Drug Recommendation System for Healthcare Professionals’ Decision-Making Using Opinion Mining and Machine Learning

This study aims to develop a web-based drug recommendation system for healthcare professionals using opinion mining and machine learning. The proposed system architecture is depicted in Fig. 1. A three-layer architecture design is implemented for this system. Two main users of this system are healthcare professionals and admin. For frontend of the web application, React JS and Material UI (MUI) library were utilized. React JS, a widely used JavaScript library, was utilized alongside the Material UI Library to develop the user interface. React’s component-based architecture, virtual DOM, and responsive design make it ideal for frontend development. The Material UI Library offers prebuilt components adhering to material design guidelines, simplifying UI development by eliminating the need to create components from scratch. Integration between React and Material UI is seamless, allowing for customized components to meet specific UI requirements. Typescript, a superset of JavaScript, was employed as the programming language for frontend development.

To fulfil the web application’s API functionalities and incorporate machine learning models, FastAPI was selected. It serves as the backend implementation for all the system’s functionality. Furthermore, the database used for this project is Mongo DB which is a NoSQL database. Mongo DB stores data in BSON format, which is compatible with the structure of the drugs, users, and forums data in the system. FastAPI acts as an intermediary between the frontend, machine learning models, and the database. When users request access to data, they make API requests. Similarly, for utilizing machine learning models, API requests are made to FastAPI, passing the necessary parameters. The backend then retrieves the model results and returns them to the frontend for display to the user.

There are four main modules in the proposed system consisting of: (1) user management module, (2) drug review analytics module, (3) recommendation module, and (4) drugs management module. The target customers for this proposed system are healthcare professionals in hospitals. They can use this system to find the best drugs for a certain diagnosis and recommend them to their patients. The admins manage the overall system modules. All healthcare professionals and admins who use this system must be registered. The user management module is important in collecting user behavior data for the drug recommendation module.

For the drug review analytics module, patient drug reviews are classified as positive, neutral, or negative based on sentence polarity. Sentiment analysis is used to identify key features from reviews and then use them to assess the polarity of the review. Feature extraction and feature orientation identification are used in the sentiment analysis process. Depending on the orientation of the feature, the polarity of a review is classified as positive or negative. The features from the reviews are extracted using the Term Frequency–Inverse Document Frequency (TF-IDF) and Bio Bert. Several algorithms are employed to determine the reviews’ sentiment, including perceptron, logistic regression, and long short-term memory networks, along with the Bio Bert and TF-IDF feature extraction algorithms. For this solution, the No Free Lunch (NFL) Theorem is used, in which all of these models are evaluated using the appropriate metrics and the best performing model is selected based on the metrics score. Following the sentiment analysis, each medicine receives an effectiveness score. This effectiveness score is used to rank drugs in the system when healthcare professionals search for drugs. Drugs with higher effectiveness scores will be given more important rate. Sentiment analysis may not be relevant for patients because they are not making decisions about the type of that they need to take. But it is crucial for healthcare professionals to gain a better understanding of drugs and their effectiveness in order to select the best drug to prescribe to a patient for a given diagnosis.

The next module is the drug recommendation module. Cold start issues and overspecialization issues might arise in a recommender system [12]. To address these issues in the proposed system, a hybrid content-based and collaborative filtering recommender system is implemented. The cold start problem is solved by using content-based filtering techniques to recommend drugs to healthcare professionals based on how similar the drugs are to one another. This approach of filtering is primarily concentrated on the characteristics of each drug utilizing an item profile characterizing and looks for similarities to previous drugs that healthcare professionals liked. Healthcare professionals receive recommendations regardless of their user profiles. Term frequency-inverse document frequency (TD-IDF) algorithm is used to weigh the feature from the dataset first, and then it generates step-by-step cosine similarity tables. Drug side effects are one of the features that are considered for content-based filtering. As a result, the lack of drug reviews and ratings is no longer an issue. Furthermore, collaborative filtering is used to tackle the overspecialization problem by enhancing healthcare professionals’ preferences. Drug recommendations are made by looking at the preferences of a group of users who are similar to them. By grouping healthcare professionals who access or buy similar products in the system, the system converts the behaviors of healthcare professionals into implicit rating weightage. After that, a collaborative rating is used to determine the user’s rating. The ratings of healthcare professionals are then matched to those of the target user in order to identify those who share the same preferences. This approach might not be very effective at first, but after many healthcare professionals start using the system, it becomes a very reliable recommendation method.

