Institut Riset dan Publikasi Indonesia (IRPI)

MALCOM: Indonesian Journal of Machine Learning and Computer Science
Journal Homepage: https://journal.irpi.or.id/index.php/malcom
Vol. 5 Iss. 3 July 2025, pp: 738-745
ISSN(P): 2797-2313 | ISSN(E): 2775-8575

Smart Prescription Reader: Enhancing Accuracy in
Medical Prescriptions
Ragil Yulianto
Faculty of Computer Science, Universitas Nusa Mandiri, Indonesia
E-Mail: 14240007@nusamandiri.ac.id
Received Jan 25th 2025; Revised Mar 22th 2025; Accepted Apr 12th 2025; Available Online Jun 19th 2025, Published Jun 22th 2025
Corresponding Author: Ragil Yulianto
Copyright ©2025 by Authors, Published by Institut Riset dan Publikasi Indonesia (IRPI)

Abstract
Reading a doctor's handwritten prescription is a challenge faced by most patients and some pharmacists, which in some
cases can lead to negative consequences due to misinterpretation of the prescription. The "Doctor's Handwritten
Prescription BD Dataset" on Kaggle contains segmented images of handwritten prescription words from BD
(Bangladesh) doctors. This dataset, intended for machine learning applications, includes 4,680 individual words
segmented from prescription images. This study introduces a Handwriting Recognition System using Convolutional
Neural Network (CNN) developed to identify text in prescription images written by doctors and convert the cursive
handwriting into readable text. Two models were evaluated in this study: CNN and MobileNet. Based on the experiments,
MobileNet showed better results compared to CNN alone. From the dataset of 4,680 words, 3,120 were used for training,
780 for testing, and 780 for validation. The study achieved a training accuracy of 97%, a testing accuracy of 88%, and a
validation accuracy of 83%. The developed model was successfully implemented in a web application.
Keyword: Convolutional Neural Network, Machine Learning, MobileNet, Prescription Images

1.

INTRODUCTION
Medication errors are a significant public health issue, with an estimated 44,000 to 98,000 emergency
hospital patients dying each year due to such errors, making them one of the leading causes of death in the
United States [1]. Some studies even suggest higher figures, potentially making medical errors the third
leading cause of death in the country [2], [3]. The healthcare system faces significant challenges in patient
safety and quality improvement. Despite extensive efforts, sustainable progress in reducing preventable harm
has been elusive [4]. Systemic weaknesses, including lack of standardization, inadequate staff wellbeing, and
organizational culture issues, contribute to persistent safety concerns [5], [6].
These systemic weaknesses manifest in various types of medication errors, compromising patient
safety. A report from King’s College Hospital NHS Foundation Trust, London, highlighted that, out of
12,006 reported medication incidents, 1,568 involved wrong-dose errors, with 44.8% due to prescribing
errors, 14.2% due to dispensing errors, and 41% due to administration errors [7]. The findings further noted
that overdoses accounted for 59% of wrong-dose cases, emphasizing the critical nature of accurate
prescription handling [7]. However, this report is merely a snapshot and cannot be used to generalize facts, as
the data only covers one hospital and may not reflect the situation elsewhere. In the context of the King’s
College Hospital NHS Foundation Trust report, most errors occurred due to systemic failures rather than
individual negligence[8], [9]. Individual negligence, such as errors in writing prescriptions or misreading
dosages, can indeed occur, but they are often influenced by systemic factors such as high workload, lack of
training, or non-standard procedures. These human factors contribute to individual negligence, which can
lead to medication errors. A systems-based approach, including pharmacist involvement in prescription
review, can significantly reduce error rates [10].
In addition to systemic weaknesses, Prescription errors often stem from human factors such as fatigue,
high workload, and inadequate system support. Handwritten prescriptions often present challenges in
legibility and completeness, leading to potential medication errors and the need for clarification calls to
prescribers. Illegible handwriting, lack of pharmacological knowledge, distractions, performance deficits, and
failure to follow protocols further exacerbate the risk [1]. To address this issue, interventions such as
improved education and training, automated systems, and periodic audits are recommended [11], [12].

