Transforming medicine through AI-enabled healthcare

There are serious challenges in creating algorithms of sufficient complexity and nuance to deal with disparate clinical data drawn from myriad sources.

In other areas of medicine, some research groups are using ‘Markov state-space models’ (HMMs) to model cancer progression. These statistical, probabilistic models are relatively easy to compute and interpret, but the problem with them, says van der Schaar, is that they “compact the patient’s entire clinical history into a single current state, thereby ‘forgetting’ what has happened to this particular patient.” This is a problem because in medicine, patient history – such as the order and duration of a variety of health events – really matters.

Meanwhile, powerful deep learning tools such as recurrent neural networks (RNN) are able to deal with longitudinal data – i.e. the patient’s history. But they, too, are not powerful enough, by themselves, to deal with the sheer complexity of medicine. So van der Schaar first built a ‘Disease Atlas’ in which she applied a mechanism called ‘attention’, which enables the RNNs to discern which events in the patient’s history are the best predictors of how their disease will progress. (This paper was published in the Machine Learning for Healthcare Conference 2018 and received the best paper award at that year’s IJCAI-BOOM Workshop.) While the predictions made by this sort of RNN are pragmatic and dynamic, these neural networks are essentially black boxes, and so cannot explain, for example, why a particular event has happened, or what the effect of a treatment might be. In short, RNNs do not allow the incorporation or extraction of clinical knowledge.

The key to solving these problems was to keep the useful probabilistic structure and interpretability of HMMs, but use RNNs to model the dynamically evolving state of the patient – the best of both worlds. To this end, in 2018 van der Schaar’s team unveiled a machine learning system called the ‘Attentive State Space’ model. (This method was recently accepted in the top machine learning conference – NeurIPS 2019.) Unlike Markovian state-space models, in which the dynamics are ‘memoryless’, this model uses an ‘attention’ mechanism to create ‘memoryful’ dynamics, whereby relative weighting of ‘attention’ determine the dependence of future disease states on past medical history. “It combines ideas from statistics with ideas from deep learning. It is this wonderful synergy that gives rise to a variety of good solutions,” she says. “What drives the model is both ‘static’ information acquired about the patient, such as gender or genetic information, as well as the evolution of the patient over time.”

"[This work] combines ideas from statistics with ideas from deep learning. It is this wonderful synergy that gives rise to a variety of good solutions.”

Mihaela van der Schaar, Turing Fellow

In the first half of 2019, the collaborators built a demonstrator system, using real, but anonymous data from the NDRS on England’s breast cancer patients – all of those currently alive in the country and many who have died – and a variety of machine learning algorithms built by van der Schaar’s team.  

The demonstrator takes information from the pathology report of an individual diagnosed with breast cancer and uses it to make an initial estimate of mortality risk over time, using a technology created by van der Schaar’s team, called AutoPrognosis. Another piece of the suite of technologies (a method called INVASE) can tell which variables are important for this particular patient at this moment in time. But that’s just the beginning. As the patient continues on their healthcare pathway, with further tests, treatments and other health events such as the appearance of another tumour, their data become richer, and the dynamic predictions more accurate. This enables the system to make individualised predictions about the patient’s healthcare pathway, including competing risks (e.g. mortality due to primary cancer, secondary cancer, or risks unrelated to cancer such as cardiovascular risk). Moreover, the ML-AIM system can go beyond predictions of risk: it can also recommend which treatments would be most effective for the specific patient, instead of relying on one-size-fits-all recommendations (see Fig 1).

Needles in haystacks

“In a clinic, what the doctor wants to know is, how similar are you to other patients?” says Rashbass. After all, that’s how most doctors work: they treat patients based on their knowledge of similar patients. But similarity is hard to determine, says Rashbass. “You can do it very simply, based on the age and sex of the patient, the tumour type and cancer stage. But that isn't good enough.” In the demonstrator, a clinician can say: ‘based on the hundreds of patient variables we've got in the model, find me the patients who are most similar to the patient sat in front of me’. “That is the start point for everything else in medicine,” says Rashbass. “The beauty of the technology we have brought to this, is that it allows us to do deep predictive clustering of individual patients.” (see Fig. 2)

This clustering will allow clinicians to explore a whole range of questions beyond the prognosis for a given patient, by allowing them to see the various probabilities that a patient will follow particular healthcare pathways, based upon the most similar patients in the country, how their cancer progressed and the outcomes of their treatment choices. This source of decision-support for clinicians and patients will be transformative and deliver truly personalised medicine.