Diagnostic Deep Learning Framework for Heart Failure

doi:10.21203/rs.3.rs-4005522/v1

Download PDF

Research Article

Diagnostic Deep Learning Framework for Heart Failure

https://doi.org/10.21203/rs.3.rs-4005522/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

In the precision medicine era, leveraging advanced technology, including deep learning, has the potential to enhance diagnostic accuracy across various diseases. However, chest radiography (CXR), pivotal for heart failure (HF) diagnosis, currently has limited precision.

Methods

Through a retrospective cohort study encompassing 144 participants from the RAMA dataset at Ramathibodi Hospital (spanning January 1, 2010, to December 31, 2019), 240 HF CXR images were scrutinized and annotated by cardiologists and radiologists. Clinical diagnosis was confirmed by cardiologists using HF signs and symptoms, pulmonary capillary wedge pressure, natriuretic peptide, and ejection fraction. The developed model, HFNet, was trained on the RAMA dataset and incorporated these clinical factors.

Results

This study evaluated the performance of the HFNet model in predicting radiographic findings related to heart failure and achieved excellent results. The model demonstrated high precision (for cardiomegaly, 1.0; for pulmonary edema, 0.9; for pleural effusion, 0.8) and good accuracy (for cardiomegaly, 0.9; for pulmonary edema, 0.6; and for pleural effusion, 0.7), coupled with respective AUC values of 1.00, 0.96, and 0.49. Concomitant recall figures stood at 0.9, 0.6, and 0.7, while F1 scores were 0.9 for cardiomegaly and 0.7 for both pulmonary edema and pleural effusion. These findings highlight the potential of HFNet to aid clinicians in the precise detection and diagnosis of HF-associated radiographic cues.

Conclusions

The development of the HFNet model introduces a promising tool for clinicians, facilitating accurate and precise diagnosis of HF-related radiographies.

Artificial Intelligence (AI)

Chest X-ray (CXR)

Deep Learning

Heart Failure (HF)

Pulmonary Capillary Wedge Pressure (PCWP)

The rapid advancement of artificial intelligence (AI) has revolutionized medical technology and has had a significant impact on clinical practice. Machine learning algorithms, such as deep learning algorithms, can analyze large amounts of data, learn from them, and evaluate clinical findings. In the medical field, AI models are often compared to human performance using receiver operating characteristic (ROC) curves, with accuracy assessed using the area under the curve (AUC). This approach has enabled the development of highly accurate models for various medical conditions, including heart failure, as demonstrated by recent studies that have used chest X-ray (CXR) imaging. (1–7)

Heart failure (HF) is a clinical syndrome characterized by typical symptoms and signs of cardiac dysfunction and was identified from medical history and physical examination and supported by laboratory and imaging assessments of structural or functional abnormalities of the heart. Although chest radiography (CXR) is a basic investigational tool for assessing HF incidence, its diagnostic performance is limited by low sensitivity (67–68%) and specificity (76–83%). (8–10)

The diagnosis of heart failure can be aided by several clinical parameters, including the left ventricular ejection fraction (LVEF), serum N-terminal pro-B-type natriuretic peptide (NT-proBNP) level, and pulmonary capillary wedge pressure (PCWP). (8–10) Recent studies have demonstrated the efficacy of AI in detecting heart failure through CXR images, with a reported diagnostic accuracy of up to 82% and an AUC of 0.82. Nevertheless, there is a lack of reported machine learning methods that incorporate clinical data with additional parameters such as the PCWP. (9–12)

In this study, Ramathibodi Hospital (RAMA) developed an artificial intelligence (AI) model to detect heart failure (HF) by utilizing 112,000 chest X-ray (CXR) images obtained from the National Institutes of Health. The model was constructed based on the radiographic data of CheXNet, which included 14 CXR findings. Specific findings, such as edema, cardiomegaly, and pleural effusion, which are associated with heart failure, were selected for model development. (13) The model underwent rigorous training, testing, and validation before applying the transferred learning method using RAMA data to develop a preliminary HF model (DenseNet121). The model was pretrained on pulmonary edema and achieved an area under the curve (AUC) of 0.9. Finally, the transferred learning approach was applied to the RAMA dataset to develop HFNet.