DOI: https://doi.org/10.57152/malcom.v5i3.1934

738

ISSN(P): 2797-2313 | ISSN(E): 2775-8575
Technological advancements in handwritten prescription recognition offer a promising avenue to
address these challenges. Automated tools using Optical Character Recognition (OCR) and artificial
intelligence (AI) have been shown to streamline healthcare processes by converting handwritten text into
readable and verifiable digital information [11]. These tools can extract data from medical forms, synthesize
research evidence, and assist in clinical documentation [12], [13], [14], [15], [16]. However, many existing
solutions face limitations. Traditional OCR methods often struggle with the variability and complexity of
handwritten medical scripts. Additionally, most approaches rely on structured, predefined forms that may not
be adaptable to the unstructured and diverse nature of real-world handwritten prescriptions [17].
To bridge these gaps, this study proposes the development of a robust, AI-based handwritten
prescription reader. Leveraging CNNs for image classification and integrated with a user-friendly web
application built using Flask, this project aims to improve the accuracy of prescription interpretation. By
automating the reading process, this system seeks to enhance medical accuracy, minimize human error, and
contribute to overall patient safety.
The objectives of this study are to develop, train, and evaluate a CNN-based model that can accurately
read handwritten prescriptions and to create a web interface that facilitates seamless user interaction with the
AI. This approach not only addresses the technological shortcomings of traditional OCR systems but also
provides a scalable solution tailored for healthcare applications. With this approach, it is hoped that a more
reliable and efficient system for reading handwritten prescriptions can be created, thereby making a
significant contribution to improving the quality of healthcare services.
2.
2.1

MATERIALS AND METHOD
Data Pipeline
We used the publicly available "Doctor’s Handwritten Prescription BD" dataset from Kaggle for this
paper [35]. We registered on the website and downloaded images of handwritten lines along with their
annotations, which were available in CSV format. The total image data consists of 4680 images, with 60%
used for training, 20% for validation, and 20% for testing, ensuring stratified data to maintain an equal
amount of data in each class. The training directory contains 60% of the total data, the testing directory
contains 20%, and the validation directory also contains 20%, each including Excel and CSV files indexing
word names. We loaded and examined the dataset by reading the CSV file containing labels for the images
and ensuring the necessary columns were present in the CSV file.
The images were processed by loading them from the extracted folder, resizing them to 64x64 pixels,
converting them to arrays, and normalizing the pixel values. This was done for several important reasons in
data processing for machine learning, including ensuring consistent image sizes, as machine learning models,
especially neural networks, require input with consistent dimensions. By resizing all images to 64x64 pixels,
we ensure that each image has the same dimensions, allowing the model to process them efficiently.
Secondly, reducing the image size to 64x64 pixels decreases the total number of pixels the model needs to
process, which helps reduce computational complexity and speeds up the model training process without
losing important information from the images. Thirdly, converting images to arrays transforms them from a
visual format to a numerical format that the model can understand, with each pixel represented by a
numerical value indicating color intensity. Finally, normalizing the pixel values by dividing them by 255 to
range between 0 and 1 helps improve model performance, as neural networks work better with input values
that have a uniform and controlled range. Normalization also aids in faster convergence during training.
Labels were converted to numerical form using LabelEncoder and then transformed into a format
usable by the model (one-hot encoding). This was done because machine learning models, especially neural
networks, cannot work directly with categorical data (such as drug names in text form). They require data in
numerical form to perform mathematical calculations. If categorical labels were directly converted to
numbers (e.g., "Paracetamol" to 1, "Ibuprofen" to 2), the model might assume an ordinal relationship
between the labels, which is incorrect. One-hot encoding avoids this issue by converting each category into a
separate binary vector, treating all categories as equal. One-hot encoding produces a binary representation
where each class is represented by a vector with a value of 1 at the index corresponding to that class and 0 at
other indices. This is very useful for classification tasks as it allows the model to predict probabilities for
each class independently. Additionally, many loss functions used in classification, such
as categorical_crossentropy, require labels in one-hot encoded form to correctly calculate prediction errors.
2.2

Model Design
CNN are deep learning algorithms used for image analysis. A neural network qualifies as a CNN even
with just one convolutional layer. CNNs are effective in extracting deep semantic features from images and
are widely used in computer vision for object classification and identification. The convolution operation
makes CNNs robust to local variations and image modifications, allowing them to recognize objects
regardless of position, size, or orientation. CNNs play a significant role in image analysis, especially in the
medical field where data sensitivity is high. Researchers use CNNs for feature extraction, transfer learning,
MALCOM - Vol. 5 Iss. 3 July 2025, pp: 738-745

739

MALCOM-05(03): 738-745
and abnormality detection. Figure 1 shows a CNN’s basic structure, including convolution, pooling, and
dense layers. Every CNN has these essential components with some transformations or adaptations.