The final module is the drug management module. Users can search for any drugs in the system related to a particular disease. Once a user views a particular drug, the drug’s details are displayed to them. Furthermore, users have additional features like drug comparison, adding drugs to wish list and forum page. The drug comparison feature will allow healthcare professionals to make comparisons between two drugs and assist them to choose the best one. The forum page on the other hand facilitates communication between healthcare professionals. This feature allows healthcare professionals to share knowledge about the drugs and prescriptions and expand their network.

Agile methodology [14, 15] is used as the development methodology in this study since it provides a stable system delivery in a short development time. Figure 2 depicts the agile development methodology diagram used in this study. Agile development involves six main phases of planning, design, development, testing, maintenance, and deployment. In agile, the system development tasks are breakdown into smaller parts and each part is handled in a number of iterations. Each iteration involves all six phases of agile development, and several numbers of iterations are required to complete an entire system. If there are any errors, the debugging process is done during development. For every iteration, a set of system requirements is listed out and followed accordingly. Therefore, requirements changes can be made easily before a new iteration starts. This ensures that agile development methodology adapts to new changes immediately and these changes can be integrated into the system easily without having to make a lot of changes in the system. Release of the system cannot be done after a few iterations as there might not be enough functionalities. After all the modules are developed completely, they are integrated into one module. Then this module is tested to make sure it is ready for final release. Agile methodology implementation helps us to minimize development process risks and focuses on getting products to market fast.

Figure 3 illustrates the overall system flowchart for this study. The main users, who are healthcare professionals, must log in before using the system. The users need to create a new account first. After logging in, users can search for drugs by name or condition. All relevant drugs related to the search will be displayed in the search results. These drugs will be ranked according to the effectiveness score. Drugs with higher effectiveness scores will be ranked first. The user can then click on any specific drug in the search results that they want to view. When a user clicks on a specific drug, details about the drug, sentiment analysis results, and recommended drugs will be available to the user. After performing sentiment analysis on drug reviews, the system will display the results including general statistics for the user’s preferences.

Moreover, based on the content-based filtering model, the system will identify drugs that are similar to the drug being viewed or displayed to users. In the search page, another recommendation will be available for users, and they can navigate to the drug comparison page. In this page, the users can select any two drugs available in the system and view their comparison results. Furthermore, users can navigate to the wish list page. There is also a forum page where users can create, edit, and delete posts, view forums created by other users and comment on them. The main purpose of this feature is to facilitate communication between healthcare professionals. The final page assist users to navigate to add drugs data page for adding more drug data to the system or delete the existing drug data. This page and feature are only for the usage of admins. Eventually, users can either search for more drugs or leave the system.

This study used sentiment analysis to classify drug reviews as positive, neutral, or negative based on their sentiment. A machine learning-based approach is utilized to implement sentiment analysis on drug reviews, as shown in Fig. 4. This is due to the fact that machine learning-based approaches generally outperform lexicon-based approaches and have higher accuracy scores [16‐20]. The dataset being utilized for sentiment analysis is labeled, making it appropriate for this method. Positive reviews are labeled as 1, neutral reviews as 0, and negative reviews as −1. This dataset is divided into 80% training data and 20% testing data. Data pre-processing includes two steps which are data cleaning and feature extraction. Data cleaning involves removing duplicate data and unnecessary contents like repetitive words, symbols and stop words. Then, text tokenization is performed on the cleaned data where reviews are broken down to a set of words and each word undergoes lemmatization.

Following the data cleaning step, TF-IDF and Bio Bert algorithms are used for feature extraction. Feature extraction is performed to select the most important features from reviews. These features are then used to predict the polarity of drug reviews. The classification process is done in the next step. For classification, supervised machine learning models like Perceptron, Logistic Regression, and Long Short-Term Memory Network are utilized. The models are then evaluated using precision, f1-score and accuracy metrics. Finally, these three algorithms are compared, and the best algorithm is selected to implement sentiment analysis on drug reviews.

Recommendation systems can be built using different types of algorithms like associative rule, content-based filtering, collaborative filtering, and knowledge-based filtering. This study utilized a hybrid content-based and collaborative algorithm [12] to build a drug recommendation system. Figure 5 depicts the proposed hybrid filtering algorithm framework for this study, which is adopted from [12]. This study aims to resolve some of the recommendation system issues such as cold start problem, increasing customer preference, large number of drugs in market and overspecialization problems [12]. Identifying problems in a system is important for the development process. Then, data collection and pre-processing are done. The dataset used in this study are drug reviews data, drug information data and healthcare professionals’ data collected from AskaPatient database, UCI ML Drug Review dataset, Drugs.com, and DrugBank respectively. After data pre-processing, the proposed recommendation model, which combines content-based filtering and collaborative filtering is performed.

