Machine learning in medicine: a practical introduction to natural language processing

Sentiment analysis is the process of assigning subjective meaning to words, phrases or other units of text [15]. Sentiment can be categorised simply as positive or negative, or can be related to more detailed themes, like the emotions that certain words reflect. Sentiment analysis serves a similar purpose to the process of ‘coding’ in qualitative research methods such as deductive thematic analysis [16]. A simple approach to sentiment analysis is to use a lexicon, which is a list of common words or phrases that have been matched to their categorical sentiment [17]. For example, a simple lexicon might match the words “love”, “favourite” and “respect” to a “positive” sentiment and the words “hate”, “pain”, and “anguish” to a “negative” sentiment. Lexicons serve as look-up tables that can automatically check the sentiment of each word or phrase in a passage of text. By quantifying the ratio of positive to negative sentiments in a sentence, for example, it is possible to start to understand the sentiment of the sentence overall. Lexicon-based sentiment analysis has been applied for detecting and monitoring disease outbreak based on the emotive sentiment of Twitter posts [18], assessing the public response to Obamacare [19], and to investigate patterns in social media posts about diabetes [20].

Sentiment analysis can be complicated by negation and sarcasm. For example, the word “not” reverses the sentiment of the word “recommend” in the sentence “I would not recommend this hospital to a friend or family member”. One potential way to handle this is by first splitting (tokenising) the sentence into bi-grams (pairs of adjacent words), rather than individual words [21]. This can help to identify words preceded by a negating particle and reverse their polarity, or sentiment can be assigned directly to the bi-gram [22]. In this case, the bi-gram “not recommend” might be assigned a negative sentiment. This approach to detecting negation has clear limitations in terms of sentence complexity, for example, negation in the sentence “the patient did not report a history of asthma” could not be handled by bi-grams. A more sophisticated and commonly used approach to handling negation is to employ algorithms that search for negation phrases. Examples include the NegEx algorithm [23] and its successor ConText [24], which can also qualify the temporality and experiencer of common medical conditions (i.e. whether a condition was present, when it was present, and in whom it was present). The sentiment of sarcastic remarks is often more dependent on context than the words themselves, and while attempts have been made to create sophisticated “sarcasm detectors”, this still poses a challenge to sentiment analysis [25].

Supervised and unsupervised ML algorithms can also be trained to assign sentiment to passages of text either independently, or with a lexicon as a hybrid approach. These approaches can account for complex interactions between words in a sentence more intricately than purely lexicon-based approaches. This paper demonstrates the simplest and least computationally intensive form sentiment analysis (the use of a publicly available lexicon only), but more advanced techniques have been described in detail elsewhere [26, 27].

Unsupervised ML algorithms aim to find previously undefined patterns within datasets, for example by grouping similar observations into clusters. They use data that have not been “labelled” by a human supervisor (i.e., observations which have not been categorised a priori) [14].

When applied to text analysis, unsupervised machine learning can be used to identify common themes within text by clustering words or sentiments that frequently appear together. This process is called “topic modelling” and is similar to an inductive thematic analysis [16, 28]. This approach has been applied to social media posts to understand common themes in a person’s reason for staying in, or leaving, an abusive relationship [29], discovering themes relating to drug non-compliance [30], and identifying health constructs that are important to patients, but not necessarily captured in patient-reported outcome measures [31].

Supervised machine learning algorithms aim to make predictions about an outcome (dependent variable) based on a set of features (independent variables). They use labelled datasets to learn how features and outcomes are related in order to predict outcomes from new, unlabelled, datasets [14]. They do this by iteratively adjusting their parameters to reduce prediction error in a training dataset, a process referred to as “tuning”. The performance of the algorithm is then usually tested with a second dataset, where it attempts to predict known outcomes from their associated features.

Broadly, supervised machine learning algorithms fall into two categories: regression algorithms and classification algorithms. Regression algorithms aim to predict a continuous outcome (e.g., blood pressure or risk of death) while classification algorithms (or classifiers) aim to predict categorical outcomes (e.g., positive or negative, benign or malignant). It is worth noting that in ML, the term “regression” is used slightly differently to traditional statistics, where regression can also be used to predict categorical outcomes. For example, a logistic regression model that aims to predict a binary outcome would be described as a classification algorithm rather than a regression algorithm, within ML literature [14].

Supervised ML algorithms have been combined with NLP to extract patient-centred outcomes from unstructured medical records [32], to detect emergent psychosis from language used by vulnerable youths [33], and to predict Care Quality Commission inspection results, based on hospitals’ social media comments [34].