Figure 1. The basic architecture of CNN
The convolutional layer is a key component in CNNs used for processing image data. This layer
applies a kernel (filter) to the input image to extract important features. Features in this context refer to
specific patterns or characteristics in the image, such as edges, textures, or shapes, which aid in object
recognition or image classification. The filter size can vary depending on the architecture. Kernels can be
applied both pointwise and depthwise. Pointwise convolution uses a 1x1 filter applied to each point in the
image, allowing each input channel to interact with each output channel. Depthwise convolution applies a
filter to each input channel separately, which are then combined to produce the final feature map. After
applying the kernel to the image, a feature map is obtained that can be further processed.
The feature map size is larger than the input image, so another pooling layer is added to the CNN to
reduce the feature map size. Pooling layers function to reduce the dimensions of the feature map, which helps
decrease overall computational time and cost. The most popular pooling technique is called max pooling. As
the name implies, the max pooling procedure selects the maximum value from a patch of the feature map.
Dense layers have neurons that are densely connected, meaning each neuron in this layer is connected
to every neuron in the previous layer. This is also known as a fully connected layer. Its function is to receive
input (learned features) from the previous layers and produce the output. The dense layer is crucial in the
final stage of the neural network as it completes the data classification task, transforming the learned features
into the final predictions.
A CNN can solve problems by either building a network from scratch or retraining an alreadydeveloped model for new types of problems. This study uses a basic CNN model and a model with
MobileNet. The basic CNN model provides a strong foundation for feature extraction, while MobileNet,
being a lightweight model, is designed to work efficiently on devices with limited resources. The
combination of these two models allows for accurate and efficient image classification under various
conditions and on different devices.
This study focuses on solving image classification problems and categorizing images into specific
classes. The procedure adopted for this purpose uses well-developed and well-trained models on diverse
image datasets for classification. The methodology involves: (i) collecting and preparing the image dataset;
(ii) normalizing and processing the images; (iii) using basic CNN and MobileNet models to classify the
images; and (iv) evaluating the results from both models.
2.2.1 Basic CNN
The CNN model used consists of several convolutional layers followed by dense and dropout layers.
We use the Rectified Linear Unit (ReLU) activation function after each convolutional layer to introduce nonlinearity. ReLU is a commonly used activation function in neural networks that converts negative values to
zero and keeps positive values unchanged. This helps the model learn more complex patterns and speeds up
the training process. The CNN model built includes convolutional, pooling, flatten, dense, and dropout
layers. This model is implemented using the Sequential Model, which is a linear stack of neural network
layers. The Sequential Model is created by adding an input layer that specifies the input shape of 64x64
pixels with 3 color channels (RGB).
The first convolutional layer has 32 filters with a kernel size of 3x3 and uses the ReLU activation
function to introduce non-linearity. This is followed by the first pooling layer that performs max pooling with
a pool size of 2x2 to reduce the spatial dimensions of the feature maps. The second convolutional layer has
64 filters with a kernel size of 3x3 and also uses the ReLU activation function, followed by the second
pooling layer with a pool size of 2x2. The third convolutional layer has 128 filters with a kernel size of 3x3
and uses the ReLU activation function, followed by the third pooling layer with a pool size of 2x2.
The flatten layer converts the output from the last convolutional layer into a one-dimensional vector,
which is then fed into the first dense layer with 512 units and the ReLU activation function. To prevent
overfitting, a dropout layer with a rate of 0.5 is used after the first dense layer. Dropout is a regularization
technique used to reduce overfitting in neural networks. During training, dropout randomly "disables" a
number of units in the neural network layer with a certain probability (in this case, 0.5), thus preventing the

Smart Prescription Reader: Enhancing Accuracy in Medical... (Yulianto, 2025)

740

ISSN(P): 2797-2313 | ISSN(E): 2775-8575
network from becoming too reliant on specific units and improving the model's generalization ability.
Finally, the second dense layer has 78 units with the softmax activation function, which produces
probabilities for each of the 78 classes. The model is trained using preprocessed data split into training and
testing sets. After training, the model is saved and used to predict labels from test images. The learning rate
used is consistently set at 0.001 throughout the entire training process.
2.2.2 MobiliNet
MobileNet was developed by Howard et al. [18]. As the name suggests, MobileNet was created to
offer a computationally light and useful model that can be used on mobile devices. This model improves
accuracy while reducing computational latency. MobileNet employs depth-wise separable convolution,
where each channel of the input image is first subjected to a depth-wise filter before being combined by a
pointwise filter. MobileNet also uses width and resolution multipliers, two hyperparameters that help
MobileNet be smaller than conventional convolutional networks. Figure 2 illustrates how the standard
convolution filter (image a) is replaced by two layers in MobileNet V1. These layers use depthwise
convolution (image b) and pointwise convolution (image c) to form a depthwise separable filter.