For content-based filtering, drugs are recommended to healthcare professionals based on drug similarities without taking user profile into count. The TF-IDF algorithm is used to extract features from data and give weightage to them. Depending on the frequency of each word, this algorithm determines its significance. This will be then used to create a step-by-step cosine similarity table. There are a few steps involved in creating a cosine similarity table. First, TF scores are computed, and the table is normalized. Then, IDF is calculated to find the number of items for each user. Finally, the importance of items is ranked for each user by multiplying TF and IDF scores. Recommendation can be done by picking the top N similar products where N stands for the number of recommendations for a user. Content-based algorithms are chosen because they help to solve cold start problems. Therefore, recommendation of drugs can be done without obtaining any user data.

For collaborative filtering, drugs are recommended to healthcare professionals based on user behavior. First, user behavior like viewing a drug or adding a drug to a wish list is converted into implicit rating weightage. This weightage is then used to generate user rating using collaborative filtering. When user behavior is converted to rating, sparse matrices are formed because there is a lot of missing data. In order to predict the missing values, matrix factorization is utilized. Matrix factorization helps to recommend least popular products to users using the new calculated rating. Collaborative filtering was selected because it helps to solve overspecialization and increasing customer preference problems. Therefore, healthcare professionals are able to gain more knowledge on different types of drugs for different diseases in the market and at the same time least popular products is recommended to them. After the proposed hybrid model is developed, it is evaluated using error metrics like mean absolute error, root-mean-square error, and ranking metrics like precision and recall. Finally, the overall recommendation system is developed and tested before integrating it into the web application.

In this study, two sets of data are required, which are drug reviews data for developing sentiment analysis models and drug details data for the system. Drug review data are obtained from mainly two sources which are AskaPatient database and UCI ML Drug Review dataset. The dataset for UCL ML Drug Review has a total of 232000 reviews for different drugs for different conditions. This dataset has attributes like drug name, health condition, drug review, date of review, drug rating and useful count. The dataset from AskaPatient contains drug review as well, but the attributes are rating, reason, side effects, comments, sex, age, dosage, and date. This dataset is very huge as it contains thousands of drug reviews for many types of drugs in the market. Both datasets do not have positive or negative labels for the reviews. Therefore, the rating of drugs is used to label the drug review data either positive, neutral, or negative for training the model. Positive reviews are labelled as 1, neutral review as 0, and negative review as −1.

Furthermore, another important dataset for this study is the drug details data from Drugs.com and Drug Bank Online. Both datasets include all the necessary information about drugs, which is extracted and shown to healthcare professionals for their reference. Data from both of these sources are combined together into a single drug dataset to be used for Dr. Drugs website. Examples of the data that can be retrieved from these datasets are generic names, dosing information, drug characteristics, associated conditions, and adverse effects.

The sentiment analysis model is developed to classify drug reviews of patients into positive, neutral, and negative classes. There are several steps involved in developing this model. The first step is data retrieval from the selected databases. For the sentiment analysis model, only review and rating attributes were used. The preprocessing tasks that were performed on the reviews are stop words removal, special characters removal, removing white spaces, transforming text to lower case, and stemming. A machine learning approach is employed to create this model, necessitating a labeled dataset. However, the existing dataset lacked labels for every review, although it did contain rating information for each entry. The ratings data were transformed into labels as follows: Ratings higher than 7 were classified as positive and given a label of 1. Ratings lower than 4 were classified as negative and assigned a label of −1. Ratings falling between 4 to 7 were considered as neutral and labeled as 0. Table 1 shows ratings data with their respective labels.