The aim of this study was to develop an AI model to assist clinicians in diagnosing heart failure using CXR findings with other concomitant clinical parameters. The use of AI-based techniques, particularly complex machine learning methods, offers numerous opportunities in cardiovascular research and clinical practice. In the context of heart failure management, the application of ML techniques has the potential to improve patient outcomes and enhance the efficiency of HF management. (5, 7) The development of the HFNet represents a significant step forward in the diagnosis of heart failure and could lead to improved patient outcomes.

Study Samples

We enrolled 144 participants in a population-based retrospective cohort study at Ramathibodi Hospital. The data were collected from electronic medical records from January 1, 2010, to December 31, 2019.

Adults (> 18 years of age) who had a clinical diagnosis of heart failure (New York Heart Association functional class [NYHA FC] I-IV) and who underwent evaluation via chest X-ray (CXR), echocardiography, or pulmonary capillary wedge pressure (PCWP) were eligible for this study. Pregnant participants and those with complex congenital heart disease, right-sided heart failure, or end-stage malignancy were excluded.

We collected data on baseline characteristics at the time of HF presentation, including age, sex, body weight, height, body mass index, symptoms and signs of HF, comorbidities, medications, devices, blood pressure, heart rate, NYHA FC, LVEF by echocardiography, NT-proBNP, and PCWP.

Ascertainment of HF diagnosis

We used the ICD-10 codes 500, 501, and 509 for the HF diagnosis. The data for HF diagnosis were obtained from electronic medical records, where we reviewed the clinical findings of HF and evaluated the chest X-ray (CXR) data. The diagnosis of HF was made by the presence of the signs and symptoms of HF, with typical HF CXR and abnormal clinical parameters: elevated NT-proBNP, abnormal echocardiography, and/or elevated PCWP. A typical HF-CXR was defined by evidence of redistribution, interstitial edema, alveolar edema, cardiomegaly, and/or pleural effusion; abnormal echocardiography findings, such as an LVEF < 40% (HF with reduced ejection fraction), an LVEF ≥ 40% plus left ventricular hypertrophy, left atrial enlargement, or diastolic dysfunction (HF with mildly reduced ejection fraction and HF with preserved ejection fraction). An elevated PCWP, indicated by a PCWP > 18 mmHg, was determined through right heart catheterization. (9–10, 14–16)

Dataset

We obtained 240 HF CXRs from 144 participants from Ramathibodi Hospital. The RAMA dataset was reviewed and verified by two cardiologists and two radiologists, and a consensus was reached with at least two of the four reviewers in agreement. Prior to interpretation, the CXRs were labeled by consensus as showing pulmonary congestion (interstitial edema or alveolar edema), cardiomegaly, or pleural effusion. Two cardiologists verified the clinical diagnosis based on the signs and symptoms of HF, PCWP, NT-proBNP, and LVEF. The images were then resized from 1024 × 1024 to 256 × 256 pixels, and the final metric was flattened. The final images for training, testing, and validation were selected from the 240 images, with a ratio of 70:15:15 for training, testing, and validation, respectively (17, 18). Feature vectors consisted of clinical factors and dense data, and the HFNet results were used to evaluate the AUC of pulmonary congestion, cardiomegaly, and pleural effusion along with their concomitant clinical parameters (NT-proBNP, LVEF by echocardiography, and PCWP).

Deep learning modeling

Our model, HFNet, was built using Keras version 2.3.1 and Python version 3.7.9. In the initial phase of our study, we developed a pretrained model, DenseNet121, based on reference data from the CheXNet dataset. (13) Subsequently, we customized the model by removing the last layer and adding three dense layers to specifically target three layers that are associated with heart failure, namely, edema, cardiomegaly, and effusion. This approach enabled us to enhance the ability of the model to accurately identify and analyze these specific findings.

