Dynamic Causal Modeling for Fast Explainable AI

By Dr. William Jones



July 05, 2023


Dynamic Causal Modeling for Fast Explainable AI

Dynamic Causal Modeling (DCM) is a key explainable AI technology that has recently been released as a commercial grade, free and open source software package, dcEmb.

DCM is the technology behind one of the UKs leading COVID-19 models, as well as a large swathe of cutting edge neuroscientific research across the world. Limited until recently to Academia, the technology has the potential to solve several key problems in industry.

AI and Machine Learning are becoming increasingly integrated with many industries. At the same time, AI and Machine Learning solutions are also becoming larger and more complicated. For many of us this presents a problem; with this increasing integration of AI with what we do, there is mounting pressure for us to understand how our AI operates, and to be able to prove that it functions in a safe, trustworthy, and accountable way. DCM is one technique that does not suffer from these issues, with trust and accountability as intrinsic qualities of a system whose learning process is explainable and human understandable end-to-end. In neuroimaging applications, for example, DCM is prominently able to describe highly complex evolving brain activity in a human understandable way in terms of the dynamic relationships between different brain regions.

DCM explains complex evolving brain activity (right) by tying the observed FMRI brain activity to dynamic interactions between brain regions (left).

An often underappreciated aspect of explainability is the ability to deal with uncertainty. In many applications, making predictions without the context of how likely these predictions are to manifest makes risk-based planning very difficult. In the case of AI based COVID-19 modeling, for example, many AI algorithms were created that predicted the number of COVID-19 deaths. However, without the context of likelihood, these predictions are meaningless. AI algorithms that cannot quantify uncertainty around their predictions to make statements like “there is a 95% likelihood that we will need between 200-300 deaths this month” are of very limited use to civic planners. DCM is once again one of the leading technologies in this area, providing uncertainty quantification as a core part of its functionality. Combined with its intrinsic explainability, the COVID-19 DCM model was able to make useful, probability bounded predictions on a wide range of use COVID-19 quantities, including the number of ventilated patients, and the number of required hospital beds.


The COVID-19 Dynamic Causal Model is able to predict useful estimates of many important quantities, and describe their likelihood due to its qualities as an intrinsic explainable and uncertainty aware technology.

Meeting the increasing standards of explainability in industry is a very current challenge for many of us. It is also, sometimes, an opportunity. As AI models become more human understandable, they lend themselves more to being improved. When solving problems with AI algorithms we, the designers, will often have pre-existing knowledge of the problem we’re solving. The more of this we can integrate into our AI, the more focused and effective the AI model will be. The original application of DCM, neuroimaging, is an excellent example of this. Neuroimaging data is often extremely expensive to collect, and of very high complexity. Without any ability to interpret the data in the context of how we already understand how the brain works, it is extremely challenging to work with. DCM is, unsurprisingly, very strong in this area, and it is through the integration of prior knowledge that we are able to solve many problems.

Previous versions of the DCM software have been written and maintained as academic research code in MATLAB. The dcEmb package represents the first commercial grade implementation of the software. Key features include:

  • Configurable design suited to a wide range of problems, outside of COVID-19 or neuroimaging.
  • A C++ codebase using specialized linear algebra libraries for high performance. This includes scaling on high performance computing(HPC) systems.
  • Extensive documentation, making the software readily accessible to a skilled professional engineer.
  • An open source codebase with an extensive testsuite to ensure commercial robustness, transparency, and auditability.
  • A future release will include a set of bindings to support the Python community.

The dcEmb C++ implementation is substantially faster than the MATLAB baseline, making better use of available resources.

As well as the original COVID-19 and neuroimaging models, we have used dcEmb to implement carbon climate models, and some interesting physics models. dcEmb is currently a free and open source software package available on GitHub. If you wish to know more about the software, visit the Embecosm website, or get in touch.

Dr. Jones has a research background in computational neuroscience, focused on the areas of consciousness, cognition, and meta-cognition. In his current role, Dr. Jones heads AI and Machine Learning efforts at Embecosm, a consultancy that solves problems in challenging areas of open source, with a focus on compilers and tool chains. His current work is focused on Bayesian methods and statistical computing.

More from Dr. William