Early Recognition System for Adverse Drug Effects Using NLP

Model

Narmadha R. P., Balaselsiya J., Keerthana B. and Rakiniya K.

Department of Artificial Intelligence and Data Science, KIT‑Kalaignarkarunanidhi Institute of Technology, Coimbatore,

Tamil Nadu, India

Keywords: Drug Identification, Adverse Effects, Natural Language Processing, Toxicity Detection, User Feedback,

Medication Management.

Abstract: This work focuses on creating an intelligent system for improving medication therapy management in

patients, with correct identification of drugs and encouragement of safer medication use. The system works

by having the patient enter a computer screen, under which it takes a picture of a drug tablet and identifies it

through sophisticated image processing algorithms that examine the shape, colour, and distinctive markings

of the tablet. After being successfully identified, the system searches a wide drug database for retrieving

essential information such as the drug's indications, chemical structure, potential side effects,

contraindications, and the potential interaction with other medications. Besides identification of drugs, the

system further employs Natural Language Processing (NLP) in order to review medical reports and patient

histories contained in the system. This centre assists in identifying abnormal patterns or repeated signs of

drug-related toxicity and allows for early intervention to avoid harmful health complications. For convenience

in use by people of all lifestyles, the system gives feedback in the form of verbal response, breaking down

complex medical jargon into easy comprehension. By combining image processing, database management,

and NLP, this system provides an end-to-end solution to medication management with the goal of minimizing

the risks involved in off-label drug use and enhancing patient safety.

1 INTRODUCTION

Adverse Drug Reactions (ADRs) and drug-induced

toxicity are key threats to the safety of patients and

rank high on the list of healthcare challenges. In spite

of strict clinical trials and regulatory processes,

unexpected adverse effects tend to emerge only after

the drugs are in wide use. Traditional detection

systems, e.g., post-marketing surveillance and

voluntary reporting, are usually slow and passive,

leading to delayed interventions. This project sets out

to adopt an innovative strategy that utilizes Natural

Language Processing (NLP) in processing

unstructured medical data like clinical notes,

electronic health records, and patient feedback in the

detection of ADRs at an early stage. Combining

state-of-the-art machine learning with a voice-

enabled assistant, the system not only flags potential

adverse reactions but also supplies real-time, easy-to-

comprehend feedback to clinicians for better

decision-making and more safe medication use.

2 LITERATURE STUDY

State-of-the-art transformer models like BERT have

shown excellent performance in identifying adverse

drug reactions (ADRs) from clinical texts. Such

models are very efficient at learning intricate

language patterns and medical terminology and can

be used to analyse unstructured data such as patient

records and clinical notes.(Siyun Yang &Supratik

Kar 2023).

Convolutional Neural Networks (CNNs) and

Recurrent Neural Networks (RNNs) have been used

with great success on patient reviews and electronic

health records (EHRs). The combined method

improves detection accuracy, although it is

associated with challenges such as high

computational requirements and risks of overfitting,

necessitating careful optimization of the

model.(Francisca Udegbe, et, al 2024).

The drug pills can be classified accurately by

using image recognition methods, and detection of

medicines is easy with high accuracy. With the help

P., N. R., J., B., B., K. and K., R.

Early Recognition System for Adverse Drug Effects Using NLP Model.

DOI: 10.5220/0013878100004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 2, pages

105-110

ISBN: 978-989-758-777-1

105

of Natural Language Processing (NLP) in toxicity

evaluation, the method is such that the end-to-end

system can deal with visual as well as text inputs,

which are suitable with the aim of early ADR

detection.(D. Mohanapriya, et, al, 2024).

Knowledge graph-based systems can associate

drug entities with their possible adverse effects and

increase detection rates. The systems give a

structured view of drug-drug interactions, side

effects, and related biomedical concepts, which

improves the interpretability of models and decision-

making accuracy.(Anu Amorim, et, al, 2024)

Machine learning algorithms are superior to rule-

based systems in detecting ADR, especially detecting

latent patterns in clinical text. Rule-based systems are

more interpretable, reflecting the accuracy-

explainability trade-off, an important consideration

in healthcare applications.(Xinxin Qi, et, al, 2024).

Long Short-Term Memory (LSTM) networks are

appropriate for clinical narrative mining since they

can understand long-range dependencies in medical

vocabulary. These models are particularly good at

identifying rare ADRs that cannot be easily identified

by simpler algorithms.(Beichang Liu, et, al, 2023).

