AI in predicting ICD-10 codes and prescriptions medicine for heart failure patients

Lack of training and socialization, inconsistent classification, inappropriate identification, differences in interpretation, and inaccuracy in certain conditions, coupled with the non-widespread adoption of ICD-10 which can have an impact on the inefficiency of drug management in hospitals, it is necessary to integrate the ICD-10-based medical recording system with the drug prescription system, so this study was conducted. This research aims to create an AI-based application that incorporates a cloud-based Long Short-Term Memory (LSTM) model that can forecast ICD-10 diagnoses and medication prescriptions for heart failure to improve inventory management. Method: AI with LSTM model integration is applied to a dataset of heart failure drug prescriptions. First, data collection of heart failure drug prescriptions, data pre-processing, and data allocation are carried out so that two types of models are obtained, namely a model for predicting ICD-10 diagnoses and a model for predicting drugs. Furthermore, validation is carried out on both prediction models using the Binary Cross Entropy Validation method. Finally, the application development was carried out with three main modules, namely the patient's personal data module, physical and laboratory examinations, and diagnosis and drug recommendations. Results: “The results of the LSTM model integration on the ICD-10 diagnosis prediction showed a training loss of 0.5859 in the first epoch and decreased by 0.3157 in epoch 18, and the training accuracy showed 0.7867 in the first epoch and 0.8538 in epoch 18. The validation loss value was 0.4514 in the first epoch and 0.3229 in epoch 18 and the validation accuracy was 0.7962 in the first epoch and 0.8512 in epoch 18. However, after epoch 18 the validation accuracy decreased slightly. The results of the LSTM model integration on the prediction of ICD-10 secondary diagnoses show that the training loss is 0.6425 in the first epoch, and 0.0842 in the 20th epoch and the training accuracy shows 0.7139 in the first epoch to 0.9790 in the 20th epoch. The validation loss figure shows 0.3839 in the first epoch and 0.0801 in the last epoch, and the validation accuracy shows a value of 0.9800 from the second epoch to the last epoch. The results of the LSTM model integration on drug prediction show that the training loss value is 0.6747 in the first epoch and 0.0733 in the 20th epoch, and the training accuracy from 0.7672 in the first epoch to 0.9784 in the 20th epoch. The validation loss value decreased from 0.5123 in the first epoch to 0.0712 in the 20th epoch, and the validation accuracy stabilized around 0.9787 from the second epoch to the last epoch. The model showed convergence around the 12th epoch, and there were no signs of overfitting.” Conclusion: The LSTM model performed quite well in predicting ICD-10 diagnoses and prescriptions for heart failure, with an architecture optimized to analyze medical text data and provide accurate predictions.