis with preserved ejection fraction: clinical characteristics and predictors of survival. Circulation. 2013 Oct 15;128(16):1781–9. pmid:24048203
  30. 30. Jiang X, Osl M, Kim J, Ohno-Machado L. Calibrating predictive model estimates to support personalized medicine. J Am Med Inform Assoc JAMIA. 2012 Apr;19(2):263–74. pmid:21984587
  31. 31. Rajaraman S, Ganesan P, Antani S. Deep learning model calibration for improving performance in class-imbalanced medical image classification tasks. Gadekallu TR, editor. PLOS ONE. 2022 Jan 27;17(1):e0262838. pmid:35085334
  32. 32. Krittanawong C, Johnson KW, Rosenson RS, Wang Z, Aydar M, Baber U, et al. Deep learning for cardiovascular medicine: a practical primer. Eur Heart J. 2019 Jul 1;40(25):2058–73. pmid:30815669
  33. 33. Siontis KC, Noseworthy PA, Attia ZI, Friedman PA. Artificial intelligence-enhanced electrocardiography in cardiovascular disease management. Nat Rev Cardiol. 2021 Jul;18(7):465–78. pmid:33526938
  34. 34. Litjens G, Ciompi F, Wolterink JM, de Vos BD, Leiner T, Teuwen J, et al. State-of-the-Art Deep Learning in Cardiovascular Image Analysis. JACC Cardiovasc Imaging. 2019 Aug;12(8):1549–65. pmid:31395244
  35. 35. Biton S, Gendelman S, Ribeiro AH, Miana G, Moreira C, Ribeiro ALP, et al. Atrial fibrillation risk prediction from the 12-lead electrocardiogram using digital biomarkers and deep representation learning. Eur Heart J—Digit Health. 2021 Dec 29;2(4):576–85. pmid:36713102
  36. 36. Sadasivuni S, Saha M, Bhatia N, Banerjee I, Sanyal A. Fusion of fully integrated analog machine learning classifier with electronic medical records for real-time prediction of sepsis onset. Sci Rep. 2022 Apr 5;12(1):5711. pmid:35383233