Reliable Diagnostics by Conformal Predictors

Prof. Alexander Gammerman
UK, Royal Holloway, University of London
The talk reviews a modern machine learning technique called Conformal Predictors (CP) that allows us to make reliable predictions using valid measures of confidence in both batch and online modes of learning. Unlike many conventional techniques, the approach does not make any additional assumption about the data beyond the iid assumption: the examples are independent and identically distributed. The talk outlines the basic ideas of Conformal Predictors and then illustrates the technique with applications to several medical problems. One of them is early diagnosis of breast cancer using proteomics data – the high-dimensional mass-spectrometry data obtained from serum samples of breast cancer patients and collected in a university hospital. The aim of the analyses is to construct diagnostic rules to distinguish between cancer patients and healthy controls. The technique allows us to make valid diagnoses for individual patients and guarantees that the overall accuracy in diagnoses can be controlled by a required confidence level.
Another illustration is an application of Conformal Predictors to functional and structural magnetic resonance images (MRI) in order to diagnose depression and make prognostic decisions at the individual level. The advantage of CP for psychiatric classification is that they provide measures of confidence which are given to each diagnostic or prognostic decision and thus the risk of an erroneous clinical decision is known for a given individual. Moreover, the risk of error may be controlled by predetermining an acceptable level of confidence for a given clinical decision and therefore the risk of misdiagnosis is known. The conformal predictors are region predictors, and would allow us to make a number of possible diagnoses. This feature may be particularly useful in situations in which the cost of a clinical error is high.