Figure 2. Convolutional Mechanisms in MobileNet
This study uses MobileNet as its base, with pre-trained weights from ImageNet and without the top
layer. The model is built sequentially using the Keras Sequential API, which allows for layer-by-layer
construction. First, the GlobalAveragePooling2D layer converts the 2D feature maps into a 1D vector by
averaging each feature map. Next, a dense layer with 512 neurons and ReLU activation is added to provide
non-linearity. Following this, a dropout layer with a 50% dropout rate helps prevent overfitting by randomly
ignoring half of the neurons during training. Finally, a dense layer with the number of neurons corresponding
to the number of classes (78) and softmax activation produces probabilities for each class.
The model accepts input images with a shape of (224, 224, 3), meaning images of 224x224 pixels
with 3 color channels (RGB). It is designed to classify images into 78 different classes. The model is
compiled using the Adam optimizer with a learning rate of 0.001, the categorical_crossentropy loss function,
and accuracy as a metric. By leveraging the efficient and powerful MobileNet architecture, this model aims
to perform well in image classification tasks.
2.3

Front-End Integration
In this study, the Python Flask package is utilized. Flask is a lightweight web framework for Python
used to build web applications. In this context, Flask serves as the backend framework that handles serverside logic, routes, and API endpoints. The machine learning model developed using Keras is integrated with
Flask to process input data (prescriptions) and generate predictions. The web-based user interface (UI) allows
users, such as medical staff, to input handwritten prescriptions into the system. The UI is designed to be
intuitive and user-friendly, ensuring that users can easily upload images or enter text data. Once the data is
submitted, the UI interacts with Flask to send the data to the backend, where it is processed by the Keras
model and predictions are generated. The results are then displayed back on the UI, which can be accessed
through a browser, for the user to review.
2.4
Model Training dan Evaluation
2.4.1 CNN Model Training
The training process of the model was conducted over 50 epochs, where each epoch represents a full
cycle through the entire training dataset. The training dataset consists of 3,120 preprocessed images. This

MALCOM - Vol. 5 Iss. 3 July 2025, pp: 738-745

741

MALCOM-05(03): 738-745
dataset was divided into small batches, each containing 32 samples, allowing the model to update its weights
after processing each batch. A batch size of 32 was chosen because it is a commonly used size that balances
training speed and weight update stability. The loss function used is categorical_crossentropy, which is
suitable for multi-class classification problems. This loss function measures the difference between the
model's predictions and the actual labels, aiming to minimize this error during training. We chose
categorical_crossentropy because it is more effective for handling multi-class classification problems
compared to other loss functions like mean_squared_error, which is more suitable for regression, or
binary_crossentropy, which is used for binary classification.
The optimizer used is Adam with a learning rate of 0.001. An optimizer is an algorithm used to adjust
the neural network's attributes, such as weights and biases, to reduce errors. Adam was chosen because it
combines the advantages of two other optimization methods, AdaGrad and RMSProp, making it efficient and
capable of performing well without much parameter tuning. Adam is also known for its ability to adjust the
learning rate during training, helping to achieve faster and more stable convergence. A learning rate of 0.001
was chosen because it is a commonly used value that provides a balance between convergence speed and
training stability. A learning rate that is too high can cause the model not to converge, while a learning rate
that is too low can make training very slow. This training process aims to minimize the loss function and
improve the model's accuracy on the validation data. Validation was performed using 780 images from the
dataset, which were used to evaluate the model's performance during training.
2.4.2 Mobilinet Model Training
The training process of the model in the provided code involves several important steps. The model is
trained for 50 epochs, with a batch size of 32, meaning the model is updated after seeing 32 training samples.
The loss function used is categorical_crossentropy, which is suitable for multi-class classification problems.
The optimizer used is Adam with a learning rate of 0.001, which combines the advantages of the AdaGrad
and RMSProp optimization methods.
The model uses the MobileNet architecture as its base, which has been previously trained on the
ImageNet dataset, but without the top layers. The input to the model has dimensions of 224 x 224 x 3,
meaning the input images are resized to 224 pixels by 224 pixels with 3 color channels (red, green, and blue).
Therefore, the training data is preprocessed to match these dimensions, such as resizing the images to 224 x
224 pixels and ensuring they have 3 color channels.
After that, several additional layers are added, including GlobalAveragePooling2D, Dense with 512
units and ReLU activation, Dropout with a rate of 0.5, and Dense with the number of units corresponding to
the number of classes (78) and softmax activation. GlobalAveragePooling2D is a layer that simplifies the
data by averaging each feature map over its spatial dimensions, effectively reducing the data's
dimensionality.
2.4.3 Evaluation
After the training is completed, the model is evaluated using metrics such as accuracy, precision,
recall, and F1-score. Accuracy measures how often the model's predictions are correct overall (1). Precision
measures the proportion of true positives (correctly predicted positive cases) out of all positive predictions,
thus accounting for false positives (incorrectly predicted positive cases) (2). Recall measures the proportion
of true positives out of all actual positive cases, thus accounting for false negatives (actual positive cases that
were incorrectly predicted as negative) (3). The F1-score is the harmonic mean of precision and recall,
providing a balanced view of both metrics (4). This evaluation provides a more comprehensive understanding
of the model's performance across various aspects.
2.5

