VISION

We focus on the challenges posed by modern medicine. These are caused by the increasing complexity of medical data, such as high dimensionality (millions of variables can be measured for each patient), diverse measurements at multiple time points (such as data from blood samples, and scans), interactions between diseases, and the  heterogeneity of diseases and patient groups (no two patients with the same disease are identical). They are also  prompted by the new ambition of medicine: to achieve personalised treatments. To use the vast amounts of new personalised medical data to ensure for each patient the right individual treatment at the right time. 

​

The conventional statistical methods of medicine were not designed with personalised medicine in mind. They cannot cope well with high dimensional measurements (such as genomic data) and disease or patient heterogeneity.  Modern AI  tools are successful in image analysis and processing written patient records, but their performance in clinical outcome and treatment response prediction has not been great. Moreover, they cannot easily use domain information and require very large data sets for training. The situation is even more critical for rare diseases, where large data sets are not available.

​

According to a recent study, only about 14% of clinical trials lead to an approved drug. The failure cost is highest for phase III trials, which cost around 50M$ on average. These measure the benefit of a new drug in larger patient groups, after a phase II trial in which a drug effect was found. Failure at phase III implies that, while some patients may benefit from the drug, their number is too low to make the drug commercially viable. Had one been able to identify the responders prospectively, with better analytical tools, the trial could have succeeded.

​

The challenges of data analytics in modern medicine require novel mathematical and statistical approaches and a thorough understanding of the most pressing medical questions . This is why we have always engaged heavily in research, and work very closely with medical researchers. We use conventional AI methods only for relative simple problems. Instead our strategy is to combine the precision of Bayesian inference methods (so we always know exactly what we compute, and control all sources of uncertainty) with the power of mathematical techniques from theoretical physics (so that the computationally prohibitive Bayesian calculations become feasible in practice).