Drug-Resistant Epilepsy Prediction Using Deep Learning
Accelerating Diagnosis of Drug-Resistant Epilepsy with Deep Learning Models
problem
Our client, a multinational biopharmaceutical company based in Europe, focuses on researching and developing medications for various neurological conditions, including epilepsy, Parkinson’s, and Crohn’s diseases. In the case of epilepsy which affects around 65 million people globally, with a significant portion experiencing drug-resistant epilepsy (DRE), where seizures persist despite treatment. Diagnosing DRE typically requires several years of trying different medications, delaying effective treatment, and impacting patients quality of life. The challenge was to expedite the identification of drug-resistant epilepsy cases, allowing for faster intervention and improved patient outcomes.
problem
Our client, a multinational biopharmaceutical company based in Europe, focuses on researching and developing medications for various neurological conditions, including epilepsy, Parkinson’s, and Crohn’s diseases. In the case of epilepsy which affects around 65 million people globally, with a significant portion experiencing drug-resistant epilepsy (DRE), where seizures persist despite treatment. Diagnosing DRE typically requires several years of trying different medications, delaying effective treatment, and impacting patients quality of life. The challenge was to expedite the identification of drug-resistant epilepsy cases, allowing for faster intervention and improved patient outcomes.
solution
Our team initiated a research project targeting US patient data provided by Symphony Health, encompassing approximately 450,000 individuals with epilepsy. The dataset comprised de-identified information such as gender, age, diagnoses, medical procedures, and prescribed medications. Leveraging the client s existing research outcomes, which employed a feature-based model combining expert knowledge and machine learning, we adopted a deep learning approach for our project. We developed and evaluated various deep learning models, including the Hierarchical Attention LSTM network, ULMfit, plain LSTM, and SWEM concatenation model, utilizing PyTorch and Tensorflow frameworks.
solution
Our team initiated a research project targeting US patient data provided by Symphony Health, encompassing approximately 450,000 individuals with epilepsy. The dataset comprised de-identified information such as gender, age, diagnoses, medical procedures, and prescribed medications. Leveraging the client s existing research outcomes, which employed a feature-based model combining expert knowledge and machine learning, we adopted a deep learning approach for our project. We developed and evaluated various deep learning models, including the Hierarchical Attention LSTM network, ULMfit, plain LSTM, and SWEM concatenation model, utilizing PyTorch and Tensorflow frameworks.
results
By utilizing a deep learning approach, we developed an end-to-end model that learns directly from raw data, eliminating the need for human expertise. This approach yielded a DRE identification model with an accuracy of 81%, outperforming the previous expert-based method s accuracy of 77.7%. As the model continues to learn from additional data, its accuracy is expected to improve further. Our ongoing collaboration with the client involves certifying the model with the FDA, a process that requires time. Furthermore, our team has uncovered valuable insights from the model, such as the association between depression and DRE, aiding experts in identifying cases more effectively. The project was conducted securely using Azure s remote desktop infrastructure, ensuring compliance with data protection regulations.
results
By utilizing a deep learning approach, we developed an end-to-end model that learns directly from raw data, eliminating the need for human expertise. This approach yielded a DRE identification model with an accuracy of 81%, outperforming the previous expert-based method s accuracy of 77.7%. As the model continues to learn from additional data, its accuracy is expected to improve further. Our ongoing collaboration with the client involves certifying the model with the FDA, a process that requires time. Furthermore, our team has uncovered valuable insights from the model, such as the association between depression and DRE, aiding experts in identifying cases more effectively. The project was conducted securely using Azure s remote desktop infrastructure, ensuring compliance with data protection regulations.
tech stack
tech stack
