A deep learning approaches and fastai text classification to predict 25 medical diseases from medical speech utterances, transcription and intent

Abstract

The article examined the deep learning models and Fastai text classification technique to predict the medical speech utterances, transcriptions, and intent to extract the 25 medicals problems. The experimental work was conducted using a large amount of data which contains 6661.wav files and one.csv file, including 13 distinct categorization fields of medical speech utterances. Each illness's exploratory data analysis demonstrated the phrase length classes and disease categorization based on the recorded speech sound of patients for each disease. The preprocessing of the task included the wordcloud consisting of all the vocabulary words having different sizes based on the number of speech utterances in each category, eliminating Nan values, verifying for duplicates, and computing the corpus and their term index. Further, features are extracted to determine the number of words in each category, the length of phrases, and the number of words in each phrase, followed by lemmatization and tokenization. Deep learning models such as GRU (Gated Recurrent Unit), LSTM (Long Short Term Memory), bidirectional gated recurrent unit, bidirectional long short-term memory, and Fastai classifier have been used to exact category of disease from the medical speech utterances and their textual phrases. After the assessment, it was discovered that Fastai earned the most incredible precision, recall, accuracy, and lowest loss rate by 96.89%, 95.8%, 93.32%, and 0.169, respectively. In comparison, bidirectional LSTM had achieved the highest F1 score by 95.69% to predict the medical speech utterances for each category.