Handling Missing Data

To ensure robustness in real-world scenarios, we carefully addressed missing data in our model. When dealing with clinical vectors such as the PCWP, which may have missing values, we employed a data preprocessing approach. Missing data were imputed using the mean value of the respective feature, as this provided a reasonable approximation given the nature of our dataset.

Subsequently, we extended the layers of the model and performed transfer learning on the RAMA dataset. Using this approach, we were able to develop the HFNet model, which is specifically tailored to identify and analyze findings related to heart failure. The HFNet architecture consists of an input layer that takes the (N, 256, 256, 3) input tensor, where N is the number of samples. This input flows to the next convolutional layer (CONV2D) with a tensor shape (N, 252, 252, 128) for feature extraction. The next layer is a max pooling layer for downsampling to a tensor size of (N, 126, 126, 128). Thereafter, we applied a flattened layer to transform the three-dimensional tensor into a feature vector (N, 2032128) to concatenate it with another clinical feature vector. These concatenated features were connected to two dense layers of 16 neurons using the ReLu activation function, where dropout was applied. Finally, for the multilabel classification problem, the output layer was a dense layer of (N, 3) to classify the three radiological findings. Our unique feature model differs from conventional AI models in that we did not develop it solely from input data obtained from chest X-ray imaging. Instead, our model utilizes clinical vectors such as the PCWP in conjunction with imaging data. The overall architecture is illustrated in Fig. 1.

Our strategy applied mini-batch training, where the batch size was set to 16. Early stopping was performed to prevent overfitting. The model stopped learning when the validation loss was no longer improved for 20 epochs, and the best model parameters were restored from the best epoch, as shown in Fig. 2. The dynamic learning rate was applied using the ReduceLROnPlateau, which reduced the learning rate by 10% if the validation loss was no longer improved five consecutive times. The initial learning rate was set to 0.001.

Sample size calculation

A power analysis was conducted using G*Power software to determine the appropriate sample size for our study. We selected the "linear multiple regression: fixed model, R² deviation from zero" analysis type since our AI model includes three predictors related to edema, effusion, and cardiomegaly in chest X-ray for assisting heart failure diagnosis. An effect size of 0.15 was deemed small based on our understanding of the research question. The significance level (alpha error) was set to 0.05, representing the probability of rejecting the null hypothesis when it is actually true. We selected a desired power of 0.95, representing the probability of correctly rejecting the null hypothesis when it was actually false. Based on these criteria, GPower software was used to calculate a sample size of 119 participants to achieve a power of 0.95. This sample size was determined to be adequate for detecting a small effect size with a 95% confidence level for our AI model's ability to detect congestive findings via chest X-ray.

Statistical analysis

The baseline characteristics are presented as percentages for categorical variables and as the means with standard deviations for continuous variables. For continuous variables that were not normally distributed, medians and interquartile ranges (IQRs) were used for presentation. The normality of the data was assessed using the one-sample Kolmogorov‒Smirnov test. To evaluate the predictive accuracy of the models, the area under the curve (AUC) was calculated. IBM SPSS Statistics version 23 was used for conducting the statistical analyses.

Clinical and radiographic features of heart failure at baseline

In total, 144 participants were screened for eligibility. Among the study population, 70 (48.6%) were men, and 74 (51.4%) were women. Overall, the mean age was 65.1 ± 17.3 years. The clinical parameters are reported as follows: the NT-proBNP level was 5,734 (2564–20,218) pg/dL, the LVEF was 41.75% (25–60%), and the PCWP was 22 (14–27) mmHg. The baseline characteristics of the study population are presented in Table 1. The radiographic findings on chest X-rays suggestive of heart failure (HF) included redistribution in 149 (62%), interstitial pulmonary edema in 190 (79%), alveolar pulmonary edema in 73 (30%), cardiomegaly in 236 (98%), and pleural effusion in 138 (57.5%). Additional radiographic findings are presented in Table 2.

