Madhurika Situt

It is possible to observe rising health information needs and improvements in information-seeking behavior all over the world. 81 % use the Internet, according to recent studies, and 59 % say they have searched for health information about diseases, diagnoses, and various treatments online. As educated patients raise questions or discuss treatment options, such effects influence the patient-physician relationship. Thus, in the decision-making process, patients tend to become active participants. Patient empowerment is often referred to as this shift in the way of thought.

Medicine and a common treatment recommendation system can help patients or medical care providers choose the best medication for various symptoms or conditions. This system would also prescribe other typical treatments based on the conditions faced by patients.

Such recommendation systems will help improve the innovative technology available today that can contribute to more concise decisions. Based on various algorithms, several current recommendation systems are being developed.

The project focuses on a Medicine and a common treatment recommendation system based on the data provided by Drugs.com, DrugsLib.com, and WebMD.com.

Typically, recommendation systems are broken down into three types: content filtering, collaborative filtering, and hybrid filtering (a mix of the two). My project aims to outline several methods for developing recommendation systems for collaborative filtering, to recognize the advantages and the disadvantages of the various approaches, and finally to recommend the best method to build a medical approach and to recommend a common treatment.

"To empower patients with the knowledge to better manage their own healthcare and to improve consumer safety by assisting in the reduction of medication errors."

Hospitals provide access to a huge volume of medical data and criteria for their wellbeing. This has greatly expanded the available digital information for patient-oriented decision-making. This digital content is also spread through multiple platforms, which prohibits people from accessing helpful information to enhance their well-being.

Therefore, there is a need for easier ways for medical professionals to efficiently use this material. An example will be access to aggregated data from the latest database on a particular issue at the point of treatment as appropriate. In comparison, more guidelines for medications, testing, and care (e.g., evidence-based medicine or therapeutic pathways) are open to medical professionals every day. Therefore, depending on their symptoms, clinical findings, or prior medical records, it becomes more impossible for them to determine the medication to offer a patient. On the other hand, all these data can be used to aim for personalized healthcare, which is currently on the rise and is anticipated in the coming years to achieve a substantial disruptive trend in healthcare.

To fill this void and facilitate decision-making during therapy, a recommendation engine for medical use may also be used. The engine will search for individuals with similar criteria in the database based on the current health status of a patient, prehistory, current prescriptions, symptoms, and past treatments. The recommended system would eventually suggest the medications that have been most effective for similar patients.

Medical information is one of the most impacted and searched subjects on the internet. Nearly 60 % of adults are searching for adequate health information on the internet, according to the Pew Internet and American Life Project, with 35 % of respondents focused on diagnosing diseases online only. Since several studies indicate that many people die as a result of medical errors, medical practitioners who prescribe are responsible for these errors. Based on their experiences, drugs. Since most of their interactions are limited, they sometimes make mistakes. This research provides practitioners with a medication recommendation method that can be used by them when prescribing medicines. A recommended framework is an ordinary framework that allows users to get a suggestion of items they can use for their exact requirements.

Health recommendation, unlike many various forms of systems, focuses mainly on patients 'enthusiastic, physical and mental problems. A drug recommendation system, focused on patient reports, is a similar system that recommends medicines for a specific illness. In this fast-growing world of technology, which can save lives by supporting doctors, this system is very important.

Recommended systems aim to provide consumers with customized inventory and repair to change the growing over-burden disadvantage of online data. Since the mid-1990s, various recommended framework methods have been expected, and multiple types of recommended framework code have been produced for the spread of applications as of late.

Due to the evolving requirements of the organization using and implementing it, evaluating the success of the recommendation system is difficult. Generally, customer satisfaction is the most indicative metric. While it is not possible to quantify the satisfaction of users by using a heuristic formula, based on how well they can handle common problems, we can still assess the efficiency of recommendation systems.

The main challenges that exist in current recommendation systems are –

1. Cold Start -

A recommendation system does not work optimally when there is insufficient information or metadata available. It is possible to divide cold starts into two distinct subsets: cold starts of the product and cold starts of the consumer. Whenever a new item is displayed on an e-commerce platform, it goes through the cold start of the product, and due to the lack of user engagement, there is no feedback. The recommendation system would not know when to view the ad relating to that product if there are not enough user interactions.

