At NHS we strive to be leaders in innovation ensuring that patients are receiving the best care that modern technology can provide. We are proud to partner with Kortical and their platform to implement cutting edge AI tools to benefit patient care.
Blood products have a short shelf life, in fact platelets only last 7 days. Hospitals need a stock of blood in all the various types and treatments, so they can ensure they can give the patients the blood they need. Ensuring all the hospitals have a supply at all times is a complex problem which involves understanding supply, manufacturing, distribution, stock holding, logistics and hospital demand.
Kortical is using their platform to build lots of AI / ML models to predict what will happen better and allow for needs to be met more precisely, thus reducing costs and wastage, as well as costly ad hoc deliveries.
The real challenge is predicting the demand for the different blood products in all the hospitals. The more accurately we can do this the better we can plan ahead. We also predict the supply. With these factors and the live data of what’s actually happening, it’s a case of notifying donors to make up any shortfall and then optimising from supply through to arrival at the hospital.
To predict supply and demand for all the hospitals accurately we need to create machine learning models that can factor in data such as weather, among other data that affects what people do and ultimately who might show up needing blood on any given day. Using the Kortical platform, we are able to rapidly create complex models from multiple data sources and stream new data to keep them updated.
The output of the machine learning models isn’t much use to humans directly but using the microservice hosting in the platform, we’re able to quickly create a nice web app, that lets the user monitor and control the demand planning.
Demand prediction is a time series problem. To help the model understand the temporal aspects we need to create lagged data for the input rows. eg: how many type A platelet units were used this day last week, this day two weeks ago, this day last year. Now the problem is that there are many different ways to create these lagged variables but we can use the Kortical python library to create them easily. Once we’ve created the lags we can upload the data and let the Kortical platform do its magic, cleaning, encoding, normalising, etc. the data in various ways and selecting the best model and parameters for this dataset.
With Kortical taking care of so much, a lot of data-scientists wonder what they have left to do but creating features that better represent the domain to a machine learning model can make a huge difference in model performance. Different features will work better with different model types but with Kortical, the data-scientist doesn’t need to worry about that. Data-scientists being freed from the drudgery of data plumbing, model selection and tuning allows them to to rapidly iterate on many sets of features quickly. The platform tracks all the experiments automatically, so nothing is lost. For this problem the data-scientists were mainly experimenting with different forms of lagged variables, to get the most accurate predictions.
Using the Kortical Cloud API we can host our code that uses the models and displays the UI easily. We just have to publish the best model via the UI, copy the REST API code and we can call that model from our app. We’ve managed to build an entire AI app in a day to win a hackathon but this application consists of multiple services, lots of UI elements and functionality, so it’s a much longer build. The Kloud API hosts the code, managing the app for scale, failover, redundancy, etc. making building cloud based AI apps / microservices very straightforward.