AI Wonder Girls: ICU-OPS

• 

  The problem

• 

  The team

• 

  User interface

• 

  User Interface Diabetes Prediction

• 

  User Interface Question and Answering system

• 

  Pipeline workflow

• 

  User Interface Model details

• 

  Thank you!

Inspiration

Artificial Intelligence has become a pervasive and empowering technology in all sectors of life, be it retail, finance, healthcare, manufacturing and much more. In the healthcare sector, AI-enabled tools can help identify diseases more quickly, predict clinical outcomes, gain valuable insights from tons of medical reports, clinical trials, medical literature and patient records. AI has penetrated every field of medicine, including drug development, treatment decisions, patient care, and financial and operational decisions. The AI wonder Girls team members are experienced data scientists and are passionate about contributing their AI skills to Healthcare Domain addressing the UN SDG 3- Good Health and Wellbeing.

Various studies in the COVID-19 pandemic have suggested the high risks associated with Diabetes Mellitus condition, during the disease and post-covid phase. High levels of glucose also seem to take longer recovery.
The project investigates the recent research studies that highlight the Diabetes Mellitus condition risks associated with other diseases especially COVID-19. It also aims to predict Diabetes Mellitus condition on patient parameters in ICU hospital admissions that will help the clinicians and hospital staff plan for better care and treatment to these patients thereby reducing mortality rate.

What it does

The project has NLP for research study analysis using Natural Language Processing. This analysis brings forth the associations with Diabetes Mellitus condition on the dataset, highlights factors related to the condition. Question and Answering NLP model is built to get the answers from the selected study papers on the subject.

The second part of the project focusses on the prediction of Diabetes Mellitus condition based on patient parameters on ICU admission. For this purpose, WIDS 2021 dataset is taken, which focuses on patient health, with an emphasis on the chronic condition of diabetes. An ICU patient's chronic conditions, such as heart disease, injuries, or diabetes, may not be readily available due to the patient's condition or if the patient is from another medical provider or system. Knowing a patient's chronic diseases can expedite clinical decisions about their care and ultimately improve their health outcomes. The speed-up of patient outcomes relieves Intensive Care Units (ICUs) struggling with overload from critical COVID-19 pandemic cases. Our AI solution, ICU-OPS, is rapidly channeling medical emergencies in the right hands, leading to the best possible patient outcomes.

How we built it

NLP : The NLP dataset is made from a few of study papers available on the net on the subject of Diabetes Mellitus. Topic Modelling, LDA analysis is performed to get the relations, topics on the condition. BERT model is custom trained on the NLP dataset on MLRun for Question Answering system. A serving is made with the trained BERT model and deployed which is the endpoint for the QnA system. The User Interface application sends query on the REST api to QnA Serving function that sends back the related answers to the UI.
Prediction Model: The dataset is preprocessed, trained on LGBM model and deployed as serving function with REST api endpoint. The UI sends the patient data for prediction to the lightgbm serving based on the patient data. Nuclio serverless functions are developed for preprocessing, training, serving the models, these can be deployed on the cluster for distributed processing and computation.

Kubernetes workflow is made by hooking all the components of the pipeline and is deployed with CD from GitHub with a change in the pipeline and the results are posted to Slack connected by hooks User Interface for the application is built on Streamlit that shows the NLP models analysis, query as API to deployed QnA serving function and shows the results. The UI also shows the trained models LGBM, XGB metrics and the explainability of the project is shown as visualisations. The UI showcases the prediction model that sends the patient ICU parameters to the deployed serving function on REST api that returns prediction of diabetes condition from the model.

Challenges we ran into

Understanding of the platform, getting the first flow working took time on the platform, the dashboard UI sync

Accomplishments that we're proud of

End to end multimodal ML pipeline with NLP and prediction model built on the MLRun/Iguazio platform along with user interface showcasing the features of the pipeline. Seamlessly bring the ML workflow into a production environment with MLRun and Iguazio platform

What we learned

Nuclio serverless functions, Kubernetes workflow and CD with Github/slack integration, User Interface with Streamlit, creating multimodal ML application with NLP, prediction models

Creating models, productionizing the ML pipelines based on data, application or business needs with the ability to scale, monitor, control and model governance typical of MLOps flow is seamlessly achieved with the useful features of MLRun/Iguazio MLOps platform

We hope that this ML application brings forth the studies in the field and aims to reduce the load on the stressed pandemic healthcare system in the pandemic times like COVID-19.

What's next for AI Wonder Girls:

The application can be easily extended to more research studies by updating the dataset. We hope to bring value-added applications in healthcare and AI for good sectors that address a current need and create positive impact on the society at large in alignment with UN-SDG goals

Built With

• azure
• bert
• docker
• github
• mlrun
• nuclio
• python
• streamlit