Deep bidirectional LSTM for disease classification supporting hospital admission based on pre-diagnosis: a case study in Vietnam

Abstract

Overcrowding in hospitals in Vietnam has caused many disadvantages in receiving and treating patients. Especially at the stage of receiving and diagnosing procedures taking patients to the treatment departments in the hospital takes up much time. This study proposes a text-based disease diagnosis using text processing techniques (such as Bag of Words, Term Frequency- Inverse Document Frequency, and Tokenizer) combined with classifiers (such as Random Forests (RF), Multi-Layer Perceptron (MLP), Embeddings and Bidirectional Long Short-term memory (LSTM)) on symptoms. As observed from the results, deep Bidirectional LSTM can reach 0.982 in AUC in the classification of 10 diseases on 230,457 samples of pre-diagnosis collected from Vietnam hospitals used in the training and testing phases. The proposed approach is expected to provide a way to automate patient flow in hospitals to improve healthcare in the future.

PhoBERT: Application in Disease Classification based on Vietnamese Symptom Analysis

“…Compared to the bag of words approach in [24], with the same dataset, Logistic Regression achieved an accuracy of 79.1 % to classify ten common diseases in Vietnam. In addition, another study in [25] obtained 87.30 % in accuracy with the dataset filtered the noisy samples. In our study, the PhoBERT model achieved an accuracy of 87.8 % (as exhibited in Table VI).…”

“…Finally, the authors used the training set fetched into the learning algorithms to build an experimental set evaluation model with 79.1 %, 79.9 %, 79.1 %, 79.5 % and 92.1 % in Accuracy, Precision, Recall, F1-score, and Gini, respectively. The work in [25] performed integration between various text processing techniques and some classification algorithms on the dataset. They obtained the best performance with Bidirectional LSTM.…”

“…As stated in [29], a bidirectional layer can be leveraged in text classification tasks aiming to capture the past and future context of each word or token in a sequence. The number of trainable parameters (448 970 parameters (params)) of the proposed LSTM is lesser than 7 % of the work in [25] (482 138 params).…”