Cognizant Technology Solutions Corporation
07/31/2024 | Press release | Distributed by Public on 07/30/2024 23:31
Our research shows deep neural networks aid clinical trials
\r\nJuly 31, 2024
July 31, 2024
Our research shows deep neural networks aid clinical trials
A Cognizant study demonstrates how to better predict the success of clinical trials-and allocate resources accordingly.
\r\n"}}" id="text-f340f4c5d8" class="cmp-text">A Cognizant study demonstrates how to better predict the success of clinical trials-and allocate resources accordingly.
Approximately 90% of drug candidates fail in clinical trials-this after pharmaceutical companies spend an average of 10 to 15 years and $1 billion on research and development.
\r\nThis high failure rate raises the question: What if research teams could predict the success of clinical trials before they start? Are there ways to identify trial design aspects, management techniques or recruitment parameters that would indicate the outcome of the study before tremendous time and money were spent?
\r\nWith the help of deep neural networks-computer systems modeled after the human brain that use multiple layers of interconnected nodes to rapidly analyze data and identify complex patterns-the answer is yes.
\r\n"}}" id="text-fb9d11fb89" class="cmp-text">Approximately 90% of drug candidates fail in clinical trials-this after pharmaceutical companies spend an average of 10 to 15 years and $1 billion on research and development.
This high failure rate raises the question: What if research teams could predict the success of clinical trials before they start? Are there ways to identify trial design aspects, management techniques or recruitment parameters that would indicate the outcome of the study before tremendous time and money were spent?
With the help of deep neural networks-computer systems modeled after the human brain that use multiple layers of interconnected nodes to rapidly analyze data and identify complex patterns-the answer is yes.
Deep neural networks enable researchers to better predict trial outcomes based on the design and management of the trial itself. This allows teams to adapt some aspects of the clinical trial to improve the likelihood of success, and to focus resources on activities most likely to have a positive result.
\r\n"}}">Deep neural networks enable researchers to better predict trial outcomes based on the design and management of the trial itself. This allows teams to adapt some aspects of the clinical trial to improve the likelihood of success, and to focus resources on activities most likely to have a positive result.
\r\nTo assess the efficacy of deep neural networks in clinical trials, a team within Cognizant's Life Sciences' internal solutions group set about developing a model that would accurately forecast trial outcomes. The results have been extremely encouraging, pointing to a future in which these trials are both more efficient and more effective-thus speeding the introduction of new treatments.
Beginning the work
\r\nAs a first step, our team converted unstructured text (pulled from 400,000 files sourced from previous trials published on Clinicaltrials.gov) into structured data. The team then used the following advanced techniques to clean the data, analyze the information, and produce actionable insights:
\r\n- \r\n
-
Natural language processing (NLP). This venerable tool uses machine learning to enable computers to understand and communicate with human language.
\r\n
\r\n - \r\n
-
Word embedding. In NLP, a word embedding is a representation (usually a vector) of a word that essentially helps group words with similar meanings.
\r\n
\r\n - \r\n
- The bidirectional encoder representations from transformers (BERT) language model. The BERT machine-learning (ML) framework understands context better than other NLP techniques thanks to its ability to analyze the relationships between words in bidirectional fashion.
- \r\n
Different techniques were then applied based on the format and condition of the data. Next, our team applied key variables to the data, such as the target disease and why the trial was stopped, to identify patterns. In our model, the output was limited to 1 or 0, which represented success or failure.
\r\n
Our model was able to take patterns found in previous trials and identify the same or similar markers in a new trial, thus effectively predicting its success or failure.
\r\nBeyond forecasting success or failure rates, these models also offer valuable insights to guide decision-making in clinical trial design and management, helping teams improve the way they allocate resources, adapt plans, and prioritize efforts.
\r\n"}}" id="text-a10ec23d9b" class="cmp-text">
To assess the efficacy of deep neural networks in clinical trials, a team within Cognizant's Life Sciences' internal solutions group set about developing a model that would accurately forecast trial outcomes. The results have been extremely encouraging, pointing to a future in which these trials are both more efficient and more effective-thus speeding the introduction of new treatments.
Beginning the work
As a first step, our team converted unstructured text (pulled from 400,000 files sourced from previous trials published on Clinicaltrials.gov) into structured data. The team then used the following advanced techniques to clean the data, analyze the information, and produce actionable insights:
-
Natural language processing (NLP). This venerable tool uses machine learning to enable computers to understand and communicate with human language.
-
Word embedding. In NLP, a word embedding is a representation (usually a vector) of a word that essentially helps group words with similar meanings.
- The bidirectional encoder representations from transformers (BERT) language model. The BERT machine-learning (ML) framework understands context better than other NLP techniques thanks to its ability to analyze the relationships between words in bidirectional fashion.
Different techniques were then applied based on the format and condition of the data. Next, our team applied key variables to the data, such as the target disease and why the trial was stopped, to identify patterns. In our model, the output was limited to 1 or 0, which represented success or failure.
Our model was able to take patterns found in previous trials and identify the same or similar markers in a new trial, thus effectively predicting its success or failure.
Beyond forecasting success or failure rates, these models also offer valuable insights to guide decision-making in clinical trial design and management, helping teams improve the way they allocate resources, adapt plans, and prioritize efforts.
Real-world application: A case study for fast-tracking chest X-ray predictions through deep neural networks
To leverage our forecasting technology in the real world, our team developed a use case to help accelerate cancer detection using chest X-rays to train a convolutional neural network (CNN)-a deep learning technique that is particularly well-suited for image and visual data processing.