End-to-end NLP workflows that combine named

entity recognition (NER) and sentiment analysis

make it possible for real-time monitoring of adverse

drug events. These systems can provide real-time,

actionable insights to healthcare professionals,

averting possible harm to patients.(Alexander

Tropsha, et, al 2023).

Combining image and text data through multi-

modal learning enhances ADR detection. Learning

image embeddings and text features jointly improves

model accuracy, serving as a strong solution for drug

toxicity and adverse reaction identification from

different data sources.(OladapoOyebode& Rita Orji

2023).

Pre-trained biomedical NLP models can be

transferred to ADR detection, allowing models to

generalize well to novel drugs with small amounts of

training data. This approach solves the problem of

limited data and accelerates the creation of reliable

ADR monitoring systems.(Jianxiang Wei, 2023)

NLP-powered voice-based systems can query

medical databases and provide real-time data about

drug safety to healthcare professionals. Natural

language queries enhance user interaction and make

decision-making easier, as it is easy to assess

potential drug risks through natural

language inputs(Lalitkumar Vora, et, al, 2023).

3 FINDINGS FROM THE

LITERATURE SURVEY

High Computational Cost and Complexity: Deep

learning algorithms, though capable, are

computationally intensive to train and implement.

Processing high amounts of unstructured medical

data, including patient reviews and clinical notes, is

time-consuming. Model architecture optimization or

cloud computing can be utilized to balance accuracy

with efficiency and make real-time ADR detection

more practical.

Sufficient Database Integration Requirement;

Combination of NLP models with trustworthy drug

databases, such as DrugBank, significantly enhances

ADR detection accuracy. A linked database offers the

system new drug profiles, established side effects,

and toxicity data that enable the model to provide

timely and accurate information. The combination

increases the usability of the model in real-world

applications by providing health professionals with

complete and updated information on drug safety.

4 DISADVANTAGES OF

CURRENT ALGORITHM

Limited Interpretability: Complex models are

"black boxes," whereby it is difficult for healthcare

workers to understand and have faith in the

predictions of the system.

Overfitting on Small Datasets: Models tend to have

difficulties with infrequent ADR events, memorizing

noise rather than informative patterns, which

constrains generalizability to novel data.

High Computational Complexity: Deep learning

models consume enormous computational resources

and time, thus rendering real-time ADR detection

challenging without specialized hardware.

Failure to Accommodate New or Unusual Drugs:

Algorithms are weak in the scenario of recently

released or rarely prescribed medications, especially

without continuous learning or updated databases.

Conclusion of Findings: Conclusion of the

Literature Survey early detection of drug adverse

reactions, the results indicate that approximately 50%

of the studies employ machine learning and deep

learning models in the detection of ADRs with non-

homogeneous accuracy between 50% and

90%despite being effective, the models are plagued

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

106

with drawbacks like high computational complexity,

overfitting, and lack of interpretability. The system

under development circumvents these by blending

models like Logistic Regression, Decision Trees, and

CNN with an NLP voice assistant. This method

increases precision, raises patient safety, and

facilitates quicker, better-qualified decision-making

for practitioners.

5 EXISTING SYSTEM

Currently used ADR detection and drug-induced

toxicity detection systems heavily rely on machine

learning and artificial intelligence approaches.

Decision Trees, Support Vector Machines (SVM), k-

Nearest Neighbors (KNN), and ensemble models like

Random Forests are among the widely utilized

algorithms for structured medical data

analysis.Image processing models are also used to

detect drugs by analysing visual attributes such as

shape, colour, and imprint.

5.1 Performance and Limitations

Current algorithms demonstrate accuracy between

50% and 90%, they come with significant limitations

that impact their reliability and effectiveness.

Computational Costs: High-performance models,

such as transformers and convolutional networks, are

computationally intensive, which can restrict

deployment in real-world applications.

Lack of Real-Time Feedback: The majority of

systems work with past data, and hence it is difficult

to provide instant notifications to the medical

professionals in case of critical cases.

Data Imbalance Issues:ADRs are relatively rare,

resulting in class imbalances making it difficult for

models to recognize less common but

dangerous side effects.

6 PROPOSED SYSTEM

The system to be proposed identifies the drawbacks

of current adverse drug reaction (ADR) detection

techniques and endleavors to overcome them by

combining Natural Language Processing (NLP),

machine learning, and voice-based technology. This

system increases the ability of early detection,

enhances the interaction with the user, and offers

real-time information to medical practitioners.

