van der Schaar Lab

Using Machine Learning to Individualise Treatment Effect Estimation: Challenges and Opportunities

This work has been realised in a close collaboration of CCAIM, the van der Schaar Lab’s PhD student Alicia Curth and AstraZeneca.

The van der Schaar lab stands out for its vibrant commitment to collaboration, striving to tackle healthcare challenges through the transformative power of machine learning. Among other things, we are leading the pursuit of novel machine learning methods for the development of effective pharmaceuticals and that necessitates a precise estimation of treatment effects.

Not only is the estimation of an expected conditional average treatment effect (CATE) an exciting machine learning problem but it is also having a real measurable impact on pharmacology and, in the long run, the outcome for patients.

We assembled an interdisciplinary team to review the state of the ML CATE estimation literature and to discuss some of the challenges and opportunities for machine learning to estimate CATE. This team included strong machine learning expertise with PhD student Alicia Curth (van der Schaar Lab) and CCAIM director Mihaela van der Schaar, Eoin McKinney (CCAIM faculty) contributing a clinical perspective, as well as Jim Weatherall (AstraZeneca) and Richard Peck (CCAIM faculty) coming from a pharmacological background.

The use of data from randomised clinical trials to justify treatment decisions for real-world patients is the current state of the art. It relies on the assumption that average treatment effects from the trial can be extrapolated to patients with personal and/or disease characteristics different from those treated in the trial. However, clinical pharmacologists understand that patients are unique. Therefore, leveraging machine learning to estimate CATE using larger and more inclusive observational datasets offers the potential for more accurate estimation of the expected treatment effects in each individual patient based on their observed characteristics.

Conventional machine learning methods developed for standard prediction problems often lack the explicit design features for forecasting treatment effectiveness, particularly in the context of determining whom to treat, when, and with which intervention. As we want to answer the question “What would be the outcome if I were to modify the current treatment policy?”, we need to tread new ground.

With this in mind, the authors discuss the challenges and opportunities for the use of ML for forecasting individual treatment responses in their paper:

Machine learning holds great potential for individualising treatment effect estimation. However, conventional prediction methods are not natively equipped to handle them. On the back of this comprehensive review, the Centre and its collaborators have their future work cut out for them to lead the exploration and to revolutionise ML methods to realise its great opportunities.

Prof Richard Peck and author Alicia Curth discussing CATE:

Here is what our team has to say about their study:

I’ve found it very interesting to study heterogeneous treatment effect estimation as a machine learning problem not only because it is ubiquitously important in many medical applications but also because it comes with unique technical challenges that go beyond what you’d encounter in standard prediction problems — meaning that it’s usually not enough to apply existing machine learning methods off-the-shelf without coming up with smart ways to address these challenges!

Alicia Curth

Today, machine learning is revolutionising so much of the world we live in. We can see opportunities to increase productivity, drive sustainability, improve healthcare, and more. As we consider this potential, it is vital that we turn our attention to addressing challenges in the utilisation of machine learning and enable these new technologies to deliver benefits for patients, society and the planet.

In this collaboration with CCAIM, some of the world’s brightest minds have turned their attention to the challenges around estimating individual response to treatment. The application of proposed solutions offer significant improvements in patient benefit through more accurate treatment decisions and more personalised care.

Jim Weatherall

For nearly ten years, the van der Schaar lab has been at the forefront of exploring treatment effect estimation through real-world observational data, harnessing the transformative power of machine learning. As we embark on this new chapter with CCAIM, our fusion of cutting-edge research and in-depth clinical and pharmacological insights is not just about translating theory into practice; it’s about pioneering new frontiers. This collaboration marks a pivotal moment in our journey, promising to revolutionise personalised treatment strategies. We stand at the cusp of an era where research on individualised treatment effects can start shaping the future of personalised medicine.

Mihaela van der Schaar

You can find the full publication here.

Want to learn more? See our website on individualised treatment effects and our tutorial on treatment effect inference.

Andreas Bedorf