About the Client
Multinational biopharmaceutical company from Europe. Their primary focus is research and development, specifically those involving medications centered on epilepsy, Parkinson’s, and Crohn’s diseases.
Around 65 million people worldwide have epilepsy. It is the most common serious neurological condition, characterized by recurrent seizures. The condition of some people may be improved with drugs, which frees them from seizures with minimal or no side effects at all. However, there are a number of people that are drug-resistant (DRE), meaning that they continue having seizures despite appropriate treatment with antiepileptic drugs. 1 in 3 people with epilepsy has the kind that is resistant to medication. In order to diagnose DRE, a doctor should try various medications first and only after a certain amount of time, which usually takes around 3 years, a patient can be diagnosed with DRE.
Specializing in epilepsy treatment, our client wanted to research ways to help doctors identify drug-resistant epilepsy cases faster – around 1 year, which in its turn could allow patients to get a better quality of life.
The research project conducted by the Quantum team focused on the US patient’s data. The client had access to claims data, provided by Symphony Health, which is around 450 000 people suffering from epilepsy. The data included the following de-identified information: gender, age, diagnosis (ICD code), a list of procedures taken, medicines prescribed with exact names and doses.
Before the start of the project, the client has shared with us the outcomes of his research (which still continues) on this topic using a feature-based model that mixes expert knowledge with learning from data with the help of simple linear ML models.
Since the client possessed data that was accumulated for 5-10 years, this has allowed us to implement a deep learning approach, i.e. creating an end-to-end model that learns purely from raw data. In this case, no human expertise is used, as the model learns everything directly from its own observations. The approach we’ve used has allowed us to create a model that identifies DRE with 81% accuracy, while the previous expert-based method showed 77,7% accuracy. The deep learning method we have used will increase the accuracy score as more data will be fed into the model. The client is now certifying the model with the FDA, but the certification process takes time.
Nonetheless, our team continues to work hard and has managed to retrieve new insights from our model that can help experts identify DRE or non-DRE, for example, people prone to depression are more likely to have DRE.
All the project was executed via remote desktop on Azure at the client’s computing facilities, thus we have managed to follow all the rules when working with sensitive data.
Let's discuss your idea!
The baseline was created using the XGBoost classifier. Then, we iterated through a series of deep learning models. These include the Hierarchical Attention LSTM network, ULMfit, plain LSTM, and the SWEM concatenation model. All deep learning solutions were created using PyTorch and Tensorflow.