When a user creates an account for the first time, the cold-start behavior happens and does not have any product preferences or background available to base recommendations. For new or current users, the cold start issue persists. We found that the Bayes classifier is most used when evaluating the metrics and methods for cold-start recommendations. Graphical models used in probability and artificial intelligence are Bayesian models. A type of Bayesian reasoning, whether it is content- or collaborative-based, is likely to be implemented in model-based recommendation systems. The Naive Bayes model is the most common way of using Bayesian models. It has proven to be the most reliable, considering its simplicity. Different attributes are mutually independent characteristics of the items in the Naive Bayes classification.

2. Data Sparsity-

Data sparsity arises from the fact that only limited items are intended by users to rate. Much of the recommendation system groups similar users' ratings; however, due to the lack of incentives or user awareness to rate products, the recorded user-item matrix has empty or unknown ratings (up to 99 %). Recommendation systems may also provide those that provide no reviews or ratings with unfair recommendations.

To mitigate this challenge, Non-Matrix Factorization is one of these algorithms based on dimensionality reduction techniques.

Other challenges with the existing recommendation systems are to recognize that it is important to define the features that make an outstanding recommendation and to identify how to quantify a recommendation system.

By implementing multiple models, this project mainly focuses on mitigating the above challenges and comparing them using metrics such as recall, accuracy, accuracy, ROC curves, similarity score, and F-measure to determine the efficiency of the recommendation system to get the best recommendation system.

• To develop a system of recommendations to recommend medicines and optional treatments based on patient reviews and conditions.

• Different collaborative-driven filtering methods to provide recommendations based on patient reviews would be introduced to provide recommendations based on circumstances. The approaches will also consider and overcome the problems facing the current medicine recommendation framework.

• Between clustering-based algorithms, matrix factorization-based algorithms, and deep learning, I will research and compare the various models using metrics such as RMSE, MSE, ROC / AUC, Similarity score, Novelty.

The proposed medicine and treatment recommendation system uses machine learning recommendation approaches, and the best one will be selected for the recommendation system to achieve metrics such as good accuracy, scalability, and model performance.

Below data sources will be used to build this recommendation system –

• Drugs.com (https://www.drugs.com/ ) -

1. The most reliable, up-to-date drug knowledge is offered by this source. This would be used to obtain the patients' updated reviews and additional details such as conditions relevant to the medication.

2. Additional or optional care recommendations from this source will be scrapped based on the conditions.

3. Features -

• • • • • • • DrugName

            • Condition

            • Review

            • Rating

            • Date

            • UsefulCount

4. Data Size - 107 MB

5. Data Format - TSV files

• Druglib.com (http://www.druglib.com/ ) -

1. It is another source that will be used to provide the drug feedback and extra reviews such as effectiveness, side effects, scores.

2. Features-

   1. • 1. • UrlDrugName

           • Rating

           • Effectiveness

           • SideEffects

           • Condition

           • BenefitsReviews

           • SideEffectsReview

           • CommentsReview

3. Data Size- 3MB

4. Data Format- TSV files

• https://archive.ics.uci.edu/ml/datasets/Drug+Review+Dataset+%28Druglib.com%29 -

1. This source provides the data in the CSV formats referred to in the points above.

• WebMD (https://www.webmd.com/ ) -

1. This source offers the patients' review along with the age, useful information of the count.

2. This will mainly be used for the analysis of exploratory data.

3. Features-

   1. • • • Age

          • Condition

          • Date

          • Drug

          • DrugId

          • Easeofuse

          • Effectiveness

          • Review

          • Sex

          • Sides

          • UsefulCount

4. Data Size- 172 MB

5. Data Format - CSV files

Below characteristics of Data are going to analyze -

• Top 20 conditions that have the highest number of medicines available.

• Bottom 20 conditions that have the least medicines available.

• The number of reviews over time and looking into how it performed.

• Exploring ratings received by customers to similar conditions.

• Exploring co-relation between rating and useful count.

• Annual ratings for medicines for specific conditions.

I want to use patient reviews to identify the most popular prescription medication and recovery types.

1. Drugs.com -

These stats are based on Drugs.com reviews (.csv), which include a rating, a brief descriptive review, and other customer comments such as yes or no responses to how helpful a review was. DrugName, Condition, Review, Scores, Date, and Usefulcount are some of the features included in this dataset.

