Director, External Affairs, Science Museum Group; Coauthor, Supercooperators; Frontiers of Complexity
Actionable Predictions

You might be forgiven for thinking that this is so blindingly obvious that it is hardly worth stating, let alone arguing, that it should become a popular meme. After all, "pre-" means “before,” so surely you should be able to take action in the wake of a prediction to change your future—like buying an umbrella when a deluge is forecast, for example.

Weather forecasting is indeed a good example of an actionable prediction, a beautiful marriage of real-time data from satellites and other sensors with modeling. But when you shift your gaze away from the physical sciences towards medicine, these predictions are harder to discern.

We are a long way from doctors being able to make routine and reliable actionable predictions about individual patients—which treatments will help them the most, which drugs will cause them the fewest side effects, and so on.

The answers to many simple questions remain frustratingly elusive. Should I take an antibiotic for that sore throat? Will immunotherapy work for me? Which of that vast list of possible drug side effects should I take seriously?  What is the best diet for me? If only we could predict the answers for a particular patient as reliably as meteorologists predict tomorrow’s weather.

Today there is much talk of Big Data providing all the answers. In biology, for example, data from the human genome project once kindled widespread hope that if we sequenced a patient’s DNA we would get a vivid glimpse of their destiny.

Despite the proliferation of genomes, epigenomes, proteomes, and transcriptomes, that crystal ball looks cloudier than at first thought, and the original dream of personalized medicine in genomics has been downgraded to precision medicine, where we assume that a given person will respond in a similar way to a previously-studied group of genetically similar people.

Blind gathering of Big Data in biology continues apace, however, emphasizing transformational technologies such as machine learning—artificial neural networks, for instance—to find meaningful patterns in all the data. But no matter their "depth" and sophistication, neural nets merely fit curves to the available data. They may be capable of interpolation, but extrapolation beyond their training domain can be fraught.

The quantity of data is not the whole story either. We are gathering a lot, but are we gathering the right data and of sufficient quality? Can we discern a significant signal in a thicket of false correlations? Given that bodies are dynamic and ever-changing, can data snapshots really capture the full complexities of life?

To make true actionable predictions in medicine, we also need a step change in mathematical modeling in biology, which is relatively primitive compared to physics and for entirely understandable reasons: Cells are hugely complicated, let alone organs and bodies.    

We need to promote interest in complex systems so we can truly predict the future of an individual patient, rather than infer what might be in store for them from earlier population studies of who responded to a new treatment, who did not, and who suffered serious side effects. We need deeper insights, not least to end making diagnoses post mortem, and to prevent tens of thousands of people perishing at the hands of doctors every year through iatrogenic effects. 

Ultimately we need better modeling based on mechanistic understanding in medicine so that, one day, your doctor can carry out timely experiments on a digital Doppelgänger before she experiments on you. Modern medicine needs more actionable predictions.