Table 1

Baseline clinical characteristics of study participants
Characteristics*	ALL (N = 144)
Age (years), Mean ± SD	65.1 ± 17.3
Male, n (%)	70 (48.6)
Symptoms of heart failure, n (%)
Breathlessness	93 (64.6)
Orthopnea	90 (60.3)
Paroxysmal nocturnal dyspnea	41 (28.5)
Reduced exercise intolerance	26 (18.1)
Fatigue	59 (41)
Increased time to recovery after exercise	17 (11.8)
Ankle swelling	65 (45.1)
Signs of heart failure, n (%)
Elevated JVP	39 (27.1)
Hepatojugular reflux	4 (2.8)
Laterally displaced apical impulse	12 (8.3)
S3 gallop	11 (7.6)
Extremities edema	74 (51.4)
Clinical parameters, Median, IQR (Q1-Q3)
Systolic blood pressure (mmHg)	126 (107–145)
Diastolic blood pressure (mmHg)	65.5 (55.8–80)
Heart rate (bpm)	84 (72.5–100)
BMI (kg/m²)	23.8 (20.7–26.7)
Creatinine (mg/dL)	1.25 (0.9–1.9)
eGFR (mL/min/1.73m²)	52.7 (28.7–71.2)
NT-ProBNP (pg/dL)	5,734 (2,564–20,218)
LVEF (%)	41.75 (25–60)
PCWP (mmHg)	22 (14–27)
NYHA FC I	28 (19.4)
NYHA FC II	38 (26.4)
NYHA FC III	48 (33.3)
NYHA FC IV	19 (13.2)
Comorbidities, n (%)
Type 2 Diabetes Mellitus	53 (36.8)
Hypertension	86 (59.7)
Dyslipidemia	59 (41)
Coronary artery disease	46 (31.9)
Chronic kidney disease	53 (36.8)
Stroke	17 (11.8)
Atrial fibrillation	44 (30.6)
Obstructive sleep apnea	10 (6.9)
Medications, n (%)
ACEI/ARBs	48 (33.4)
ARNIs	11 (7.6)
BBs	85 (59)
MRAs	38 (26.4)
SGLT2 inhibitors	11 (7.6)
Ivabradine	8 (5.6)
Hydralazine/Nitrates	13 (9)
Digoxin	11 (7.6)
Furosemide	92 (63.6)
Devices, n (%)
CRT-P	1 (0.7)
CRT-D	8 (5.6)
AICD	6 (4.2)
PPM	6 (4.2)
Values are n (%) for dichotomous variables, Median, IQR (Interquartile Range) (Q1-Q3) for continuous variables. *JVP jugular venous pressure, LVEF Left ventricular ejection fraction, PCWP Pulmonary capillary wedge pressure, NYHA New York Heart Association, ACEI Angiotensin-converting enzyme (ACE) inhibitors, ARNI Angiotensin Receptor Blocker/Neprilysin Inhibitors, BBs. Betablockers. MRAs Mineralocorticoid Receptor Antagonists, SGLT2 inhibitors Sodium-Glucose Cotransporter 2 inhibitors, CRT-P cardiac resynchronization therapy with pacemaker, CRT-D cardiac resynchronization therapy defibrillator, AICD Automated Implantable Cardioverter-Defibrillator, PPM Permanent Pacemaker

Table 2

Radiographic Findings of Heart Failure in Chest X-ray Images
Findings*	ALL (N = 240)
Typical HF Chest X-ray Findings, n (%)
Redistribution	149 (62)
Interstitial pulmonary edema	190 (79)
Alveolar pulmonary edema	73 (30)
Cardiomegaly	236 (98)
Pleural effusion	138 (57.5)
Other findings, n (%)
Atelectasis	33 (13.8)
Consolidation	19 (7.9)
Patchy or Ground glass opacities	10 (4.1)
Fibrosis	4 (1.6)
Nodule or Mass	0
Presence of ETT, catheter, or devices	100 (41.7)
Values are n (%) for dichotomous variables. *HF Heart Failure, ETT Endotracheal tube