Here, I began by looking at variables starting with the uniqueID given in the dataset to see if the consumer has left several reviews for a certain medication.

It appears there is only one review by each customer.

Then I looked at the details to see how many medications were available for each condition. I merged the training and testing datasets for this study and discovered the top 10 diseases for the most medications available.

As you can see from the picture above, minimum of 85 drugs are available for top 10 conditions individually.

Also, I investigated average rating for the top 10 drugs available in dataset and found the below findings –

From the above analysis, it appears Phentermine drug has the highest rating of 8 out of the top 10 drugs.

Now looking into the count of reviews over time, it was found the number of reviews has drastically increased over time as presented below.

I also carried out the distribution plot to explore the ratings count and its distribution over the data and below distribution is observed for all the ratings.

2. Druglib.com -

User ratings on individual medications, as well as associated diseases, are included in the dataset, as well as a 10-star patient rating indicating overall patient satisfaction.

The data is structured in such a way that a patient with a uniqueID orders a prescription that is appropriate for his condition and then reviews and rates the drug on the date of purchase. If other people read the summary and find it interesting, they will press usefulCount, which will increase the variable by one.

To explore the dataset in more details, firstly I carried out a count of drugs that are available for the top 30 conditions -

I also looked into the user ratings distribution across the complete dataset and found the below distribution –

3. Webmd.com -

The dataset provides user reviews on specific drugs along with related conditions, side effects, age, sex, and ratings reflecting overall patient satisfaction.

To further analyze the data, I looked into the unique value count of all the features present in the data set and found the below observations –

I also observed the different satisfaction levels based on the top 5 drugs with higher reviews count and found below analysis –

From the above plot, I found out that ‘lexapro’ drug has the most reviews present in the dataset with the satisfaction level 5.

To further analyze the data, I looked into other feature and carried out the review count for different age group and how those medications were effective for different age group based on the effectiveness feature present in dataset and found the below analysis –

1. Missing values removal -

Conditions had about 1000 missing values, so they were deleted from the dataset after further analysis.

2. Conditions Preprocessing -

It was discovered that the conditions included some meaningless entries, such as "0 /span> users found this comment helpful." These entries are not useful for the recommendation system. As a result, such conditions are removed from the results. Also, conditions with only one medicine available have been removed, as the recommendation system will continue to recommend the same medication regardless of consumer feedback.

3. Reviews Preprocessing -

Preprocessing in Review involves deleting stop words, stemming, and cleaning text. I've used nltk's pre-existing stop words in this case. There are also words like "needn't" that include the word "not." I removed these terms from the list of stop words because they are necessary components of emotional analysis.

The process of reducing a word to its base word or stem in such a way that similar words are grouped together under a single stem is known as stemming. I'm using nltk's snowball stemmer, also known as Porter2 stemmer, which is a more advanced variant of Porter Stemmer.

Other steps involve deleting HTML text, adding space, changing to lower case, removing stop words, stemming words, and connecting them with space again.

COLLABORATIVE FILTERING -

Collaborative filtering filters information by using the interactions and data collected by the system from other users. It is based on the idea that individuals who have agreed on certain items in their assessment are likely to agree again in the future.

The so-called similarity index-based approach is applied by most collaborative filtering systems. In the neighborhood-based method, based on their similarities to the active user, multiple users are chosen. Active user inference is made by measuring the weighted average of the selected user ratings.

The relationship between users and objects is the subject of collaborative-filtering systems. The similarity of the items is determined by the similarity between the users who have rated all items and the ratings of those items.

There are two classes of Collaborative Filtering:

1. User-based, which measures the similarity between target users and other users.

2. Item-based, which measures the similarity between the items that target users’ rate or interact with and other items.

Types of collaborative filtering –

The CF techniques are broadly divided into 2-types:

1. Memory Based approach.

2. Model-Based approach.

Model-Based approach -

In this method, CF models are built to predict the medicine/drug-using machine learning algorithms. The algorithms in this method can further be broken down into 3 sub-types-

I built a sentiment column in the dataset to determine if the feedback and ratings given are positive or negative. In order to do so, I measured the polarity and subjectivity of the customer's feedback.

