About
AutoPrognosis 2.0
written by Andreas Bedorf and Mihaela van der Schaar
The van der Schaar lab has been developing AutoPrognosis, a cutting-edge framework and associated open-source package which enables easy creation of new risk scores, personalised diagnostics, and prognostics using state-of-the-art machine learning methods.
By erasing the need for significant technical expertise, we democratise predictive modelling, making it accessible for a wide array of expert and non-expert clinicians, healthcare practitioners, researchers, industrial developers as well as anyone interested in building a risk score for a disease or predicting an event of interest. AutoPrognosis provides not only predictions and forecasts, but also a wide-range of interpretability methods that enable users with the ability to understand and debug the resulting machine learning models.
You can find the software via https://github.com/vanderschaarlab/AutoPrognosis as an open source Python package.
We first presented an early version of the AutoPrognosis framework in a paper at the 2018 edition of ICML. Its goal was to automate the development of predictive models that may be used to guide screening and treatment decisions by predicting how patients' clinical conditions will develop in the future.
The core component of AutoPrognosis is an algorithm that automatically configures and optimises machine learning pipelines, each of which comprises a combination of algorithms for missing data imputation, feature processing, model selection and training, as shown schematically to the right.
Since its first introduction in 2018, AutoPrognosis has already been applied in a number of clinical settings, including predicting outcomes for cardiovascular disease, cystic fibrosis, and breast cancer. Most recently, AutoPrognosis was adapted into a tool for hospital capacity planning as part of the U.K. National Health Service’s response to COVID-19.
The accuracy of AutoPrognosis has repeatedly outperformed that of both cutting-edge machine learning models and routinely used statistical techniques.
Now, we are taking the next big leap forward: AutoPrognosis 2.0.
We are proud to introduce AutoPrognosis 2.0 (you can find the paper here), which provides an exceptional leap in the scope, performance, and accessibility of AutoPrognosis for the healthcare community and beyond. Making use of recent developments in AutoML and beyond, making some major advances:
Expanded scope to classification, regression, and time-to-event (survival) analysis
Enhanced optimisation of machine learning pipelines for all stages of model development
Directly incorporated model explainability and debugging tools
Greatly improved usability, making the software increasingly straightforward
Enabled translation of new models into web-based tools without requiring specialised knowledge
This opens the field to clinical domain experts, such as clinicians.
AutoPrognosis 2.0 takes either raw or curated medical datasets, manages missing data, selects the optimal ML pipeline, and provides a report detailing the predictive model, along with explanations of how the model comes to its conclusions. Finally AutoPrognosis 2.0 translates the final predictive model into a web-based interface for clinicians to use.
If you would like to explore what algorithms are currently included in AutoPrognosis 2.0, please refer to the table to see them grouped by pipeline stage. Numbers in brackets correspond to the number of hyperparameters optimised over by AutoPrognosis. Our framework is readily extendable to additional methods, algorithms, and hyperparameters.
We believe that AutoPrognosis 2.0 has the potential to fundamentally transform the field and provide clinicians with a level of machine learning empowerment that has never been seen before.
Below, we outline the various challenges which AutoPrognosis can address.
AutoPrognosis 2.0 is a healthcare-focused AutoML framework that is innately adaptable. New algorithms can be added to the mix whenever needed or when new advances in machine learning methods are made. Using the data from all pipelines, AutoPrognosis 2.0 can either select the best performing pipeline or create an ensemble of pipelines based on the empirical likelihood that that pipeline will produce the best results.
To see how AutoPrognosis 2.0 can be used in practice, we provided an illustrative application where we construct a prognostic risk score for diabetes using the UK Biobank. While we have provided an illustrative example of how AutoPrognosis can be used, the key finding reported here is not the performance of a single illustrative model, but rather the way in which it was built.
In summary, AutoPrognosis 2.0 is designed in such a way that both expert and non-expert users can use it, thereby democratising model construction, understanding, debugging, and sharing. We believe AutoPrognosis 2.0 is a necessary development in the journey towards widespread adoption of ML systems in clinical practice and hope that researchers will engage with this tool. Rather than marginalising healthcare experts, we believe AutoPrognosis places them at the centre and empowers them to create new clinical tools. As part of this journey, we will continue to add new features and improve AutoPrognosis. Stay tuned….
Finally, while the focus and motivation for AutoPrognosis is medicine, it has not escaped our notice that AutoPrognosis can be used to construct predictive models and risk scores for applications beyond healthcare, such as social sciences, education, finance, energy, and more.
Endorsements
“We have built AutoPrognosis, a tool aimed at democratising machine learning for anyone interested in developing new risk scores, personalised diagnostics, and prognostics using state-of-the-art, interpretable machine learning techniques. We hope that, irrespective of their knowledge of machine learning, numerous clinicians, researchers, and healthcare professionals will use AutoPrognosis to build powerful analytics that will empower them to better support patients.”
“Machine learning has great potential to change how we practise medicine. However that potential cannot be realised unless cutting edge algorithms are used to address important questions posed by those with clinical knowledge. AutoPrognosis makes this possible, allowing clinical experts to access powerful tools in a robust, reproducible way that has not so far been feasible without substantial technical know-how.”
“AutoPrognosis is the first comprehensive, low-code framework for developing and deploying clinical risk prediction tools. With a straightforward interface, AutoPrognosis brings the power of state-of-the-art methods to machine learning experts and non-experts alike, freeing the healthcare community to focus on building the right risk prediction tools for the most pressing questions. AutoPrognosis is likely to become your co-pilot in any clinical risk prediction task.”