HFNet performance

As the learning progressed, HFNet achieved the following prediction performance, as indicated by the area under the curve (AUC): pulmonary edema (0.96), cardiomegaly (1.00), and pleural effusion (0.49), as shown in Fig. 3. The results of our study indicate that the model achieved a high level of precision (1.0) in detecting cardiomegaly while achieving good precision (0.9 and 0.8) in detecting pulmonary edema and pleural effusion, respectively. The accuracy of the model was 0.9 for cardiomegaly and 0.6 and 0.7 for pulmonary edema and pleural effusion, respectively. The recall of the model was 0.9 for cardiomegaly and 0.6 and 0.7 for pulmonary edema and pleural effusion, respectively.

Furthermore, the F1 score, which provides a balance between precision and recall, was high for detecting cardiomegaly (0.9) and for detecting pulmonary edema and pleural effusion (0.7). Notably, the AUC for pulmonary edema was particularly high at 0.96, indicating excellent performance in this area.

Overall, our results demonstrate that the model performs well in detecting cardiomegaly and has good performance in detecting pulmonary edema and pleural effusion; however, there is room for improvement in terms of accuracy and recall for these latter conditions. These findings have important implications for improving the accuracy of clinical diagnosis and treatment planning.

In this study, we developed the HFNet AI model, which integrates clinical parameters with deep learning to improve the accuracy of HF diagnosis via radiographic findings. Our model's ability to predict useful radiographic findings of HF can assist clinicians in diagnosing HF, especially in centers without specialized personnel for difficult cases. Unlike recent studies that rely solely on radiographic findings or elevated PCWP, our model incorporates clinical parameters, such as PCWP and chest X-ray imaging, to improve the performance of the AI model for heart failure diagnosis. A comprehensive diagnostic approach, which included clinical presentation, physical examination, laboratory tests, and imaging studies, was used to verify the diagnosis of HF. This approach minimized the risk of misdiagnosis and ensured a robust diagnosis of HF.(19–22)

Our study investigated the performance of the HFNet model in predicting imaging findings of pulmonary edema, cardiomegaly, and pleural effusion. We found that as learning progressed, the model achieved high prediction performance for cardiomegaly, with an AUC of 1.00, and for pulmonary edema, with an AUC of 0.96, indicating high accuracy in distinguishing between positive and negative cases. However, for pleural effusion, the AUC was only 0.49, indicating poor performance in predicting this condition.

Furthermore, our study demonstrated that the model had high precision for detecting cardiomegaly, with a score of 1.0, and good precision for detecting pulmonary edema and pleural effusion, with scores of 0.9 and 0.8, respectively. The accuracy of the model for detecting cardiomegaly was 0.9, while for pulmonary edema and pleural effusion, it was 0.6 and 0.7, respectively. The recall of the model for detecting cardiomegaly was 0.9, while for pulmonary edema and pleural effusion, it was 0.6 and 0.7, respectively. The F1 score, which provides a balance between precision and recall, was high for detecting cardiomegaly (0.9), while for detecting pulmonary edema and pleural effusion, it was 0.7 for both.

Based on our findings, the HFNet model shows promising potential as a tool for accurately detecting cardiomegaly and pulmonary edema. However, the model's poor performance in detecting pleural effusion indicates the need for further improvements to enhance its sensitivity and accuracy in detecting this condition. Notably, recent guidelines have reported that chest X-ray findings in HF patients are primarily associated with congestion, such as pulmonary venous congestion, interstitial edema, alveolar edema, and cardiomegaly. (9, 10) Therefore, our model's poor performance in detecting pleural effusion is consistent with the clinical setting of heart failure patients.

