Authors and Affiliations
# Authors
Sang Mee Lee1,2; Myung Jin Chung1,3,4; Jin Oh Choi5; Kyoung Min Park5; Mi Rae Shin6;Taeyoung Kim1*; Darae Kim5*
#Affiliations
1Medical AI Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea
2Department of Health Sciences and Technology, SAIHST, Sungkyunkwan University, Republic of Korea
3Department of Data Convergence and Future Medicine, Sungkyunkwan University School of Medicine, Seoul,Republic of Korea
4Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea
5Division of Cardiology, Depart of Medicine, Heart Vascular Stroke Institute, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea
6Division of Cardiology, Depart of Medicine, Heart Vascular Stroke Institute, Samsung Medical Center, Seoul, Republic of Korea

Background
Heart Failure(HF) is one of the most prevalent cardiovascular disease globally. There are often no specific symptoms of heart failure, leading patients to visit the emergency room when acute exacerbation occurs. Furthermore, diagnosing heart failure can take few days, resulting in poor prognosis. To address this, we suggest an Artificial Intelligence(AI)-based diagnostic model for heart failure that relies on electrocardiogram(ECG) to immediately identify subtle changes and discover hidden features within the data.

Methods
We developed an AI model for normal and heart failure detection from standard 10-second, 8-lead ECG(leads I, II, and V1–6). We excluded the other 4 leads derived from leads I and II, which duplicated information. Performances of two classification models were compared: long short-term memory(LSTM) and 1-D convolutional neural network(CNN) architectures. The patients were included in the control group based on their NT-proBNPlevels being below 125 and their ejection fraction (EF) being above 50%. The heart failure patients were hospitalized following their heart failure diagnosis, and we used the ECG records closest to their admission date, ranging from 7 days before to 7 days after the admission. A total of 3,072 ECGs, with a 1:1 ratio of control (1,536) and heart failure (1,536)patients, were divided into 80% training and 20% test set.

Results
1D-CNN presented better classification performance with the area under the receiver operating characteristic curve(ROC-AUC), sensitivity and specificity of 0.925, 0.942 and 0.799, respectively. Figure 2 shows that the 1D-GradCAM for true positive cases suggests that the model gives considerable importance to the QRS segment when classifying to heart failure.

Conclusion
We propose an immediate, accurate diagnostic deep learning model for heart failure using ECGs, which provide precise, non-invasive, rapid, and easy to measure records of a patient's heart rhythm. We expect that prompt diagnosis of heart failure using ECG in primary care may potentially prevent exacerbations and reduce healthcare costs.

**Authors and Affiliations**
**# Authors**
Sang Mee Lee^1,2^; Myung Jin Chung^1,3,4^; Jin Oh Choi^5^; Kyoung Min Park^5^; Mi Rae Shin^6^;Taeyoung Kim^1*^; Darae Kim^5*^
**#Affiliations**
^1^Medical AI Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea
^2^Department of Health Sciences and Technology, SAIHST, Sungkyunkwan University, Republic of Korea
^3^Department of Data Convergence and Future Medicine, Sungkyunkwan University School of Medicine, Seoul,Republic of Korea
^4^Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea
^5^Division of Cardiology, Depart of Medicine, Heart Vascular Stroke Institute, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea
^6^Division of Cardiology, Depart of Medicine, Heart Vascular Stroke Institute, Samsung Medical Center, Seoul, Republic of Korea

**Background**
Heart Failure(HF) is one of the most prevalent cardiovascular disease globally. There are often no specific symptoms of heart failure, leading patients to visit the emergency room when acute exacerbation occurs. Furthermore, diagnosing heart failure can take few days, resulting in poor prognosis. To address this, we suggest an Artificial Intelligence(AI)-based diagnostic model for heart failure that relies on electrocardiogram(ECG) to immediately identify subtle changes and discover hidden features within the data.

**Methods**
We developed an AI model for normal and heart failure detection from standard 10-second, 8-lead ECG(leads I, II, and V1–6). We excluded the other 4 leads derived from leads I and II, which duplicated information. Performances of two classification models were compared: long short-term memory(LSTM) and 1-D convolutional neural network(CNN) architectures. The patients were included in the control group based on their NT-proBNPlevels being below 125 and their ejection fraction (EF) being above 50%. The heart failure patients were hospitalized following their heart failure diagnosis, and we used the ECG records closest to their admission date, ranging from 7 days before to 7 days after the admission. A total of 3,072 ECGs, with a 1:1 ratio of control (1,536) and heart failure (1,536)patients, were divided into 80% training and 20% test set.

**Results**
1D-CNN presented better classification performance with the area under the receiver operating characteristic curve(ROC-AUC), sensitivity and specificity of 0.925, 0.942 and 0.799, respectively. Figure 2 shows that the 1D-GradCAM for true positive cases suggests that the model gives considerable importance to the QRS segment when classifying to heart failure.

**Conclusion**
We propose an immediate, accurate diagnostic deep learning model for heart failure using ECGs, which provide precise, non-invasive, rapid, and easy to measure records of a patient's heart rhythm. We expect that prompt diagnosis of heart failure using ECG in primary care may potentially prevent exacerbations and reduce healthcare costs.

Deep Learning Algorithm for Detection of Acute Heart Failure Using Standard ECG waveforms