User Interface Design and UX Goals
The UI of the Prescription Detector is designed with a minimalistic layout to ensure ease of use and
avoid overwhelming users (see Figure 3). Key action areas, such as the file upload zone and the "Upload"
button, are prominently displayed for intuitive interaction. The UI supports both drag-and-drop functionality
and a file picker, catering to diverse user preferences. Feedback mechanisms are prominently integrated,
providing immediate results and confidence levels after uploading, ensuring clarity without requiring
additional navigation. Additionally, a "Help" button at the top offers quick access to documentation or FAQs,
supporting users who may need guidance during the process.
The process flow of the Prescription Detector begins with users selecting or dragging their
prescription image into the designated upload area, ensuring a user-friendly and straightforward interaction.
Once uploaded, the system validates the file's format and suitability for prediction, checking aspects such as
size and type. The validated image is then processed by a backend AI-powered model, which analysis the
prescription details and generates a prediction result. The output, including the prediction and a confidence
level, is displayed in a clear, readable format. For example, a confidence level of 0% may indicate an issue
with the uploaded file or data, prompting users to retry or seek assistance.

Smart Prescription Reader: Enhancing Accuracy in Medical... (Yulianto, 2025)

742

ISSN(P): 2797-2313 | ISSN(E): 2775-8575

Figure 3. Prescription Detector Upload Form
The system provides comparative information on whether prescribed medication meets the patient's
needs and offers alternative medications based on the patient's condition and medical history. This helps
users make informed decisions regarding treatment. The classification results are clearly displayed for easy
viewing and saving. The primary goal is to minimize medication errors and enhance healthcare quality.
3.

RESULT AND ANALYSIS
The performance of the Handwriting Recognition System was evaluated using a test dataset consisting
of 780 prescription images. Table 1 shows the performance comparison between the CNN and MobileNet
models. The CNN model achieved an accuracy of 63%, with precision, recall, and F1-score of approximately
65%, 62%, and 63% respectively. On the other hand, the MobileNet model achieved an accuracy of 83%,
with precision, recall, and F1-score of approximately 84%, 82%, and 83% respectively.
Table 1. Performance Comparison of CNN and MobileNet Models
Matric
F1
Precision
Recall
Accuracy