This paper contains practical examples of common text analysis techniques, which we perform on a freely available dataset that contains over 200,000 patient drug reviews (in the form of text and an associated numerical rating, out of 10). The dataset is similar to what might be obtained by scraping online patient forums or healthcare provider review websites.

We provide code that can be modified and applied to similar analyses in other datasets. We aim to demonstrate to clinicians and qualitative researchers the type of text analyses that are easily performed with open source software, and provide practical understanding to academics that wish to apply these techniques in their own research.

The methods section of this paper is structured into four parts, which in turn cover:

We present samples of code written using the R Statistical Programming Language within the paper to illustrate the methods described, and provide the full script as a supplementary file. At points in the analysis, we deliberately simplify and shorten the dataset so that these analyses can be reproduced in reasonable time on a personal desktop or laptop, although this would clearly be suboptimal for original research studies.

While this paper is intended for readers who are relatively new to the field, some basic familiarity with the R programming language and machine learning concepts will make this manuscript easier to follow. Before attempting to recreate these experiments, readers may wish to read our introductory paper to machine learning in general, which covers R programming, supervised machine learning and model interpretation in more detail than is described in this paper [14].

In recent years, the R and Python programming languages have become extremely popular for machine learning tasks [35]. They are both open-source, with thousands of free pre-programmed packages that can be used for statistical computing, and large online communities that provide support to novice users. R and Python have similar capabilities and are becoming increasingly interoperable, with many important machine learning packages now available for use in both languages.

We have performed these analyses using R version 4.0.3 in the RStudio computing environment. The experiments have been designed to run relatively quickly on a personal device. We used a 2017 MacBook Air, running macOS Mojave Version 10.14.6 on a 1.8 GHz processor with 8 GB of random-access memory.

In these experiments, we used the Drug Review Dataset from the University of California, Irvine Machine Learning Repository [36]. The dataset was obtained by scraping pharmaceutical review websites and contains drug names, free text patient reviews of the drugs, and a patient rating from 1 to 10 stars, among other variables. The dataset comes pre-split into training and test sets. We have randomly selected 5000 records from the training dataset to start with, in order to reduce computational demand.

In English text, many different combinations of characters can be used to mean the same thing. For example, “won’t”, “will not”, “Will not”, and “will not!” all use a different set of characters to convey very similar meanings. The main goal of data cleaning in NLP is to standardise text so that these variations are interpreted as the same feature by the machine learning models downstream.

There are a number of NLP techniques for standardising the free text comments [37]. We expanded contractions (e.g., replaced words “don’t” with “do not” and “won’t” with “will not”), removed non-alphanumeric characters, and converted all characters to lower case.

We also performed stemming with an English language stemmer. Stemming is the use of algorithms to reduce similar words to a common stem, for example by removing suffixes [38]. In our data cleaning pipeline, we have used the simple and freely available Porter algorithm for stemming, which largely works by removing inflexional suffixes. For example, the Porter algorithm would convert the words “learning”, “learned”, and “learns” to their common stem “learn” [39].

Next, we removed English “stop words” (common and usually unimportant words such as “the”, “and” and “is”) [40], and words with 3 or fewer characters. This dramatically reduces the number of features in the dataset, and allows algorithms to focus on the most meaningful elements of text. This stage of data cleaning is based on a principle known as Zipf’s Law, which states that the occurrence of a word within a body of text is inversely proportional to its rank in a frequency table. This means that the most commonly occurring word (often “the” in English language) occurs approximately twice as frequently as the second most common word, three times as frequently as the third most common word, and so on [41]. In keeping with Zipf’s law, 135 repeated words make up half of the one million words in the Brown University Standard Corpus of Present-Day American English [42]. For the linguistic analyses described in this paper, it is generally accepted that the most commonly used words are the least informative.

Next, we assigned a sentiment to each word in the dataset using a freely available lexicon known as “Bing”, first described for use in consumer marketing research by Minqing Hu and Bing Liu in 2004. The Bing lexicon ascribes either a “positive” or “negative” sentiment to 6786 different English words [17].

We chose to evaluate the overall sentiment of reviews for four different drugs, which we chose based on their diverse indications and pharmacological properties: Levothyroxine, Viagra, Oseltamivir and Apixaban. Levothyroxine and Viagra are used to treat hypothyroidism and erectile dysfunction respectively. Oseltamivir is used to treat and prevent influenza infections, and Apixaban is used to treat and prevent blood clots. Many people who take Levothyroxine and Viagra experience relief of troublesome symptoms, while people taking Oseltamivir or Apixaban prophylactically may not experience any appreciable benefit other than a lower risk of developing influenza or blood clots. We tested the hypothesis that Levothyroxine and Viagra are reviewed more positively than Oseltamivir and Apixaban.