In our study, we considered the potential issue of data leakage, where patients may have undergone multiple imaging studies. However, we took steps to prevent any overlap between the dataset groups by ensuring that each patient was included in only one group. Our model was trained via transfer learning on the Chest X-ray14 dataset and fine-tuned on the Rama dataset via the DenseNet-121 architecture. (13) We also integrated early stopping during the training process to mitigate the risk of overfitting, as evidenced by the comparable performance results between the training and validation sets. The early stopping mechanism was triggered if the model’s performance did not improve after more than 20 epochs, as illustrated in Fig. 2. Overall, our findings suggest that our model has been effective at circumventing overfitting concerns, which may have implications for future research in this area.

In previous studies, AI models have been developed to detect HF on CXRs without clinical factors, with a reported AUC of 0.82. (11) However, these models have limitations in their accuracy, as they do not consider important clinical factors that can aid in the diagnosis of HF. In contrast, our study incorporated clinical parameters, such as the PCWP, along with radiographic findings to develop an artificial intelligence (AI) model, resulting in significantly improved prediction performance, with an area under the curve (AUC) of 0.96 for the detection of pulmonary congestion.

The results of our study underscore the significant improvement in diagnostic accuracy and potential patient outcomes achieved through the integration of clinical factors with radiographic findings. These findings offer valuable insights into the development of effective AI models for diagnosing heart failure (HF), with implications for both future research and clinical application.

The integration of clinical parameters with radiographic data has proven to be a pivotal advancement in AI-assisted HF diagnosis. Our model's incorporation of crucial clinical factors, such as pulmonary capillary wedge pressure (PCWP), has yielded superior predictive performance compared to studies solely relying on chest radiography (CXR) images (19). This demonstrates the power of a comprehensive approach in HF detection.

Acknowledging the limitations inherent in our study is crucial. The single-center design, while yielding insightful results, may temper the broad application of our findings, particularly in centers with distinct imaging protocols. Nonetheless, our research holds significance in contexts marked by resource constraints, offering a valuable perspective on AI model development for HF diagnosis. Despite our study's modest sample size, the inclusion of pertinent clinical parameters bolstered the model's performance, emphasizing its reliability. To establish its robustness and applicability, further validation using larger and diverse datasets is imperative.

Our investigation underscores the capacity of the HFNet as an adjunctive tool, substantially elevating the accuracy of detecting CXR-based HF-related radiographic findings. Harnessing the insights gained from clinical features such as PCWP, our research amplifies the capabilities of deep learning models, particularly in identifying cardiomegaly and pulmonary edema. While recognizing the limitations inherent to our study, including its small sample size and single-center design, we accentuate the need for rigorous validation on broader datasets to ensure the model's robustness and applicability. The development of HFNet as an AI-based diagnostic tool signifies a significant advancement in HF diagnosis, necessitating further exploration of its clinical utility and potential to influence patient outcomes.

ACEIs, angiotensin-converting enzyme (ACE) inhibitors

AICD, automated implantable cardioverter-defibrillator

AI, artificial intelligence

ARNI, angiotensin receptor blocker/neprilysin inhibitor

AUC, area under the curve

BBs, Beta-blockers

CRT-D, Cardiac resynchronization therapy defibrillator

CRT-P, Cardiac resynchronization therapy with pacemaker

CXR, Chest radiography

ETT, Endotracheal tube

HF, Heart failure

ICU, intensive care unit

CCU, Cardiac Care Unit

IQR, Interquartile Range

JVP, Jugular venous pressure

LVEF, Left ventricular ejection fraction

MRAs, mineralocorticoid receptor antagonists

NT-ProBNP, N-terminal pro-B-type natriuretic peptide

NYHA FC, New York Heart Association functional class

PCWP, pulmonary capillary wedge pressure

PPM, Permanent Pacemaker

RAMA, Ramathibodi Hospital

ROC, Receiver operating characteristic

SGLT2 inhibitor, sodium-glucose cotransporter 2 inhibitor

Ethics approval and consent to participate

The study was approved by the Human Research Ethics Committee, Faculty of Medicine Ramathibodi Hospital, and was conducted according to the International Conference on Harmonization of Technical Requirements for Pharmaceuticals for Human Use (ICH), Good Clinical Practice Guidelines and the Declaration of Helsinki. All patients provided written informed consent.