Polarity is a float in the range [-1,1], with 1 denoting a positive review and -1 denoting a negative review. The terms "opinion," "emotion," and "judgment" all refer to subjective details, while "objectivity" refers to empirical data.

Subjectivity is also a float with a value between 0 and 1. I calculated the polarity and subjectivity of each analysis using the text blob library and saved the results in a new column inside the dataset. If the polarity score is equal to or less than 0, the sentiment of the review is negative, and the value of the sentiment column is 0; if the polarity score is greater than 0 and the value of the sentiment column is 1, the sentiment of the review is positive, and the value of the sentiment column is 1. This will be our goal column, which will tell us whether the drug is successful or not based on the sentiment score.

To tokenize and create a vocabulary for drug reviews, I used the CountVectorizer vectorizer, which will be our function for training the machine learning model for our recommendation system.

The CountVectorizer is a simple tool for tokenizing a collection of text documents and creating a vocabulary of known words, as well as encoding new documents with that vocabulary.

I have implemented multiple models for the recommendation system, which are-

1. Logistic Regression

2. Multinomial Naïve Bayes

3. Bernoulli Naïve Bayes

4. LinearSVC

5. Decision Tree

6. Random Forest

7. Simple Neural Network

8. Convolutional Neural Network

For logistic regression, multinomial naïve Bayes, Bernoulli naïve Bayes, linearsvc, decision tree, and random forest algorithms, I am using GridsearchCV library provided by Sklearn to find the best parameter for the models and metrics used to calculate the best algorithm are –

1. Accuracy–

Accuracy is the ratio of the number of correct predictions to the total number of input samples.

2. ROC-AUC–

It demonstrates how accurate our model is at ranking predictions. It tells us what the chances are that a positive review chosen at random would be rated higher than a negative review chosen at random.

3. Root Mean Square Error–

The RMSE is the square root of the residuals' variance. It shows how well the model fits the data in terms of absolute fit–how similar the observed data points are to the model's expected values.

1. Mean Square Error–

MSE is calculated by taking the average of the square of the difference between the original and observed values of the data.

2. Precision –

Precision is a metric that measures how many accurate positive predictions have been made. It's the proportion of correctly predicted positive reviews to total positive reviews predicted.

3. F1_score –

The weighted average of Precision and Recall is the F1 Score. As a result, this score considers both false positives and false negatives.

4. Recall score –

Recall score is the ratio of correctly predicted positive observations to all observations in the actual class.

All the above metrics are used from Sklearn.

Output observed for the above algorithms are as below –

From the above output, it is clear that LinearSVC performed better than all the other algorithms with an accuracy of 90%, roc_auc_score of 0.93, RMSE score of 0.49, and MSE score of 0.24.

Neural Networks –

The ability to learn a more complex, non-linear function is one of the reasons for the performance improvement from a deeper network. This makes it easier for the networks to distinguish between different classes if they have enough training data.

For this recommendation system, I have implemented a simple sequential neural network and convolutional neural networks using Keras Tensorflow.

1. Simple sequential linear neural networks –

The Simple sequential neural network model is a linear stack of layers. I have used 5 dense layers and 3 batch normalization layers to normalize the inputs of each layer in such a way that they have a mean output activation of zero and a standard deviation of one. For activation functions, I have used Relu and Sigmoid. Sigmoid is used to predict the probability as an output since the probability of anything exists only between the range of 0 and 1 and it will help to predict if the drug for the specific condition is good or not for the specific condition.

Model summary-

For this model, I have received accuracy of 90% which is good accuracy for simple neural network.

The tradeoff of loss and accuracy is visualized below –

2. Convolutional Neural Network -

A convolutional neural network (CNN) is a type of feed-forward neural network which applies convolution operation in place of general matrix multiplication in at least one of its layers.

A typical CNN consists of three components that transform the input volume into an output volume, namely, convolutional layers, pooling layers, and fully connected layers.

These layers are stacked to form convolutional network architectures.

1. Convolution: As being the core operation, convolutions aim to extract features from the input. Feature maps are obtained by applying convolution filters with a set of mathematical operations. Here, I have used Conv1D as core since my input is a single number or a one-element tuple.

2. Nonlinearity: In order to introduce nonlinearities into the model, an additional operation, usually ReLU (Rectified Linear Unit), is used after convolution operation.