To handle the large dataset, it was downsized to contain 60,000 rows of data only. Out of 60,000 rows, 20,000 rows were labelled as Positive, another 20,000 as Neutral, and the remaining 20,000 as Negative. This selection ensured a balanced representation of different sentiment categories. Subsequently, only this reduced dataset was utilized for the subsequent modelling tasks. Then, the dataset was split into two parts: 80% for training data and 20% for testing data. This division allowed for the model to be trained on a majority of the data while preserving a separate subset for evaluating its performance. After the completion of data splitting, the feature extraction process was applied to both the training and testing datasets. Two distinct algorithms were employed for this purpose: TF-IDF and Bio BERT. Bio BERT is an adapted version of the BERT algorithm designed specifically for biomedical text analysis. It undergoes pre-training on a vast collection of biomedical literature and clinical text to acquire a deep understanding of word and sentence contexts. By learning from this extensive corpus, Bio BERT gains the ability to generate contextualized representations of biomedical terms and phrases. Once the dataset has been transformed into vectors, the next step involves training the sentiment analysis models. For this task, three distinct algorithms were employed: Logistic Regression, Perceptron, and LSTM (Long Short-Term Memory). A total of six sets of models were developed for sentiment analysis. These models were created by combining two feature extraction algorithms with the three aforementioned training algorithms. This approach allowed for a comprehensive exploration of various combinations, enabling the comparison and evaluation of the different feature extraction techniques and training algorithms for sentiment analysis. Grid search was applied to determine the optimal parameters for training each model across all the training algorithms. The resulting models were then evaluated, and their accuracy, macro-average f1-score, and macro-average precision scores were recorded. Table 2 summarizes and compares the performance metrics of the trained models.

Based on the results shown in Table 2, it is evident that the models utilizing the TF-IDF vectorizer for feature extraction generally outperformed the Bio BERT-based models. Among the TF-IDF models, the Logistic Regression algorithm demonstrated the highest performance compared to Perceptron and LSTM. It achieved an accuracy of 0.7441, an F1-score of 0.7449, and a Precision of 0.7444 respectively. Based on these findings, to develop an effective sentiment analysis model, the TF-IDF algorithm is chosen for feature extraction, while Logistic Regression is selected as the preferred model training algorithm.

The best trained sentiment analysis model and vectorizer are converted into.pkl files and incorporated into the backend, which utilizes FastAPI. This model serves two purposes within the system. Firstly, when an administrator adds a new drug to the system, all the accompanying reviews uploaded for that drug are automatically classified according to their sentiment using the model. This allows for the categorization of reviews as positive, neutral, or negative. Secondly, healthcare professionals are given the option to contribute new reviews for any specific drug, which are also subjected to sentiment classification using the trained model. Before prediction, each review undergoes pre-processing similar to the data cleaning process applied to the dataset. By employing this model, the system ensures that all newly added reviews, whether uploaded by administrators or healthcare professionals, undergo sentiment analysis to provide valuable insights into the sentiment associated with specific drugs. This enables the assignment of a sentiment rating to each drug. A sample interface for the sentiment analysis is illustrated in Fig. 6.

To provide drug recommendations based on similarity, the system incorporates a content-based filtering model. This model analyses the characteristics of drugs to identify similarities and make recommendations. When a user views a specific drug, the model suggests other drugs that are similar to the one being viewed. The development of this model involves several steps, starting with the acquisition of dataset containing drug information and details. The drugs data used for this model includes the following specific attributes: Generic Name, Brand Name, Common Side Effects, Description, Dosage, Food Interaction, Indications, Ingredients, Manufacturer, Price, Rating, and Things to Avoid. All these attributes are important to determine the similarity between drugs. After reading the data, the data undergoes several text pre-processing steps. Since all data is in string format, the text pre-processing steps include conversion to lower case, removal of stop words, removal of special characters and lemmatization. After these pre-processing steps, the individual attributes of each drug are joined together to create a corpus.

Each drug in the database has its own corpus consisting of its attributes. Then the TF-IDF algorithm is used to perform feature extraction on these corpuses. All corpora are converted to vectors. In order to find similar drugs for a given drug, attributes of that drug undergo text preprocessing, and are then combined to create a corpus as well. This corpus is subsequently vectorized using the TF-IDF algorithm, transforming it into a numerical representation. After completing these steps, the similarity between drugs is determined. The Cosine Similarity algorithm is employed for this purpose. Both the vectors of drugs in the database and the vector of the drug being viewed are fitted into the Cosine Similarity algorithm. The corpus of the drug being viewed is compared with the corpuses of all other drugs in the database and the similarity scores of each of the drugs is computed. These drugs are then sorted based on ranking using the argsort() function depending on the similarity score. The indices of the top 5 drugs are retrieved based on the similarity scores. Using these indices, the corresponding drugs are retrieved from the database. Essentially, this approach constitutes a content-based recommender system, where drugs being viewed by users are compared with the rest of the drugs in the database. The top five most similar drugs are recommended to the user based on this comparison.

