Enhancing Patient Experience Using Machine Learning and Deep Learning Based Sentiment Analysis

In the modern world of healthcare, the quality of care is no longer measured solely by medical outcomes or the speed of a diagnosis. Increasingly, the experience of the patient themselves has become a central metric for success. When a person visits a hospital, they leave behind a trail of thoughts, complaints, and praise, often shared in unstructured text on websites, forums, and social media. These thousands of individual stories contain a wealth of information about what is working and what is failing, but they are too vast and messy for human staff to read and analyze one by one. To make sense of this digital noise, researchers turn to a field of computer science known as sentiment analysis. This is the process of teaching machines to read text and understand the emotion behind the words, distinguishing between joy, anger, and indifference. By automating this task, healthcare providers can move from guessing what patients feel to knowing exactly what they feel, allowing them to fix specific problems and tailor their services to real human needs.

A team of researchers from several Indian universities set out to apply this technology to the complex reality of patient care. They gathered a massive collection of approximately 10,000 written reviews from 21 different hospitals across various cities in India. These reviews were a mix of positive feedback, negative complaints, and neutral observations. The goal was not just to count how many people were happy or unhappy, but to build a sophisticated system that could understand the deep meaning of these words and uncover the hidden topics that drove those feelings. To do this, the researchers tested a wide array of computer models, ranging from traditional statistical methods to advanced artificial intelligence systems designed to mimic the way the human brain processes language.

The process began with cleaning the raw data, a necessary step to prepare the text for analysis. The researchers stripped away punctuation, converted all letters to lowercase, and removed common words like "the" or "and" that carry no emotional weight. They also standardized the language, ensuring that variations like "running" and "ran" were treated as the same concept. Once the text was organized, they fed it into several different types of algorithms. First, they used established machine learning techniques, including Random Forest, Decision Trees, and a powerful tool called XGBoost. These models worked by looking at the frequency and importance of specific words to guess the sentiment of a review. Among these, the XGBoost model proved to be the most effective, correctly categorizing the reviews with an accuracy of nearly 90 percent. It was particularly good at identifying the nuances of positive feedback, though it faced slightly more difficulty with the more complex negative and neutral categories.

To go deeper, the team employed a second layer of technology known as deep learning. These models are designed to understand the context and sequence of words, much like how a person understands that the meaning of a sentence changes based on the order of its words. They tested four different architectures: Simple RNN, LSTM, GRU, and Bidirectional LSTM. These systems were trained using two different methods of representing words as numbers, known as Word2Vec and GloVe, which allow the computer to understand that certain words are semantically related. The results showed that the deep learning models significantly outperformed the traditional ones. The best performer was the Bidirectional LSTM model, which achieved an accuracy of over 96 percent. This model was unique because it read the text in both directions, from start to finish and from finish to start, allowing it to capture the full context of a sentence more effectively than models that read in only one direction.

While the high accuracy of these models was impressive, the researchers wanted to know more than just whether a review was good or bad. They wanted to know what the reviews were actually about. To uncover these hidden themes, they used a technique called topic modeling. This method groups words that frequently appear together to reveal the underlying subjects of the conversation. When they applied this to the negative, positive, and neutral reviews, the model identified specific clusters of words characteristic of each sentiment category. The researchers selected the top contributing words for each topic to describe the key themes occurring within the different sentiments, allowing them to compare the subjects that drove positive feelings against those that caused dissatisfaction. This ability to separate the emotional tone from the specific subject matter allowed the researchers to see exactly which aspects of hospital care were driving patient satisfaction or dissatisfaction.

The study also explored the use of a very advanced model called BERT, which is based on a pre-trained system that has already learned a vast amount of language from the internet. The researchers explicitly applied this Transformer model for comparison purposes against the other algorithms tested, detailing its architecture and training setup to benchmark its performance. While BERT is known for its ability to understand complex context, the study focused its reported performance metrics on the custom-built deep learning models and traditional machine learning classifiers. The researchers concluded that by combining these powerful classification tools with topic modeling, hospitals can gain a comprehensive view of their performance. They can move beyond simple satisfaction scores to understand the specific reasons behind them, whether it is the cost of care, the attitude of a receptionist, or the quality of the medical treatment.

This work demonstrates that the vast, unstructured voices of patients can be transformed into clear, actionable insights. By using a combination of machine learning to sort the emotions and topic modeling to identify the causes, healthcare providers can pinpoint exactly where their services need improvement. The findings suggest that the future of patient experience lies in listening to the data not just as a number, but as a collection of stories that, when read correctly, reveal the path to better care. The researchers noted that while their current work focused on online reviews, the same methods could be applied to feedback gathered through phone surveys or social media, further expanding the reach of this understanding. Ultimately, the study provides a blueprint for how technology can bridge the gap between patient experience and hospital management, ensuring that the voices of those receiving care are heard and understood.