International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 11 | Nov 2025
p-ISSN: 2395-0072
www.irjet.net
COMPARATIVE EVALUATION OF MACHINE LEARNING AND DEEP LEARNING APPROACHES FOR SENTIMENT CLASSIFICATION IN PATIENT DRUG REVIEWS Nasrullah Makhdom1, H N Verma2, Arun Kumar Yadav3 1M.Tech. Student, Department of CSE, ITM University Gwalior, Madhya Pradesh, India
2Associate Professor, Department of CSE, ITM University Gwalior, Madhya Pradesh, India 3Professor, Department of CSE, Sharda University Agra, Uttar Pradesh, India -------------------------------------------------------------------------***-----------------------------------------------------------------------Abstract: The exponential growth of user-generated health content on online platforms has introduced new opportunities
for extracting valuable insights from patient drug reviews. Sentiment analysis enables automated assessment of patient perceptions toward drugs, side effects, and treatment efficacy, supporting pharmacovigilance and clinical decision-making. Traditional lexicon-based and classical machine-learning (ML) models have been widely applied in healthcare text mining; however, their limited contextual understanding and reliance on handcrafted features constrain performance on complex medical narratives. This study presents a comparative evaluation of ML and deep-learning (DL) models for sentiment classification of patient drug reviews using the UCI Drug Review dataset. A collection of classical machine-learning frameworks, including Support Vector Machines (SVM), Naive Bayes classifiers, Logistic Regression, and random forests - was compared to the latest deep-learning systems. Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and the transformer-based Bidirectional Encoder Representations from Transformers (BERT) model were in the deep-learning category. Text data were preprocessed, vectorized through TF–IDF and word embeddings, and evaluated using accuracy, precision, recall, and F1-score metrics. Results revealed that DL models significantly outperformed traditional ML approaches, with BERT achieving the highest accuracy of 91.4% and F1-score of 0.90. The findings confirm that transformerbased models capture nuanced contextual and semantic information in patient language, leading to more reliable sentiment classification. This study contributes to healthcare informatics by identifying the most effective computational strategies for analyzing patient feedback, facilitating better pharmacovigilance, patient experience evaluation, and informed medical decision-making. Keywords: sentiment analysis, machine learning, deep learning, BERT, patient drug reviews, pharmacovigilance
1. INTRODUCTION The rapid expansion of digital health technologies has changed how people share their medical experiences and opinions. Patients today actively use online health platforms such as Drugs.com, WebMD, and various discussion forums to describe how they feel about specific treatments, their effectiveness, and possible side effects [1,2]. These platforms host millions of user-generated comments and reviews, which together form a massive source of real-world evidence. The information contained in these reviews helps researchers and healthcare professionals understand how people respond to medications in their daily lives. Such insights are highly valuable for pharmacovigilance, post-market drug safety monitoring, and for improving personalized medical treatment plans [3,4]. However, while these data sources are rich in information, they are also extremely unstructured and diverse. Reviews often include slang, abbreviations, mixed emotions, and complex sentence structures, making manual analysis almost impossible at scale [5]. To address this problem, researchers rely on Natural Language Processing (NLP) techniques, which can automatically process, interpret, and extract patterns from large amounts of text. Among these techniques, sentiment analysis is one of the most effective tools for identifying emotions or attitudes expressed in text [6]. Sentiment analysis allows classification of text as positive, negative, or neutral, helping to measure overall public or patient perception. In the field of healthcare, this technology plays a critical role by enabling automated understanding of patient emotions and satisfaction toward drugs, hospitals, or medical devices [7]. For example, if a large number of patients describe positive experiences about a new painkiller, the system can highlight this feedback for researchers or pharmaceutical companies. Conversely, if negative comments dominate, it could serve as an early warning signal for potential drug safety concerns. Early versions of sentiment analysis mainly depended on lexicon-based methods, such as SentiWordNet and VADER, which assign polarity scores to words based on pre-defined dictionaries [8,9]. Although such models are simple, transparent, and
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 959