Automatic emotion recognition in healthcare data using supervised machine learning

Data collection

To build the required dataset, we collect English text from different online sources including Health News sites e.g., Medisys (Rortais et al., 2010) and Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT) (Donnelly, 2006), Discussion forums & Communities. Further, we also collect some text samples from Google News (http://news.google.com/) related to the relevant disease by manually searching with disease keywords. Further details about dataset collection can be seen in our previous paper (Azam, Tahir & Mehmood, 2020a). We crawled the dataset from different online sources and save the text into an excel file. There are currently a total of 4,202 text samples in our dataset related to different prominent diseases and we are also working to expand it further.

Disease classes

For this purpose, first of all, we target the dominant diseases of Pakistan. According to the Seasonal Awareness and Alert Letter (SAAL) (National Institute of Health, Government of Pakistan, 2020) of October 2019–April 2020 issued by the Ministry of National Health Services, Regulations & Coordination, Government of Pakistan, National Institute of Health (NIH) Islamabad Pakistan, Measles, Dengue, Typhoid, Malaria, Acute Hepatitis, HIV, Ebola, Cancer, and some others are the dominant diseases of Pakistan. Moreover, we also include the disease which is trending in the world 2020, i.e., Coronavirus or COVID-19. Therefore, we prepare our own emotionally labeled dataset allocated on the chosen diseases. There are a total of eight disease classes in the EmoHD dataset which are HIV/AIDS, Dengue, Hepatitis, Malaria, Influenza, Coronavirus, Cancer, and others respectively. Other class means if text sample is not related to the mentioned seven classes then we gave that text to another class label. We can say in other class, text can be related to any other disease but not from these seven mentioned disease classes. In this way, at the point of data collection, we already know the disease class of that text sample as we are fetching the picked out disease type data from online resources.

Emotion class labelling

The Emotion label is given to the text sample by using Parallel Dots API (Komprehend, 2020). It is an Artificial Intelligence (AI) platform created using advanced Deep Learning (DL) and Machine Learning (ML) techniques. This API labels the text into 6 classes as follows: (1) Happy (2) Sad (3) Angry (4) Excited (5) Bored (6) Fear. To use this API, we need an API_Key and it returns a JSON response having a confidence score for each of the emotion labels (Happy, Sad, Angry, Excited, Bored, or Fear). After getting the probability of every six emotions from the Parallel Dots API, we then assign the final emotion label to every text on the basis of the higher probability of emotion class. So now in our excel file, we have text along with disease label and final label of emotion against each text. Also, we have the probability of every emotion class against each text. In this way, we labeled the dataset containing 4202 samples.

Dataset splits

We randomly split our dataset into the training and testing with the ratio of 80% and 20%. After splitting our dataset into the training and testing part, there are a total of 3362 samples in training dataset and 840 total samples in the test dataset.

Text specialities of EmoHD dataset

We did an analysis on our collected dataset and it can be seen in the Table 1 that there are a total of 1634319 number of words in our dataset and if we talk about unique number of words then it is around one lac that is 91988. Similarly. the total number of characters, number of numerics, unique Bigrams, and Trigrams can also be seen in Table 1.

Emotion class wise distribution of EmoHD dataset

Table 2 shows the emotion class-wise distribution for each text sample. It is evident in Table 2 that the Angry emotion class has a high number of text samples that are 1,343 and the Bored emotion class has the least number of text samples that are only 22. Sad emotion class has also a significantly less number of samples than other emotion classes. Since, we are talking about emotions of patients during a specific disease then it is obvious that if someone has a disease then there will be less chance that the person will feel bore during that disease. It is natural that is why our dataset has less number of bored emotion class samples. Moreover, It can be seen from the given Table 2 that the class imbalance problem is present in our dataset. This is because of text we got against each disease and we can overcome on class imbalance problem by applying oversampling or under-sampling techniques.

Disease class wise distribution of EmoHD dataset

Statistics of distribution of disease class of EmoHD dataset can be seen in the Fig. 1. It can be seen in given Fig. 1 that the Cancer class has a high number of samples that are 660. It is evident from the Fig. 1 that the class imbalance problem does not explicitly exist in disease class wise distribution of the EmoHD dataset but we have seen in Table 2 that when we see emotion class wise distribution of these samples then class imbalance balance problem is present there as we have discussed the reason in the previous section.

Lowercase conversion

The first step of pre-processing is the lower case conversion. For this purpose, we clean the data by transforming the text into the lower case to avoid having multiple copies of the same words. For instance, ‘Happy’ and ‘happy’ must be considered as a single word but due to case differences they are counted as different words. Therefore, we apply the lower case conversion technique to avoid such ambiguities.

Punctuation removal

The next step is to remove punctuation from the text, as it does not add any useful information while training the model. Removing all punctuation occurrences from text data will help the model to learn functional features instead of non-functional instances.

Removal of stop words

Stop words should be removed from the text data as they do not add much meaning to the machine learning model. For this purpose, we use a predefined library of python, i.e., nltk.corpus (Pythonspot, 2016) for English stop words.

Common words removal

Common words are removed from the textual data because of the high similarity in the vocabulary. The overlap in vocabulary affects the evaluation scores during the training of machine learning classifiers. These words are removed as their presence will not be of any use in the classification of our text data.

Rare words removal

Similarly, just as we removed the most common words, we also remove rarely occurring words from the text. They are so rare, hence, the association between them and other words is dominated by noise.

