Meta Learning for Enhanced Disease Prediction from EHR Data

The healthcare sector is becoming more dependent on electronic health records (EHR) for disease forecasting, risk evaluation, and mortality analysis. Although AI-driven models have enhanced disease prediction, they frequently focus on common diseases and face difficulties with new or rare diseases. Furthermore, these models require large datasets for better accuracy, posing challenges in diverse or limited-data scenarios. To solve these issues, this research proposes a novel Long Short-Term Memory (LSTM)-Attention network-based meta-learning framework for prediction tasks using time-series data from EHRs. The framework is designed to address challenges such as limited sample sizes, imbalanced labels, and the ability to predict unseen diseases. The proposed model is capable of handling multiple tasks related to irregular patterns and anomalies in time-series signals. The meta-learning approach enables the system to leverage knowledge from previous tasks, enhancing its ability to predict new and previously unseen diseases from ECG data. The proposed LSTM-Attention model is evaluated against conventional models like Support Vector Machine (SVM), Random Forest (RF), and XGBoost. Experimental results demonstrate that the proposed model outperforms these models, achieving superior performance in predicting HRV, arrhythmia, and abnormalities from ECG signals. The LSTM-Attention model achieves the highest accuracy (0.92), precision (0.90), recall (0.91), F1 score (0.91), and ROC-AUC (0.93). Moreover, the prediction time for the proposed model is 95 seconds, significantly faster than other models.