3. Pooling (Subsampling): Pooling reduces the dimensionality of the feature maps to decrease processing time. I have used GlobalMaxPooling1D as it set the pool size equal to the input size so that the max of the entire input is computed as the output value.

4. Classification: The output from the convolutional and pooling layers represents high-level features of the input.

Model Summary -

For this model, I have received an accuracy score of 73%, which is not enough for our recommendation system. I have also tried to increase the number of layers, although it didn’t affect the accuracy of the model.

The tradeoff of loss and accuracy is visualized below -

I calculated the Mean Square error score by carrying out the average of all the mse scores from the epochs and found the below trade off between simple sequential and convolutional neural network mse scores.

The following is the final report table with all of the metrics data for all of the algorithms -

Looking at the above observations it appears, simple sequential linear neural network and LinearSVC will be the best for our recommendation system based on the accuracy. Further looking into the MSE scores, It can be concluded simple sequential neural network will provide the best results among all the models.

Carrying out the predictions/recommendations using a simple sequential linear neural network.

To carry out the recommendations, I have carried out the predictions using simple sequential neural network model. Only top 5 values from predictions are selected with high probabilities values. I have merged the predicted values to testing data to carry out the drug name.

Further, I have sorted the data based on the ratings present in the testing data and below medicine and treatments were recommended by the recommendation system for the condition 'Anxiety'-

Now to further evaluate if the recommendation system is performing well, I am using the different metrics from recmetrics library.

Recmetrics - It is a python library of evaluation metrics and diagnostic tools for recommender systems.

For this, I have created 2 more recommendation system -

1. Popularity recommender - It is a simple popularity recommender to demonstrate recommender metrics in action. The popularity recommender simply recommends the top 5 drugs. To carry out predicted values I have selected the first value of testing data and assigned the same to the complete dataset.

2. Random recommender - This recommender takes the random values from the actual target values.

After implementing the popularity recommender and random recommender, our predicted values looks like below-

I have used 2 metrics from recmetrics to check if our fitted medicine and treatment recommendation system using a simple sequential linear neural network algorithm is able to recommend appropriate recommendations.

1. Mean Average Recall -

Mean Average Recall at K (Mar@k) measures the recall at the kth recommendations. Mar@k considers the order of recommendations, and penalizes correct recommendations based on the order of the recommendations.

From the above plot, it is clear that our collaborative filtering recommendation system has a better recall score than other random recommender systems.

2. Coverage -

Coverage is the percent of items that the recommender is able to recommend. It referred to as prediction coverage and it's depicted by the next formula.

Where 'I' is the number of unique items the model recommends in the test data, and 'N' is the total number of unique items in the training data.

From above plot, it seems collaborative filtering recommendation system is capable of providing recommendations for the dataset.

From above both the metrics, it is clear that our fitted medicine and treatment recommendation system providing appropriate recommendations.

To that end, I believe that the best type of medicine and treatment recommendation is likely a simple sequential linear neural network model.

The neural network model is the best because it captures the advantages of both types while also mitigating the disadvantages.

The recommender will now have the advantage of providing novel recommendations that could help customers explore drug options that are new and different but are still likely to be considered.

Future development steps-

1. Creating a Hybrid recommender-

In this project, I'm looking into different model-based collaborative filtering approaches. Hybrid recommenders, on the other hand, provide a solid base for pursuing newer opportunities such as contextualizing recommendations, using parallel hybrid algorithms, and processing broader datasets, among other things.

2. Mitigating limitations of Collaborative filtering approach -

Since the embeddings are automatically learned, this method does not require domain knowledge. The dot product of the corresponding embeddings is the model's prediction for a given (user, item) pair. As a result, if an object isn't seen during training, the system won't be able to construct an embedding for it or question the model for it. The cold-start problem is a term used to describe this concern.

With a hybrid recommendation method, this problem can be solved.

3. Avoiding repeated recommendations for the same customers-

This recommendation system gives out repeated recommendations to consumers who have given the medication a positive or negative review. I would change the referral framework so that users who have already provided drug reviews don't get recommendations for the same drugs again.