To do this we tabulated the positive and negative sentiments assigned to all reviews of each drug, and calculated the percentage of sentiments that were positive.

Text data can be structured into a semantic hierarchy. A “token” is a series of characters that form the smallest semantic unit. In the experiments described in this paper, each word stem is its own token. If we had tokenised the drug reviews into bi-grams (to handle negation, for example), then each token would be two adjacent words. A “term” is a class of token with the same characters [38]. A “document” is a collection of tokens that appear together to convey a collective meaning, and a “corpus” is a collection of documents [37]. For example, within the corpus “Romeo and Juliet” the document “good night good night parting is such sweet sorrow that I shall say good night till it be morrow” might contain 19 tokens (words), with the term “night” appearing 3 times.

Before we can apply statistical or machine learning models to our text, we must first convert it into numeric data in a meaningful format. This can be achieved by creating a data table known as a document term matrix (DTM), sometime also referred to as a term document matrix (TDM) [14]. In the DTM, each row represents a document, and there is a column for each term used within the whole corpus. In a TDM, the orientation of rows and columns is switched. The cells contain numerals representing the number of times each term was used within a document. It is common for most cells in a DTM to contain the value “0”, as there are often many terms in a corpus, but these are not all used in each document.

For our illustration of unsupervised machine learning, we chose to explore patterns between drugs based on the words used to describe them. To do this, we constructed a DTM where all the reviews of a given drug were combined into one document, representing the description of that drug by multiple reviewers. We then removed terms that were absent from ≥ 99.5% of documents (Fig. 1). The resulting DTM (Fig. 2) had 1111 different rows (i.e., 1111 different drugs, each representing a document) and 1948 columns (terms used within the corpus).

The values in our DTM represent term frequency, but it is also possible to weight these values by scaling them to account for the importance of a term within a document. A common way to do this, that readers should be familiar with, is the term frequency – inverse document frequency (TF-IDF) index. The term frequency is the number of times a term appears in a document. The inverse document frequency is the natural logarithm of the total number of documents, divided by the number of documents with a given term in it. The inverse document frequency gives an impression of the “importance” of a term within a corpus, by penalising common terms that are used in lots of documents. The TF-IDF is the product of the term frequency and the inverse document frequency, and provides insight into document-level semantics by measuring the frequency of a term within a document, and weighing this against the “importance” of the term within the corpus [38].

To cluster similar documents (drugs) together, we used an unsupervised machine learning technique called latent Dirichlet allocation (LDA). The LDA algorithm clusters terms into a predefined number of “topics” based on the probability of those terms being used together within a document. It then predicts which topic a document will belong to, based on the terms in that document.

Unlike other forms of clustering, such as k-means, it is possible for a term to belong to more than one topic in an LDA analysis [28]. This generally makes LDA a more appropriate tool for topic modelling, as in most cases it will be possible for a document to meaningfully belong to more than one topic. For example, the drug citalopram could belong to both a topic representing drugs that treat depression and to a topic representing drugs that cause nausea. We chose LDA as an easy-to-implement technique for basic topic modelling, although other methods have been described, including dynamic topic modelling (which maps topic structure within a corpus over time) [43], and structural topic modelling (which uses meta-data, such as author name and date to enhance cluster formation) [44].

We programmed the LDA to cluster terms into 3 topics (Fig. 3). The number of topics was chosen for illustrative purposes. Choosing the number of clusters for an LDA-based topic model can be challenging. Where a number of clusters are expected based on an understanding of the corpus content, this number can be chosen (similarly to a deductive thematic analysis). Where the analysis is exploratory, the process can be repeated iteratively, and different models assessed for real-world plausibility. There are also statistical approaches to determining topic number, for example the rate of perplexity change, which relates to how well the model fits hold-out data [45].

To explore themes within the terms, we then identified the 10 terms most likely to belong to each topic. Next, we identified the 10 documents (drugs) most likely to belong to each topic. We hypothesised that similar drugs would be described in similar ways, and therefore cluster together.

Next, we rearranged the dataset into a DTM where each review was an individual document. Sparse terms were removed, resulting in 808 remaining features (terms), which were weighted by TF-IDF. We randomly selected a sampled 1000 reviews to further reduce computational burden. Clearly, researchers aiming to generate robust models should use as much data as possible, although this can add to the computing time and hardware requirements.

For each review, we took the associated star rating and dichotomised it into a binary outcome. A star rating of 1–5 was categorised as a “Bad” review and a star rating of 6–10 was categorised as a “Good” review. We then aimed to train 3 different supervised machine learning algorithms to predict whether a review was “Good” or “Bad”, depending on the words that were used in the review.