CNN
0,63
0,65
0,62
0,63

MobiliNet
0,83
0,84
0,82
0,83

The evaluation process was conducted through several steps. First, a dataset of prescription images
was collected and prepared for model training and testing. These images were then processed and labeled to
ensure that each image had the correct annotations. Following this, the CNN and MobileNet models were
trained using the training dataset, and their performance was tested using the test dataset. In terms of case
studies, the model showed notable success in identifying certain words. For instance, it accurately identified
the word "Aceta" from a prescription image with a high confidence score of 85%, thanks to the clear and
distinct handwriting. Similarly, the word "Amodis" was correctly recognized with a confidence score of 80%,
attributed to the relatively simple and legible handwriting style.
However, the model also encountered challenges. For example, it struggled with the word
"Azithrocin" due to the cursive and overlapping handwriting style, resulting in a low confidence score of
45% and incorrect identification. Another difficult case was the word "Baclon," which was misinterpreted as
"Baclofen" due to the similarity in character shapes and the presence of noise in the image.
The test results suggest that while MobileNet model performs well on clear and distinct handwriting,
it faces difficulties with cursive and overlapping styles. These challenges can be attributed to the variability
in handwriting styles, noise and artifacts in the prescription images, and the complexity of the model. Despite
these challenges, the study demonstrates the potential of using MobileNet for handwriting recognition in
medical prescriptions, with opportunities for further improvements to enhance accuracy and reliability.

MALCOM - Vol. 5 Iss. 3 July 2025, pp: 738-745

743

MALCOM-05(03): 738-745
4.

DISCUSSION
The Handwriting Recognition System shows potential in handling real-world variabilities, particularly
in the legibility of handwritten prescriptions. The use of CNN with MobileNet is effective in recognizing
clear text, but its performance decreases with more complex handwriting. Additionally, the model is limited
in handling multilingual prescriptions because the dataset used primarily consists of English words written by
Bangladeshi doctors. To address this, future research could involve training the model on a more diverse and
multilingual dataset, thereby enhancing its robustness and versatility.
Additionally, the system's performance on extremely illegible text remains a significant challenge.
Noise and artifacts in the prescription images can obscure the text, leading to misinterpretation. Enhancing
the pre-processing steps to better clean and segment the images could help mitigate this issue. Moreover,
incorporating advanced techniques such as attention mechanisms or transformer-based models could improve
the system's ability to focus on relevant parts of the image and better handle complex handwriting styles.
In conclusion, while the Handwriting Recognition System shows potential in converting doctors'
handwritten prescriptions into readable text, there are several areas for improvement. Addressing the
variability in handwriting styles, expanding the model's capability to handle multilingual prescriptions, and
enhancing its performance on extremely illegible text are crucial steps towards developing a more robust and
reliable system. Future research and development efforts should focus on these aspects to further enhance the
accuracy and applicability of the system in real-world scenarios.
5.

CONCLUSION AND FUTURE DIRECTIONS
This study has made significant contributions to enhancing healthcare efficiency and accuracy by
developing a Handwriting Recognition System using CNN to interpret doctors' handwritten prescriptions.
The system, particularly the model with MobileNet, demonstrated a notable improvement in recognizing and
converting cursive handwriting into readable text. With a validation accuracy of 88% and a test accuracy of
83% on the test set, the model shows promise in reducing the risk of misinterpretation of prescriptions, which
can lead to negative consequences for patients. The successful implementation of the model in a web
application further underscores its practical applicability. Despite the challenges faced with more complex
and less legible handwriting, the system's ability to correctly identify a significant portion of the test samples
highlights its potential utility in real-world healthcare settings.
For future research, several avenues can be explored to enhance the system's performance and
applicability. One promising direction is the integration of natural language processing (NLP) techniques to
not only recognize handwritten text but also understand the content of prescriptions. This could enable the
system to provide additional context and ensure the correct interpretation of medical instructions. Another
important area for future work is the collaboration with healthcare facilities to access larger and more diverse
datasets. This would allow the model to be trained on a wider variety of handwriting styles and languages,
improving its robustness and generalizability. Additionally, incorporating advanced machine learning
techniques, such as transformer-based models, could further enhance the system's ability to handle extremely
illegible text and complex handwriting patterns.
In conclusion, while the current study has laid a strong foundation for the development of a
handwriting recognition system for medical prescriptions, ongoing research and development are essential to
address its limitations and expand its capabilities. By leveraging advancements in machine learning and
collaborating with healthcare professionals, the system can be refined to provide even greater accuracy and
reliability, ultimately contributing to safer and more efficient healthcare delivery.
REFERENCES
[1]
N. Alrabadi et al., “Medication errors: a focus on nursing practice,” Journal of Pharmaceutical Health
Services Research, vol. 12, no. 1, pp. 78–86, Feb. 2021, doi: 10.1093/jphsr/rmaa025.
[2]
L. C., “New Questions on the Road to Safer Health Care,” Pediatrics, doi: 10.1542/peds.2013-0647.
[3]
H. Carole, H. M., E. S. Susan, and D. U. Beth, “Who killed patient safety?,” J Patient Saf Risk Manag,
doi: 10.1177/25160435221077778.
[4]
R. Mannion and J. Braithwaite, “False Dawns and New Horizons in Patient Safety Research and
Practice,” Int J Health Policy Manag, vol. 6, no. 12, pp. 685–689, doi: 10.15171/ijhpm.2017.115.
[5]
L. J. Donaldson, C. Lemer, D. J. Noble, F. Greaves, and M. Fletcher, “Finding the Achilles’ heel in
healthcare,” J R Soc Med, vol. 103, no. 2, pp. 40–41, doi: 10.1258/jrsm.2009.090417.
[6]
H. H. Louise, J. Judith, W. Ian, T. A., and B. O. Daryl, “Healthcare Staff Wellbeing, Burnout, and
Patient Safety: A Systematic Review,” PLoS One, doi: 10.1371/journal.pone.0159015.
[7]
G. F. Cavell and D. Mandaliya, “Magnitude of error: a review of wrong dose medication incidents
reported to a UK hospital voluntary incident reporting system,” European Journal of Hospital
Pharmacy, vol. 28, no. 5, pp. 260–265, Sep. 2021, doi: 10.1136/ejhpharm-2019-001987.
[8]
L. L., “A systems analysis approach to medical error.,” J Eval Clin Pract, doi: 10.1046/J.13652753.1997.00006.X.