NLP-Driven Text Analysis: The system utilizes

sophisticated NLP models to process unstructured

medical information, including clinical notes,

electronic health records (EHRs), and patient

reviews. This enables the detection of latent patterns

and linguistic signals that suggest possible ADRs or

drug-induced toxicity.

Machine Learning Algorithms: Logistic

Regression, Decision Trees, Random Forest, SVM,

and deep models such as CNNs and LSTMs are

utilized to increase the detection rate. Models are

trained to identify drugs, classify the toxicity level,

and forecast side effects.

Figure 1: Drug Toxicity Detection Flow.

Image Processing for Drug Identification: The

system has an identification module that is drug

image processing based. The users are able to scan a

drug tablet, and the system reads the image and

identifies the drug type and matches it to a database

Early Recognition System for Adverse Drug Effects Using NLP Model

107

for further analysis.

Voice-Enabled Assistant: To facilitate better

access, the system comes equipped with a voice

assistant based on NLP. The medical professionals

can pose questions to the system in natural language,

e.g., querying about possible side effects or recent

ADR reports. The assistant offers voice answers,

offering instant and convenient access to

information.

Database Integration: The system interfaces with

drug databases (such as DrugBank) to fetch current

drug profiles, established side effects, and toxicity

data. The model is thus ensured to operate using the

most recent pharmaceutical data for increased

accuracy and reliability. Figure 1 shows the drug

toxicity detection flow.

7 WORK FLOW DIAGRAMS

The proposed system process is intended to simplify

the identification of drug-induced toxicity and

adverse drug reactions (ADRs) using image

processing, machine learning, and Natural Language

Processing (NLP). Below is an overview of the step-

by-step process of the system (figure 2):

Drug Image Input: The system begins by the user

either scanning or uploading an image of a drug tablet

using a camera or mobile device. Image data is

obtained and preprocessed with libraries such as

OpenCV for quality improvement and accurate drug

identification.

Drug Identification via Image Processing: The

preprocessed image is input into a Convolutional

Neural Network (CNN) to identify the drug by its

shape, color, and imprint. The drug name identified

is taken and forwarded to the next level for further

processing.

Query to the Drug Database: The discovered drug

is compared with an extensive drug database (e.g.,

DrugBank). Relevant information, including drug

composition, usage, potential side effects, and known

ADRs, is retrieved.

Text Data Analysis with NLP: Unstructured

medical data, such as clinical notes, EHRs, and

patient reviews, is processed using NLP models (like

BERT or BioBERT). The NLP model scans for

adverse effect mentions, linguistic patterns, and

correlations between drug names and toxicity reports.

Machine Learning-Based ADR Prediction: The

extracted text is then fed into machine learning

models (such as Random Forest, SVM, LSTM) to

foresee potential ADR risks. The model gives a risk

score or prediction of whether the drug is associated

with any side effects.

Figure 2: Drug Identification and Feedback System

Workflow.

Voice-Enabled Feedback: It has a voice assistant

that reads out the results in sound feedback.

Physicians or consumers can ask questions like

"What are the side effects of this drug?" and receive

instant, natural language responses.

8 EXPECTED OUTCOMES

The proposed system is expected to significantly

enhance drug safety surveillance using the

application of image processing, Natural Language

Processing (NLP), and machine learning for on-time

detection of drug-induced toxicity and adverse drug

reactions (ADRs). By integrating a combination of

different technologies, the system will provide rich,

real-time data to health professionals, allowing them

to make decisions and prevent patient harm.

Accurate Drug Identification: The module for

processing images will reliably identify drug pills

with high accuracy and minimize chances of

medication mistakes (figure 3).

Early Detection of ADRs: Adverse reactions

detection via NLP will effectively spot issues from

patient testimonials, clinician notes, and electronic

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

108

patient data, ensuring intervention at the early stage.

Figure 3: Tablet Toxicity Finder – User Interface

Dashboard.

Table 1: Ml Algorithms: Accuracy & Use Cases.

Algorithm/Technology

Accuracy

Advantages

CNN (Convolutional

Neural Networks)

90%

CNNs are

good at

processing

sophisticated

images and

can learn

automatically

sophisticated

features.