Consent for publication

Informed consent was obtained from all individual participants included in the study. The authors affirm that the human research participants provided informed consent for publication of the data in Table 1 and Table 2.

Data availability statement

The data supporting the findings of this study are available within the article. The dataset of X-ray images that supports the findings of this study is available from the authors, but restrictions apply to the availability of these data. The dataset was used under license from the Ramathibodi Hospital for the current study and is not publicly available. However, the data are available from the authors upon reasonable request and with permission from the Ramathibodi Hospital.

Competing interests

There are no conflicts of interest related to this study. The authors collected and accessed all the data and take responsibility for the integrity of the study and for the data analysis.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Authors’ contributors

WC, TP, RT, SP, AS, and CN contributed to the conception and design of the study. WC organized the database. WC, TP, AS, and CN verified the database. RT and SL developed the AI model.

Acknowledgments

We extend our gratitude to the patients who participated in this study. Special thanks to the BMC editors and reviewers for their valuable contributions in scrutinizing and enhancing this manuscript.

Lal A, Pinevich Y, Gajic O, Herasevich V, Pickering B. Artificial intelligence and computer simulation models in critical illness. World J Crit Care Med. 2020;9(1):13–9. DOI: 10.5492/wjccm.v9.i1.13.
Yasmin F, Shah SM, Naeem A, Shujauddin SM, Jabeen A, Kazmi S, Siddiqui SA, Kumar P, Salman S, Hassan SA, Dasari C. Artificial intelligence in the diagnosis and detection of heart failure: the past, present, and future. Rev Cardiovasc Med. 2021 Dec 22;22(4):1095-113. DOI: 10.31083/j.rcm2204119.
Zaharchuk G, Gong E, Wintermark M, Rubin D, Langlotz CP. Deep learning in neuroradiology. AJNR Am J Neuroradiol. 2018;39(10):1776–84. DOI: 10.3174/ajnr.A5544.
Choi DJ, Park JJ, Ali T, Lee S. Artificial intelligence for the diagnosis of heart failure. NPJ Digit Med. 2020 Apr 8;3(1):54. DOI: 10.1038/s41746-020-0264-4.
Javeed A, Khan SU, Ali L, Ali S, Imrana Y, Rahman A. Machine learning-based automated diagnostic systems developed for heart failure prediction using different types of data modalities: a systematic review and future directions. Comput Math Methods Med. 2022 Feb 3;2022. DOI: 10.1155/2022/2643803.
Malik A, Brito D, Vaqar S, Chhabra L. Congestive heart failure. In: StatPearls [Internet]. StatPearls Publishing; 2022 May 5.
Averbuch T, Sullivan K, Sauer A, Mamas MA, Voors AA, Gale CP, Metra M, Ravindra N, Van Spall HG. Applications of artificial intelligence and machine learning in heart failure. Eur Heart J-Digital Health. 2022 Jun 22;3(2):311-22. DOI: 10.1093/ehjdh/ztab 028.
Cardinale L, Volpicelli G, Lamorte A, Martino J, Andrea Veltri. Revisiting signs, strengths, and weaknesses of standard chest radiography in patients of acute dyspnea in the emergency department. J Thorac Dis. 2012;4:398–407.
McDonagh TA, Metra M, Adamo M, et al. 2021 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure. Eur Heart J. 2021 Sep 21;42(36):3599-726. DOI: 10.1093/eurheartj/ehab368.
