Artificial intelligence experimental design for healthcare data
The PhD research will investigate efficient ways of applying artificial intelligence in healthcare. Artificial intelligence and, in particular, machine learning algorithms have proved extremely efficient in dealing with data generated by healthcare practitioners.
However, in order to make the best of these approaches, it is important to design the machine learning experiments in a way that is optimal in terms of resources involved and in terms of accuracy of the results.
In the machine learning community, a test/train split of data is something done in a standard way, mostly random, with a certain distribution. For clinical data, such split in the experimental phase might not make sense. The data could be split by various attributes: by individual patient, by subsets of patients with similar characteristics, by condition, in a temporal way, and so on. Identifying the right statistical setting for performing the experiments is a challenging task. There are two challenges that this research will investigate, in a statistical design of artificial intelligence experiments for particular types of clinical data:
– Identifying the most suitable machine learning method to be applied for a particular dataset (data might come from multiple sources)
– Optimising the choosing of the hyperparameters for the machine learning model
– Identifying the best learning strategy (which data should be used for training and which for testing) to avoid bias.
The data to work with is temporal (multidimensional time series records) or spatio-temporal. The postholder will work closely with healthcare providers and domain experts.