Bespoke Learning of Disease Progression Using Inter-Network Distance: Application to Haematology-Oncology: Joint Work with Dr. Kangrui Wang, Dr. Akash Bhojgaria and Dr. Joydeep Chakrabartty

In certain real-world problems, supervised learning of the functional relation is sought between a multivariate time series and associated system properties, that serve respectively as the output and input variables. Such learning will then require training data comprising pairs of designed value of the input and the corresponding realisation of the output—except, realisations of this output at different design points, may vary in temporal span. If truncation to the minimum of such time points is not possible, or not preferred, we require a parametrisation of the considered output, where such a parameter is unaffected by the length of the output value. Such a lossless embodiment of the time series output is bespoke learnt, using the scaled Hellinger distance between the graphical models that are learnt, for a pair of time series data sets. An application of such bespoke learning of a viable scalar parameter that stands for such a variably-long output, is presented, to learn the risk score for developing the potentially terminal disease SOS/VOD, that affects some recipients of bone marrow transplants. Following the learning of this score, its functional relation with the vector of the pre-transplant variables is sought—by learning this function as modelled with a Gaussian Process. Subsequently, we undertake prediction of the SOS/VOD score at the pre-transplant stage, for new patients, whose pre-transplant variables are known.