Github Link - https://github.com/Madhurika1292/Medicines-and-Common-Treatment-Recommendation-System

• The science and Information (SAI) Organization. (n.d.). Retrieved February 17, 2021, from https://thesai.org/

• S. Bankhele, A., E. Bresso, R., A. Cami, A., J. Castro, F., EA. Charkhat Gorgich, S., L. Chen, P., . . . Y. Zhang, M. (1970, January 01). Recommender systems in the healthcare domain: State-of-the-art and research issues. Retrieved February 17, 2021, from https://link.springer.com/article/10.1007/s10844-020-00633-6

• Call for research papers - 2021. (2011, November 23). Retrieved February 17, 2021, from http://www.ijstr.org/

• Fayyaz, Z., Ebrahimian, M., Nawara, D., Ibrahim, A., & Kashef, R. (2020, November 02). Recommendation systems: Algorithms, Challenges, metrics, and business opportunities. Retrieved February 17, 2021, from https://www.mdpi.com/2076-3417/10/21/7748

• A new algorithm for solving data sparsity Problem Based-on non negative matrix factorization in recommender systems. (n.d.). Retrieved February 17, 2021, from https://ieeexplore.ieee.org/document/6993356

• Prescription drug Information, interactions & side effects. (n.d.). Retrieved February 17, 2021, from https://www.drugs.com/

• Drug information, Research, clinical Trials, News. (n.d.). Retrieved February 17, 2021, from http://www.druglib.com/

• UCI machine Learning Repository: DRUG Review Dataset (DRUGS.COM) data set. (n.d.). Retrieved February 17, 2021, from https://archive.ics.uci.edu/ml/datasets/Drug+Review+Dataset+%28Drugs.com%29

• Better information. better health. (n.d.). Retrieved February 17, 2021, from https://www.webmd.com/

• Bartolini, L. (2016, February 1). Practice Current: How do you treat anti-NMDA receptor encephalitis? Neurology Clinical Practice. https://cp.neurology.org/content/6/1/69.

• https://daiwk.github.io/assets/Batmaz2018_Article_AReviewOnDeepLearningForRecomm.pdf

• Bartolini, L. (2016, February 1). Practice Current: How do you treat anti-NMDA receptor encephalitis? Neurology Clinical Practice. https://cp.neurology.org/content/6/1/69.

• Bleeding Disorders. Froedtert & the Medical College of Wisconsin. (n.d.). https://www.froedtert.com/benign-hematology/conditions/bleeding-disorders.

• Centers for Disease Control and Prevention. (n.d.). Centers for Disease Control and Prevention. https://www.cdc.gov/.

• Chirag M Sandesara, M. D. (2020, December 5). Atrioventricular Block Treatment & Management: Approach Considerations, Pacemaker Implantation, Long-Term Monitoring. https://emedicine.medscape.com/article/151597-treatment.

• Chris, N, S., & Shobi. (2020, January 30). What are Max Pooling, Average Pooling, Global Max Pooling and Global Average Pooling? MachineCurve. https://www.machinecurve.com/index.php/2020/01/30/what-are-max-pooling-average-pooling-global-max-pooling-and-global-average-pooling/.

• Chronic eosinophilic leukemia - Canadian Cancer Society. www.cancer.ca. (n.d.). https://www.cancer.ca/en/cancer-information/cancer-type/leukemia/leukemia/chronic-eosinophilic-leukemia/?region=on.

• Figueiredo, M. (2020, June 15). Brensocatib Named FDA Breakthrough Therapy for Non-CF Bronchiectasis. Bronchiectasis News Today. https://bronchiectasisnewstoday.com/2020/06/15/brensocatib-granted-fda-breakthrough-therapy-status-for-non-cystic-fibrosis-bronchiectasis/.

• Healthline Media. (n.d.). Medical information and health advice you can trust. Healthline. https://www.healthline.com/.

• Home. Digestive Health Centre. (2020, October 11). http://www.digestivehealth.com.au/.

• Mayo Foundation for Medical Education and Research. (n.d.). Mayo Clinic. https://www.mayoclinic.org/.

• medicinenet.com. (n.d.). https://www.medicinenet.com/.

• Primary CNS Lymphoma Treatment (PDQ®)–Patient Version. National Cancer Institute. (n.d.). https://www.cancer.gov/types/lymphoma/patient/primary-cns-lymphoma-treatment-pdq.