We split the data into training and test sets to create and evaluate our models respectively. We randomly assigned 75% of the reviews to the training set and 25% to the test set (Fig. 4).

This split resulted in a training dataset with 524 “Good” reviews and 226 “Bad” reviews. Training data with unbalanced classes can cause classifiers to predict the more frequently occurring class by default, particularly when sample sizes are small and features are numerous [46]. This can result in misleading accuracy statistics, for example if a model has a high sensitivity but poor specificity and is tested in a sample that has many more positive than negative observations.

To address the issue of class imbalance, we used the synthetic minority oversampling technique (SMOTE) [47]. The SMOTE algorithm creates new, simulated datapoints to balance the number of observations in each class. New data are simulated based on clusters that exist within the training data, using another form of machine learning algorithm known as K-nearest neighbours. Other methods to handle class imbalance are discussed elsewhere [48]. Briefly, these strategies involve oversampling the minority class, undersampling the majority class, or increasing the penalty for a majority class misspecification relative to a minority class misspecification.

The SMOTE algorithm transformed our training data into a dataset with 678 “Good” reviews and 678 “Bad” reviews. We then used this dataset to train 3 different types of supervised machine learning algorithm: a regularised logistic regression, a support vector machine (SVM), and an artificial neural network (ANN). These three types of classifier represent a spectrum of ML algorithms, ranging from relatively simple, easily interpretable and with a low number of parameters (regularised regression), to complex and difficult-to-interpret algorithms with a large number of parameters (ANN).

Regularised regression is similar to traditional regression, but applies an additional penalty term to each regression coefficient to minimise the impact of any individual feature on the overall model. Depending on the type of regularisation, and size of the penalty term, some coefficients can be shrunk to 0, effectively removing them from the model altogether. The purpose of regularisation is to prevent overfitting in datasets with many features [14].

Support vector machines aim to model a linear decision boundary (or “hyperplane”) that separates outcome classes in high-dimensional feature space. Model parameters can vary the way in which data are transformed into high-dimensional space, and how the decision boundary is drawn [14].

Artificial neural networks are so-called because they share a conceptual topography with the human central nervous system. They consist of interconnected neurons arranged in layers. Each neuron sums its inputs, multiplies this by a weight, and transforms the signal through an activation function. The weight of each neuron and their collective arrangement will affect model performance [14].

We would recommend that readers consult our previous instructional paper for a more thorough description of regularised regression, SVMs and ANNs [14]. For the purposes of this experiment, it is sufficient to understand that each model has a number of parameters which can be iteratively adjusted to improve that model’s predictive performance in samples of the training dataset.

In this study, model parameters were iteratively adjusted and tested across 10 bootstrap samples of the training dataset. When recreating this experiment, the number of bootstrap samples can be increased to improve model performance (reduce overfitting), but this will add to the computational demand. The parameters that maximised classification accuracy were chosen for the final models, which were then evaluated in the test dataset. Model performance was assessed with classification accuracy, area under the receiver operating characteristic curve (AUC) and confusion matrices.

The results of our sentiment analysis are displayed in Table 1. Levothyroxine and Viagra had a higher percentage of positive sentiments than Apixaban and Oseltamivir. The number of sentiments in the analysed dataset was low, and sentiments for each drug were negative overall.

The ten terms most likely to appear in each topic are presented in Table 2, along with their beta values, which represent the term density within that topic. The five most common terms in Topic 1 (effect, feel, start, week, month) might suggest that this topic represents drugs that have taken weeks or months for the patient to feel its effects. Topic 1 also contains the terms “anxieti” and “depress” (note the Porter algorithm has reduced the term “anxieties” to the stem “anxieti”). This is interesting because many drugs used to treat anxiety and depression are associated with a gradual onset of action [49]. Topic 2 contains several terms that might relate to the female reproductive system. Topic 3 was less clearly defined.

The ten documents (drugs) most likely to belong to each topic are presented in Table 3, along with their gamma values (the proportion of terms used in a document that belong to the given topic). In Topic 1, 9 of the 10 most common drugs are primarily used to treat mental health problems. The tenth, Levetiracetam, is an anti-epileptic. Interestingly, Levetiracetam’s use is associated with anxiety and suicidal ideation [50]. In Topic 2, all 10 drugs are hormone-based treatments primarily used as contraceptives. The drugs listed in Topic 3 did not have a clearly defined relation.

The regularised regression and ANN took approximately 10 min each to train. The SVM took approximately 40 min to train. Model performance statistics are presented in Table 4. Classification accuracy ranged from 0.664, 95% CI [0.608, 0.716] for the regularised regression to 0.720, 95% CI [0.664, 0.776] for the SVM.