Mengxi Li

Introduction

In a recent project, I applied advanced text processing and machine learning techniques to analyze sentiment in healthcare reviews. Using Python, I developed a model to classify doctor reviews into two categories: 5-star (excellent) or non-5-star ratings, based on the depth of “doctor knowledge” mentioned in the reviews.

Objective

The aim was to train and test a logistic regression model using a dataset of 80,000 training reviews and then predict the sentiment on an additional 20,000 test reviews. The challenge was to accurately classify these sentiments, which are vital for enhancing patient care and service quality.

Methodology

Data Preparation:

• Data Loading and Preprocessing: Reviews were loaded using pandas and preprocessed to shuffle and structure the dataset for analysis. Key features like the review state and posted time were extracted and converted to categorical variables.

Feature Engineering:

• TF-IDF Vectorization: I utilized TfidfVectorizer from scikit-learn to transform text data into a numeric format, emphasizing words that are significant for sentiment analysis.
• Temporal Features: The ‘postedTime’ was parsed to extract additional features such as year and hour, which were considered during modeling.

Model Development:

• Logistic Regression: I implemented a logistic regression model using statsmodels.api, which was trained on the processed features, including the TF-IDF scores and additional categorical variables.

Performance and Evaluation

The logistic regression model demonstrated robust performance with in-sample and out-of-sample accuracy rates of 91.7% and 90.94%, respectively. These results underscore the model’s effectiveness in handling and predicting complex, text-based data. Model performance was validated using confusion matrices and accuracy metrics provided by sklearn.metrics.

Implementation Details

• BERT Embeddings: Although BERT embeddings were considered, I employed dummy embeddings for simplification. Future iterations of this project could explore the integration of BERT for potentially enhanced semantic analysis.
• Text Preprocessing: Text data was meticulously cleaned and preprocessed, removing stopwords, punctuations, and numbers, and applying stemming and lemmatization to distill text to its most informative essence.

Conclusion

This project illustrates the power of combining traditional machine learning models with advanced NLP techniques to extract meaningful insights from textual data in healthcare settings. By accurately determining patient sentiments, healthcare providers can better understand patient needs, improve service quality, and enhance overall patient satisfaction.

I am eager to further refine this model and explore additional applications of machine learning and NLP in healthcare and other sectors. Stay tuned for future updates on this exciting intersection of data science and healthcare!