• Rolain, J. M., Brouqui, P., Koehler, J. E., Maguina, C., Dolan, M. J., & Raoult, D. (2004, June 1). Recommendations for Treatment of Human Infections Caused by Bartonella Species. Antimicrobial Agents and Chemotherapy. https://aac.asm.org/content/48/6/1921#:~:text=All%20immunocompromised%20patients%20with%20a,prophylactic%20antibiotic%20treatment%20with%20erythromycin%20(.

• Shubham JainI am currently pursing my B.Tech in Ceramic Engineering from IIT (B.H.U) Varanasi. I am an aspiring data scientist and a ML enthusiast. I am really passionate about changing the world by using artificial intelligence. (2020, December 23). NLP For Beginners: Text Classification Using TextBlob. Analytics Vidhya. https://www.analyticsvidhya.com/blog/2018/02/natural-language-processing-for-beginners-using-textblob/.

• sociopath00. (2018, October 16). Random Forest using GridSearchCV. Kaggle. https://www.kaggle.com/sociopath00/random-forest-using-gridsearchcv.

• Specialists in Medical Care. Neurological Specialists, P.C. 99 Hawley Lane | Home Page. (n.d.). https://www.neurologyct.com/.

• Team, G. T. E. (n.d.). How to Help Someone Who Doesn't Like Their Body. Body Image –. https://www.goodtherapy.org/learn-about-therapy/issues/body-image/get-help.

• U.S. Department of Health and Human Services. (n.d.). Chronic Inflammatory Demyelinating Polyneuropathy (CIDP) Information Page. National Institute of Neurological Disorders and Stroke. https://www.ninds.nih.gov/disorders/all-disorders/chronic-inflammatory-demyelinating-polyneuropathy-cidp-information-page.

• U.S. National Library of Medicine. (n.d.). National Center for Biotechnology Information. https://www.ncbi.nlm.nih.gov/.

• Urogynecological Risk Assessment in Postmenopausal Women. Medscape. (2014, January 15). https://www.medscape.com/viewarticle/814321_4#:~:text=Atrophic%20urethritis%20and%20trigonitis%20can,indications%20for%20systemic%20estrogen%20replacement.

• Çano, E., & Morisio, M. (2019, January 12). Hybrid Recommender Systems: A Systematic Literature Review. arXiv.org. https://arxiv.org/abs/1901.03888#:~:text=Being%20a%20hot%20topic%2C%20hybrid,%2C%20processing%20larger%20datasets%2C%20etc.

• Belhekar, C. (2020, August 28). Recommendation Systems - A walk through. Medium. https://chaitanyabelhekar.medium.com/recommendation-systems-a-walk-trough-33587fecc195.

• Harode, R. (2020, March 8). WebMD Drug Reviews Dataset. Kaggle. https://www.kaggle.com/rohanharode07/webmd-drug-reviews-dataset.

• Kurama, V. (2019, September 4). A Simple Introduction to Collaborative Filtering. Built In. https://builtin.com/data-science/collaborative-filtering-recommender-system.

• Grace-Martin, K., Ali, M. M., Kewal, Shawn, Ulrich, Mathur, N., … Grateful2U. (2020, November 2). Assessing the Fit of Regression Models. The Analysis Factor. https://www.theanalysisfactor.com/assessing-the-fit-of-regression-models/#:~:text=The%20RMSE%20is%20the%20square,an%20absolute%20measure%20of%20fit.

• Turkish Journal of Computer and Mathematics Education (TURCOMAT). (n.d.). https://turcomat.org/.

• Khorram, B., & Yazdi, M. (2019, February). A New Optimized Thresholding Method Using Ant Colony Algorithm for MR Brain Image Segmentation. Journal of digital imaging. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6382633/.

• Team, K. (n.d.). Keras documentation: The Sequential model. Keras. https://keras.io/guides/sequential_model/.

• Sharma, S. (2019, February 14). Activation Functions in Neural Networks. Medium. https://towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6.

• A review on deep learning for recommender systems: challenges and remedies. PDF Free Download. (n.d.). https://docplayer.net/155168956-A-review-on-deep-learning-for-recommender-systems-challenges-and-remedies.html.

• Brownlee, J. (2019, July 5). A Gentle Introduction to Pooling Layers for Convolutional Neural Networks. Machine Learning Mastery. https://machinelearningmastery.com/pooling-layers-for-convolutional-neural-networks/.