During the implementation of the model into the system, an initial step involves generating TF-IDF vectors for all the drugs beforehand. The resulting TF-IDF vectors for all drugs are then stored in a.pkl file for easy retrieval. When a user interacts with the system’s user interface and selects a specific drug, the TF-IDF vectors of all drugs in the database are loaded from the.pkl file. These vectors, along with the vector representation of the drug being viewed, are then fitted to the cosine similarity algorithm. Finally, the recommended drugs are retrieved from the model and displayed to the user. Whenever a new drug is added to the database, the TF-IDF vectors associated with the drug data are regenerated. Subsequently, the existing.pkl file is updated to include these new vectors. Figure 7 and 8 depict the user interface of the content-based filtering model.

As shown in Fig. 7, the focus is on the drug ‘Eletriptan’, commonly used for migraines. The top three recommended drugs are ‘Rizatriptan,’ ‘Sumatriptan,’ and ‘Venlafaxine,’ which are also used for migraine treatment. The remaining two drugs are used for different purposes but are the most similar options available to ‘Eletriptan’ within the limited database of only four drugs for migraine treatment. Another example is demonstrated in Fig. 8, where the drug under consideration is ‘Cephalexin,’ utilized for treating ‘bladder infection.’ The top two recommended drugs are ‘Nitrofurantoin’ and ‘Ciprofloxacin,’ which are also commonly used for ‘bladder infection.’ As there are only three drugs in the database for ‘bladder infection,’ the remaining displayed drugs represent the closest alternatives compared to the others in the database. This successful outcome illustrates the effective functionality of the content-based filtering model.

The collaborative filtering model has been developed to recommend drugs to users based on their similarity to other users. Data for this model is collected directly from the system, capturing user behavior such as drug views, reviews, and additions to the wish list. These interactions, which can be considered as user ratings, are stored in the database and used to construct the collaborative filtering model. The initial steps involve loading the user behavior data and drug data, followed by preprocessing, and encoding. To ensure a comprehensive set of drugs is considered for recommendations, the unique drugs from both datasets are combined. A Label Encoder is then utilized to assign numeric IDs to each drug, facilitating efficient processing of categorical drug names. The core element of the collaborative filtering approach is the user-item matrix. An empty matrix is created to store user ratings for drugs. The user interactions with drugs are assigned different weights: adding drugs to the wish list is given a weight of 2.0, viewing a drug is given a weight of 1.0, and reviewing a drug is given a weight of 3.0. These weightings reflect the importance placed on each user’s behavior. Using the drug encodings and assigned weights, the user ratings are populated in the user item matrix. Some drugs may have missing ratings, indicating that no users have rated those specific drugs. Handling missing ratings is crucial to address the sparsity of the user-item matrix. If a drug has not been rated by any user, its rating is replaced with the average rating of all drugs. This step ensures that the model can provide reasonable recommendations for unrated drugs based on collective user behavior.

Once the user-item matrix is prepared, it undergoes transformation using Singular Value Decomposition (SVD), which is a dimensionality reduction technique. The TruncatedSVD class from scikit-learn is employed to perform SVD on the user-item matrix, allowing for lower-dimensional representations. After transforming the matrix, the next step involves calculating the similarity between users. Cosine similarity is used to measure the similarity between the transformed user-item matrix representations of different users. This similarity matrix captures the extent of similarity between users’ behaviors in drug ratings. Moving on to the recommendation process, similarity scores between the user in need of recommendations and other users in the system are calculated. These scores are utilized to identify similar users based on their similarity to the target user. Drug ratings from similar users are collected, forming the basis for generating recommendations for the target user. Recommendations are based on drugs that similar users have rated, but the target user has not yet rated. Finally, the inverse transform function is applied to the encoder to retrieve the recommended drug names based on their IDs. The code snippet for the recommendation process is shown below.

The collaborative filtering model is seamlessly integrated into the system, streamlining the data loading and user item matrix transformation steps. The user item matrix is transformed and stored in a.pkl file for efficient retrieval. When a recommendation is requested for a particular user, only the user index is needed to retrieve their similarity scores. Using these similarity scores, the model identifies users who exhibit similar preferences and behaviors. It then collects drug ratings from these similar users, which are aggregated to determine the overall ratings for each drug. Based on these aggregated ratings, the drugs are sorted to prioritize the most highly recommended ones. The updated recommendations are displayed to the user in the main page of the system, specifically in a dedicated section titled “Recommended Drugs for You,” as depicted in Fig. 9. Each user in the system will receive a personalized set of recommended drugs based on their individual behaviors. For newly registered users, recommendations would not be available initially. However, once the model is trained again and they start to interact with the system, they start receiving personalized recommendations. It is important to note that recommended drugs for similar users dynamically adapt over time, reflecting their own behaviors as well as those of other users within the system.