ReSampling

It can be seen in the Table 2 that the emotion class distribution is not evenly distributed and it clearly shows the unbalanced data. To handle unbalanced data, we need to perform data re-sampling. Re-sampling technique (Tian et al., 2019) can either be Over-Sampling or Under-Sampling. We try both of these approaches and resample the data based on the majority and minority class respectively.

Moreover, after balancing the data, the analysis is performed based on two perspectives, i.e., including minority class as well as excluding minority class.

Count vectorization

In Natural Language Processing (NLP) toolbox, Count Vectorization has an important role. The number of times a word or term appears in the documents are referred to as count vectors and that is why Count vectorizer is also termed as Term Frequency vectorizer (Sekhar et al., 2020). In count vectorizer, we do not simply catch the presence of words for guaranteed report yet additionally catch how frequently it happens. Consequently, the text data is presented as a vector of words alongside the occasions of the terms that happen in that text sample.

Term frequency-inverse document frequency

Realizing exactly when a word appears in text data and when they show up consistently or inconsistently across the dataset isn’t extremely useful. Regardless of whether a word shows up habitually in certain documents yet less oftentimes in others can be truly valuable (Bounabi, El Moutaouakil & Satori, 2019). Therefore, TF-IDF comes in there and plays its vital role. TF-IDF score is composed of two terms as mentioned below: ${\displaystyle TF\left(w\right)=\frac{\text{Number of times word (w) appears in a document}}{\text{Total number of words in the document}}}$ ${\displaystyle IDF\left(w\right)=\text{log}\left(\frac{\text{Total number of documents}}{\text{Number of documents with word (w)}}\right)}$

Further, TF-IDF vectors can be generated at different levels of input tokens, i.e., Words, Characters, and N-grams (Bansal, 2019). The Word Level TF-IDF is a matrix that represents TF-IDF scores of every word in different documents or archives. Similarly, N-gram Level TF-IDF is actually the combination of N terms together and the resultant matrix describes the tf-idf scores of N-grams. On the other hand, the CharLevel TF-IDF score is based on the characters present in the corpus.

Multinomial naive bayes (MNB)

MNB is specially designed for text data and it is a particular version of Naive Bayse (Balakrishnan & Kaur, 2019). Multinomial Naive Bayes classifier utilizes multinomial dissemination for every one of the features. MNB just tells us that each probability of feature given class is a multinomial dispersion instead of some other dissemination. This functions admirably for information that can undoubtedly be transformed into counts, for example, the word includes in content.

MultiLayer perceptron (MLP)

MLP is a feed-forward neural network (FFNN). This classifier uses the backpropagation method during the learning of data (Labani et al., 2018). A MLP comprises of at least 3 layers: an input layer, a hidden layer, and a yield layer. Aside from the input layer, every node in every layer is a neuron that utilizes a nonlinear activation function. The basic architecture diagram of MLP is shown in Fig. 3.

Logistic regression (LR)

Logistic regression looks like a regression algorithm by name but actually, it is a classification algorithm. Logistic regression uses the sigmoid function. The sigmoid function can be defined as given in Eq. (1). (1)${\displaystyle g\left(x\right)=\frac{1}{1+e^{{-}^x}}}$

while x in Eq. (1) can be any number or it can be any formula. It is a linear classifier that calculates the probabilities using logistic function to find the relation between categorical dependent and independent variables (Genkin, Lewis & Madigan, 2007; Shah et al., 2020).

Support vector machine (SVM)

SVM is a supervised ML algorithm (Tong & Koller, 2001). The model is very simple and many people preferred to use this model because of less computational power and it gives significant accuracy. SVM not only performs linear classification but also performs non-linear classification very well with the help of kernel trick. The basic idea of this model is very simple, the algorithm draws a hyper-plane or line that divides the data into categories or classes. Hinge loss is used to draw a line that helps in maximizing the margin between line and data points. For an expected yield t = ±1 and a classifier score z, hinge loss can be defined as given in Eq. (2). (2)${\displaystyle Los{{s{}_H{}_i{}_n}_g}_e=max\left(0,1-t\ast z\right)}$ It can be used for both regression problems and classification problems but mostly it is used for classification problems. This model is very helpful for text categorization. That’s why we used this model to predict emotion class from given text samples of the EmoHD dataset. It performs very well when data has high dimensions but its efficiency compromises with noisy data.

Random forest (RF)

Random Forest is an administered learning calculation that is utilized for both classification and regression (Onan, Korukoğlu & Bulut, 2016). Nonetheless, it is predominantly used for classification. As we all know, timberland is made up of trees, and more trees imply a more vigorous woodland. Essentially, arbitrary woodland calculation makes choice trees on information tests and afterward gets the expectation from every one of them lastly chooses the best arrangement by methods for casting a ballot. It is an outfit technique that is superior to a solitary choice tree since it diminishes the over-fitting. Random forest calculation runs productively in enormous information bases and creates exceptionally precise forecasts by assessing missing information.

eXtreme gradient boosting (XGB)

XGB uses a gradient boosting framework and basically, it is an improved version of gradient boosting algorithm (Ria et al., 2020). It is a more regularized model that uses L1 and L2 regularization which helps to improve model generalization ability, thus overcoming the overfitting problem. This model should not be used for image recognition or if there are few training samples than number of features.