During the last decade, a new trend in medicine is transforming the nature of healthcare from reactive to proactive. This new paradigm aspires to detect diseases at an early stage and introduce diagnosis to stratify patients and diseases to select the optimal therapy based on individual observations. This paradigm transformation relies on the availability of complex multi-level biomedical data. In order to take advantage of this information, an important effort is being made to develop new mathematical and computational methods for extracting maximum knowledge from patient records. This requirement enables the use of computer-assisted Clinical Decision Support Systems for the management of individual patients.
The Clinical Decision Support System (CDSS) are computational systems that provide precise and specific knowledge for the medical decisions to be adopted for diagnosis, prognosis, treatment and management of patients. The framework and the origin of this Thesis is precisely the development of a CDSS based on Machine Learning algorithms to infer predictive models for non-invasive brain tumour diagnosis. This process began with the European project INTERPRET (2002) and went on with other two European projects eTUMOUR (2005) and HEALTHAGENTS (2008), which have endeavoured to develop an automatic diagnostic tool applied to Proton Magnetic Resonance Spectroscopy (1HMRS) data from brain tumours. A major aim was to minimize the need for an invasive histological diagnosis of a brain tumour biopsy. Machine Learning has been successfully applied to this problem providing automated analysis of 1HMRS. However, the development of brain tumour classifiers able to generalize requires a large number of cases to be acquired for each tumour type and at present the approach has only been used for a few common tumours. Cases are collected from a large number of hospitals over many years and data transferred to a centralised database. This approach has several disadvantages, ethical approval and patient consent needs to be obtained to send and store data. Distributed databases in which classifiers can be trained without moving the data from the hospital at which it was collected would provide a practical solution. The ability to retrain the classifiers as new data accumulates is also an important requirement and to meet these needs, incremental learning algorithms may give a practical optimal solution.
After the analysis of non-incremental ML approaches for automatic brain tumor diagnosis, this Thesis introduces new incremental learning algorithms of general purpose for stationary environments and particularly for adapting the predictive models to new centers in the framework of automatic brain tumour decision making using 1HMRS.
Until now, the different CDSS developed for brain tumour diagnosis have only used non-incremental classification models. Non-incremental classifiers entail an implicit assumption that learning stops when the current training set has been processed. Hence, the performance of a non-incremental automatic classifier strongly depends on the availability of a representative training set for each class. However, the gathering of these data is often expensive and time-consuming. Under these circumstances the properties of incremental learning algorithms provide an effective solution.
An incremental learning algorithm sequentially produces a new predictive model when new observations are available. The new predictive model is determined by the knowledge held in the previous model and by the information provided by the current data. Therefore, an incremental learning algorithm should be able to learn additional information from new data without completely forgetting its previous knowledge, improving at the same time the performance of the models in the course of time.
The present Thesis introduces the design, development, and evaluation of two new incremental learning algorithms for general purpose dynamic CDSS with an application to brain tumour diagnosis. Unlike many state-of-the-art incremental learning algorithms, we assume that previous data are not accessible at all, which is a common constraint in medical decision problems with distributed databases.
The first incremental learning algorithm is based on a generative weighted combination of maximum-likelihood estimators where the data are assumed to follow a multivariate Gaussian distribution. The algorithm has the ability to learn in an incremental fashion, improving the performance of the models when new information is available, and converg- ing in the course of time. Furthermore, it can incorporate new classes to its knowledge base if new diagnosis are available within the new data allowing the customization of the models at a particular clinical center. An evaluation using five benchmark databases has been used to characterize the behaviour of the algorithm and, finally, it has been applied to automatic brain tumour classification, comparing it with two state-of-the-art incremental algorithms.
The second algorithm is based on a discriminative logistic regression using a Bayesian inference paradigm where the posterior parameter distribution of one iteration is used as a prior parameter distribution for the training of the next model once the new data are available. This algorithm does not make any assumption on the underlying distribution of the data. The performance of the incremental algorithm is demonstrated by employ- ing different benchmark datasets and comparing it to the previous incremental learning algorithm using also the brain tumour database.
Both algorithms show a good behaviour obeying the definition of incremental learning algorithm and achieving the desired properties they should have. The algorithms have the ability to customize an already trained predictive model to the specific distribution of a particular hospital assuming that new information is ready for supervised classification at different times, without the need to access to the previously seen data. This ability to customize a model to a specific clinical centre could be used to improve the behaviour of a state-of-the-art CDSS for aiding brain tumour diagnosis in the future.