Titolo della tesi: Explainable Clinical Decision Support System - Opening black-box meta-learner algorithm Expert’s based
Mathematical optimization methods are the basic mathematical tools of all artificial
intelligence theory. In the field of machine learning and deep learning the examples
with which algorithms learn (training data) are used by sophisticated cost functions
which can have solutions in closed form or through approximations. The interpretability
of the models used and the relative transparency, opposed to the opacity
of the black-boxes, is related to how the algorithm learns and this occurs through
the optimization and minimization of the errors that the machine makes in the
learning process. In particular in the present work is introduced a new method for
the determination of the weights in an ensemble model, supervised and unsupervised,
based on the well known Analytic Hierarchy Process method (AHP). This method is
based on the concept that behind the choice of different and possible algorithms to
be used in a machine learning problem, there is an expert who controls the decisionmaking
process. The expert assigns a complexity score to each algorithm (based on
the concept of complexity-interpretability trade-off) through which the weight with
which each model contributes to the training and prediction phase is determined.
In addition, different methods are presented to evaluate the performance of these
algorithms and explain how each feature in the model contributes to the prediction
of the outputs. The interpretability techniques used in machine learning are also
combined with the method introduced based on AHP in the context of clinical
decision support systems in order to make the algorithms (black-box) and the results
interpretable and explainable, so that clinical-decision-makers can take controlled
decisions together with the concept of "right to explanation" introduced by the
legislator, because the decision-makers have a civil and legal responsibility of their
choices in the clinical field based on systems that make use of artificial intelligence.
No less, the central point is the interaction between the expert who controls the
algorithm construction process and the domain expert, in this case the clinical one.
Three applications on real data are implemented with the methods known in the
literature and with those proposed in this work: one application concerns cervical
cancer, another the problem related to diabetes and the last one focuses on a specific
pathology developed by HIV-infected individuals. All applications are supported by
plots, tables and explanations of the results, implemented through Python libraries.
The main case study of this thesis regarding HIV-infected individuals concerns an
unsupervised ensemble-type problem, in which a series of clustering algorithms are
used on a set of features and which in turn produce an output used again as a set of
meta-features to provide a set of labels for each given cluster. The meta-features
and labels obtained by choosing the best algorithm are used to train a Logistic
regression meta-learner, which in turn is used through some explainability methods
to provide the value of the contribution that each algorithm has had in the training
phase. The use of Logistic regression as a meta-learner classifier is motivated by the
fact that it provides appreciable results and also because of the easy explainability
of the estimated coefficients.