\r\nAs a first step, we collected a robust data set of chest X-ray images, each of which was labeled normal, pneumonia, or COVID-19. Next, we conducted data preprocessing, which consisted of resizing and normalizing the images to ensure uniformity and suitability for the CNN model.
\r\nThe team then selected a CNN model architecture that would support image classification, which in this case was Visual Geometry Group (VGG). A VGG is a type of CNN developed by the Visual Geometry Group at the University of Oxford that offers a relatively simple and effective way to perform image recognition.
\r\nIn the next phase of the program, our team set about training the model. As part of this process, the preprocessed images were fed into the CNN. This allowed the network to learn and identify features specific to each category (normal, pneumonia, COVID-19) through multiple layers of convolutions and pooling.
\r\nFollowing training, the model underwent validation and testing using a separate dataset to fine-tune parameters and prevent overfitting. This step was crucial for evaluating the model's accuracy on new, unseen images.
\r\nWith validation complete, the trained CNN model was ready for use. With oversight by our team, it was used to classify new chest X-ray images according to the three specified categories based on the patterns it had learned, accurately categorizing 89% of images.
\r\nBy leveraging deep learning techniques to automate the review and classification of X-ray images, this CNN enabled faster and more accurate diagnoses compared to conventional methods. This saved valuable time in initiating treatments and improving health outcomes.
\r\n"}}" id="text-8b80dcd380" class="cmp-text">To leverage our forecasting technology in the real world, our team developed a use case to help accelerate cancer detection using chest X-rays to train a convolutional neural network (CNN)-a deep learning technique that is particularly well-suited for image and visual data processing.
As a first step, we collected a robust data set of chest X-ray images, each of which was labeled normal, pneumonia, or COVID-19. Next, we conducted data preprocessing, which consisted of resizing and normalizing the images to ensure uniformity and suitability for the CNN model.
The team then selected a CNN model architecture that would support image classification, which in this case was Visual Geometry Group (VGG). A VGG is a type of CNN developed by the Visual Geometry Group at the University of Oxford that offers a relatively simple and effective way to perform image recognition.
In the next phase of the program, our team set about training the model. As part of this process, the preprocessed images were fed into the CNN. This allowed the network to learn and identify features specific to each category (normal, pneumonia, COVID-19) through multiple layers of convolutions and pooling.
Following training, the model underwent validation and testing using a separate dataset to fine-tune parameters and prevent overfitting. This step was crucial for evaluating the model's accuracy on new, unseen images.
With validation complete, the trained CNN model was ready for use. With oversight by our team, it was used to classify new chest X-ray images according to the three specified categories based on the patterns it had learned, accurately categorizing 89% of images.
By leveraging deep learning techniques to automate the review and classification of X-ray images, this CNN enabled faster and more accurate diagnoses compared to conventional methods. This saved valuable time in initiating treatments and improving health outcomes.
\r\nLooking to the future of deep neural networks
\r\n
As with any intelligent system in its early stages, the predictive capabilities of deep neural networks may vary. That said, the model will learn and evolve over time, producing more accurate insights as valid data is collected and the system is refined.
\r\nEvolution of the model is also likely to unlock new use cases for deep neural networks in clinical trials. For example, the fundamental concepts established in the AML study could be adapted to address challenges like patient recruitment and eligibility by leveraging similar ML techniques to analyze patient characteristics and match them with trial criteria.
\r\n
In this way, we see that deep neural networks hold incredible promise for revolutionizing clinical trial research, offering unprecedented potential to tackle critical challenges, improve outcomes, and ultimately bring new treatments to market sooner.
\r\n
Looking to the future of deep neural networks
As with any intelligent system in its early stages, the predictive capabilities of deep neural networks may vary. That said, the model will learn and evolve over time, producing more accurate insights as valid data is collected and the system is refined.
Evolution of the model is also likely to unlock new use cases for deep neural networks in clinical trials. For example, the fundamental concepts established in the AML study could be adapted to address challenges like patient recruitment and eligibility by leveraging similar ML techniques to analyze patient characteristics and match them with trial criteria.
In this way, we see that deep neural networks hold incredible promise for revolutionizing clinical trial research, offering unprecedented potential to tackle critical challenges, improve outcomes, and ultimately bring new treatments to market sooner.
Associate, Life Sciences-Internal Solutions Group
\r\n"}}">Himanshu Sharma
Associate, Life Sciences-Internal Solutions Group
Himanshu works with Cognizant's Life Sciences-ISG team. He uses machine learning, deep learning, and generative AI to drive innovation in the life sciences vertical and, ultimately, build a better future.
\r\n \r\n"}}" id="text-a9f8c2b004" class="cmp-text">Himanshu works with Cognizant's Life Sciences-ISG team. He uses machine learning, deep learning, and generative AI to drive innovation in the life sciences vertical and, ultimately, build a better future.
Follow
Manager, Life Sciences-ISG Clinical R&D
\r\n"}}">Pradip G. Injapuram
Manager, Life Sciences-ISG Clinical R&D
Pradip works with Cognizant's LS-ISG team. A pharmacist by profession, Pradip is passionate about using emerging technologies like AI/ML to drive innovation in patient engagement in clinical trials and enable better health outcomes.
\r\n \r\n"}}" id="text-f18bb863a5" class="cmp-text">Pradip works with Cognizant's LS-ISG team. A pharmacist by profession, Pradip is passionate about using emerging technologies like AI/ML to drive innovation in patient engagement in clinical trials and enable better health outcomes.
Follow