Heidenreich PA, Bozkurt B, Aguilar D, Allen LA, Byun JJ, Colvin MM, Deswal A, Drazner MH, Dunlay SM, Evers LR, Fang JC. 2022 AHA/ACC/HFSA guideline for the management of heart failure: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Am Coll Cardiol. 2022 May 3;79(17):e263-421. DOI: 10.1016/j.jacc.2022.02.020.
Matsumoto T, Kodera S, Shinohara H, Ieki H, Yamaguchi T, Higashikuni Y, et al. Diagnosing heart failure from chest X-ray images using deep learning. Int Heart J. 2020;61:781–6. DOI: 10.1536/ihj.20-031.
Seah JCY, Tang JSN, Kitchen A, Gaillard F, Dixon AF. Chest radiographs in congestive heart failure: Visualizing neural network learning. Radiology. 2019;290:514–22. DOI: 10.1148/radiol.2018180766.
Rajpurkar P, Irvin J, Zhu K, Yang B, Mehta H, Duan T, Ding D, Bagul A, Langlotz C, Shpanskaya K, Lungren MP. Chexnet: Radiologist-level pneumonia detection on chest X-rays with deep learning. arXiv preprint arXiv:1711.05225. 2017 Nov 14.
Ma TS, Bozkurt B, Paniagua D, et al. Central venous pressure and pulmonary capillary wedge pressure: fresh clinical perspectives from a new model of discordant and concordant heart failure. Tex Heart Inst J. 2011;38(6):627. DOI: 10.14503/THIJ-10-3215.
Van Aken H, Vandermeersch E. Reliability of PCWP as an index for left ventricular preload. Br J Anaesth. 1988;60:85S-89S. DOI: 10.1093/bja/60.6.85S.
Gidwani UK, Mohanty B, Chatterjee K. The pulmonary artery catheter: a critical reappraisal. Cardiol Clin. 2013;31(4):545-565. DOI: 10.1016/j.ccl.2013.07.001.
Joseph VR. Optimal ratio for data splitting. Stat Anal Data Min. 2022;15(4):531-8. DOI: 10.1002/sam.11552.
Hrehova S, Knapcikova L. The Study of Machine Learning Assisted the Design of Selected Composites Properties. Appl Sci. 2022;12(21):10863. DOI: 10.3390/app122110863.
Celik A, Surmeli AO, Demir M, Esen K, Fural O, Camsari A. The early diagnostic value of chest X-ray scanning by the help of artificial intelligence in heart failure (ART-IN-HF): The first outcomes. J Am Coll Cardiol. 2022;79(9_Supplement):395. DOI: 10.1016/j.jacc.2022.03.1020.
Błaziak M, Urban S, Wietrzyk W, Jura M, Iwanek G, Stańczykiewicz B, Kuliczkowski W, Zymliński R, Pondel M, Berka P, Danel D. An Artificial Intelligence Approach to Guiding the Management of Heart Failure Patients Using Predictive Models: A Systematic Review. Biomedicines. 2022;10(9):2188. DOI: 10.3390/biomedicines10092188.
Siegersma KR, Leiner T, Chew DP, Appelman Y, Hofstra L, Verjans JW. Artificial intelligence in cardiovascular imaging: state of the art and implications for the imaging cardiologist. Neth Heart J. 2019 Sep;27(9):403-413. DOI: 10.1007/s12471-019-01301-3. PMID: 31278552; PMCID: PMC6665107.
Hirata Y, Kusunose K, Tsuji T, Fujimori K, Kotoku JI, Sata M. Deep learning for detection of elevated pulmonary artery wedge pressure using standard chest X-ray. Can J Cardiol. 2021 Aug;37(8):1198-1206. DOI: 10.1016/j.cjca.2020.12.018. Epub 2020 Dec 31. PMID: 33573877.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Diagnostic Deep Learning Framework for Heart Failure

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Introduction

Methods

Statistical analysis

Results

Clinical and radiographic features of heart failure at baseline

HFNet performance

Discussion

Conclusion

Abbreviations

Declarations

Funding

Acknowledgments

References

Additional Declarations

Status:

Version 1