Smart Prescription Reader: Enhancing Accuracy in Medical... (Yulianto, 2025)

744

ISSN(P): 2797-2313 | ISSN(E): 2775-8575
[9]
[10]
[11]

[12]

[13]

[14]

[15]

[16]
[17]

[18]

G. A. James, “A systems approach to preventing adverse drug events.,” Stud Health Technol Inform.
C. Michelle et al., “Influence of a systems-based approach to prescribing errors in a pedia tric resident
clinic.,” Acad Pediatr, doi: 10.1016/j.acap.2014.03.018.
S. Mohsin-Shaikh et al., “The impact of electronic prescribing systems on healthcare professionals’
working practices in the hospital setting: A systematic review and narrative synthesis,” BMC Health
Serv Res, vol. 19, no. 1, pp. 1–8, 2019, doi: 10.1186/s12913-019-4554-7.
J. Memon, M. Sami, R. A. Khan, and M. Uddin, “Handwritten Optical Character Recognition (OCR):
A Comprehensive Syste matic Literature Review (SLR),” IEEE Access, vol. 8, pp. 142642–142668,
doi: 10.1109/access.2020.3012542.
S. Karthikeyan, A. G. S. de Herrera, F. Doctor, and A. Mirza, “An OCR Post-Correction Approach
Using Deep Learning for Processing Med ical Reports,” IEEE Transactions on Circuits and Systems
for Video Technology, vol. 32, no. 5, pp. 2574–2581, doi: 10.1109/tcsvt.2021.3087641.
L. V Rasmussen, P. L. Peissig, C. A. McCarty, and J. Starren, “Development of an optical character
recognition pipeline for handwritt en form fields from an electronic health record,” Journal of the
American Medical Informatics Association, vol. 19, no. e1, pp. e90–e95, doi: 10.1136/amiajnl-2011000182.
N. Gal-Nadasan, V. Stoicu-Tivadar, E. Gal-Nadasan, and A. R. Dinu, “Robotic Process Automation
Based Data Extraction from Handwritten Medi cal Forms,” in Studies in Health Technology and
Informatics, IOS Press, 2023. doi: 10.3233/shti230741.
E. Coiera and S. Liu, “Evidence synthesis, digital scribes, and translational challenges for artificial
intelligence in healthcare,” Cell Rep Med, vol. 3, no. 12, p. 100860, doi: 10.1016/j.xcrm.2022.100860.
L. V. Rasmussen, P. L. Peissig, C. A. McCarty, and J. Starren, “Development of an optical character
recognition pipeline for handwritten form fields from an electronic health record,” Journal of the
American Medical Informatics Association, vol. 19, no. e1, pp. e90–e95, Jun. 2012, doi:
10.1136/amiajnl-2011-000182.
A. G. Howard et al., “MobileNets: Efficient Convolutional Neural Networks for Mobile Vision
Applications,” 2017, [Online]. Available: http://arxiv.org/abs/1704.04861

MALCOM - Vol. 5 Iss. 3 July 2025, pp: 738-745

745