Deep Autoencoder-Based Model for Multi-Class Disease Prediction Using Symptom Data

Abstract

Critical fact: Proper prediction of diseases by use of symptoms data plays a critical role in contemporary health care systems. This paper will introduce a combined deep autoencoder based multi-class disease predictor model, which can represent complex and non-linear feature-level interactions with a high-dimensional feature symptom matrices through feature loss and diagnosis prediction in one model. In particular, to obtain sparse representations of the samples in a low-dimensional latent space, an encoder with bottleneck layer is used to conserve essential and diagnostic features and eliminate redundancies, which are then fed to fully connected layers to make predictions. We use the Adam optimizer with dynamic measures such as early termination, and learning rate decay to guarantee the best convergence and generalization. In real-world application, state of the art performance is observed on a benchmark dataset containing 4920 samples, 132 symptom features and 40+ class diseases using about 80% for training data resulting in test data with overall accuracy of 99.6% along with almost perfect precision, recall and F1-scores across most classes. Further tests using confusion matrix, ROC curve, precision-recall curve and calibration plot show the strength, reliability and discriminative capability of the developed model. These findings indicate that the presented framework can be successfully used to make proper predictions of the disease, and, therefore, it can be utilized to a maximum in the real-life clinical decision support systems.