KNN (K-Nearest

Neighbors)

90%

KNN is easy

to use and

works well

with small

datasets.

Logistic Regression,

Naïve Bayes, Random

Forest, Decision Tree,

Support Vector

Machine

>95%

Prediction of

toxicity

Linear Regression,

Random Forest

>92%

Recommended

Dosage and

Usage

Guidelines

ANN (Artificial

Neural Network)

90%

Versatile

ANNs can

efficiently

manage

almost any

audio-related

task.

High Accuracy Predictions: Medications will be

determined based on risk profiles with high accuracy

rates of predicting potential ADRs and toxic levels by

drug models.

The voice assistant: The voice assistant will provide

immediate verbal feedback to healthcare providers,

whereas the system delivers real-time warnings for

high-risk medications to allow timely

clinical intervention. Table 1 shows the ML

Algorithms: Accuracy & Use Cases.Figure 4 shows

the Voice-Driven Drug Safety Interface. Figure 5

shows the performance comparison between existing

and proposed systems.

Figure 4: Voice-Driven Drug Safety Interface.

Figure 5: Performance Comparison Between Existing and

Proposed Systems.

9 CONCLUSIONS

The system of early detection of ADRs and drug-

induced toxicity proposed in this work integrates

image processing, NLP, and machine learning to

improve patient safety and assist clinical decision-

making. With automated drug identification, medical

text analysis, and real-time feedback via a voice

assistant, the system presents an active method for

monitoring drug safety. This approach overcomes the

Early Recognition System for Adverse Drug Effects Using NLP Model

109

shortcomings of the conventional methods,

minimizes medication-related harm risk, and works

towards safer and more effective healthcare

practices.

REFERENCES

Alexander Tropsha, OlexandeIsayev, Gisbert Schneider &

Artem Cherkasov (2023) – This research highlights

how deep learning and QSAR modeling are changing

drug discovery, making the process smarter and more

precise. Found in the Journal of Discovery.

Anu Amorim, Luiz Piochi, Ana Gaspar, Antonio Preto,

Nicia Ferreira & Irina Moreira (2024) – A glimpse of

how computational models can anticipate drug toxicity

at an early stage in the development process, making

drugs safer even before they reach patients. It is in the

Journal of Chemical Research in Toxicology.

Beichang Liu, Guoqing Cai, Jili Qian, Tiambo Song &

Quan Zhang (2023) – A team of scientists developed a

machine learning-based system for the identification of

drugs more accurately and efficiently. Published in the

International Journal of Computer Science and

Information Technology.

D. Mohanapriya, Jyothi Chepur, Deepali Hirolikar, K.

Sundareswari, Kiran Pokkuluri& R. Subbulakshmi

(2024) – This research centers on the application of

machine learning to classify patient drug reviews and

identify adverse drug reactions. Published in the

Journal of Measurement Sensors.

Francisca Udegbe, OgochukwuEbulue, Charles

Ebulue&ChukwunonsoEkesiobi (2024) – A dive into

using machine learning in drug discovery, its

advantages, and the challenges to researchers. Featured

in the Journal of Computer Science & IT Research.

Jianxiang Wei, Tianling Hu, Jimin Dai, Ziren Wang, Pu

Han &Weidong Huang (2023) – This article discusses

how AI is able to identify mentions of adverse drug

reactions from text through the use of deep learning and

NLP methods.

Lalitkumar Vora, Amol Gholap, Keshava Jetha, Raghu

Thakur, Hetvi Solanki & Vivek Chavda (2023) –

Review of the new advancements in drug research, i.e.,

how AI is making drug side effects more identifiable

and understandable. Published in the Journal of

Pharmaceutics.

OladapoOyebode& Rita Orji (2023) – A fascinating study

on how AI and natural language processing (NLP) can

analyze patient reviews on social media to detect

possible drug reactions. It is in the Journal of Health

Informatics.

Siyun Yang &Supratik Kar (2023) – This article explores

how AI and machine learning can be used to detect

adverse drug reactions and toxicity at an early stage,

leading to improved patient safety. It is in the Journal

of Artificial Intelligence Chemistry.

Xinxin Qi, Yuanchun Zhao, Zhuang Qi, Siyn Hou &Jiajia

Chen (2024) – This article discusses the application of

machine learning to drug discovery and encompasses

its potential, practical applications, and some of the

challenges that researchers must face. It is in Journal of

Moleculars.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

110