Heart failure is a problem that affects the patient’s normal oxygen supply. Many people around the world suffer from this disease and the number is increasing due to factors of modern life. The aim of the project is to use neural networks to integrate clinical data and echocardiographic structures to develop an accurate heart failure detection system. Long-term memory modeling (LSTM) was used to process clinical data from the heart failure prediction dataset and MobileNet architecture was used to analyze echocardiogram frames from the EchoNet-Dynamic dataset developed at Stanford University. The results of these models are then combined using model averaging to improve overall performance. The LSTM model achieved 91% accuracy, 94% recall and 93% F1-score based on clinical data. On the other hand, the MobileNet architecture achieved 84% accuracy, 66.15% recall and 65.47% F1-score on echocardiogram frames. The combined model achieved an average of 88% accuracy, 80% recovery and 79% F1-score, highlighting the benefits of integrating multiple sources. This method not only improves diagnostic accuracy, but also provides an efficient solution for heart failure detection, providing valuable clinical decision support in cardiology.