Practical machine learning approaches for decision support in dementia diagnosis

  • Niamh McCombe

    Student thesis: Doctoral Thesis


    This Ph.D. thesis contributes towards the development of practical applications of data science that may be utilised to improve the dementia care pathway in the presence of real-world constraints. The thesis begins with an overview of the dementia care pathway, with discussion of the gaps between recommendations and what occurs in routine clinical practice. This is followed by a discussion of machine learning techniques that have been applied to dementia, and a review of related computational techniques used in the thesis.

    The thesis has led to four original research contributions. The first contribution deals with the widespread problem of extreme missingness in training and test data in clinical settings. Imputation and classification workflows are evaluated on an Alzheimer’s disease dataset of cognitive and functional assessments, in order to find the most useful workflow for a dementia diagnostic support system. The results support the use of iterative imputation on the training dataset combined with a reduced-feature classification model. In the second contribution, the problem of evaluating the accuracy of data imputation is addressed. A highly linear relationship is found between the principal component loadings and assessment feature imputability, allowing imputation accuracy to be estimated even where the correct value for missing data is unknown. The third and fourth contributions use clinician assessment time as a practical cost constraint for feature selection from assessments used in dementia diagnosis, generating accuracy-time optimised subsets of assessment items that perform better than other standard assessments. The cost optimization algorithm is then generalized for multiple types of costs. To encourage clinical engagement and adoption, graphical user interfaces are created for the algorithms.

    Overall, the thesis provides machine learning solutions for use in real-world clinical settings that have been optimised for practicality, in terms of addressing the key challenges of diagnostic accuracy, computational cost, and available clinical assessment time.
    Date of AwardFeb 2023
    Original languageEnglish
    SupervisorKongfatt Wong-Lin (Supervisor), Paula McClean (Supervisor), Girijesh Prasad (Supervisor) & Xuemei Ding (Supervisor)


    • Dementia
    • Data science
    • Feature selection
    • Cost-sensitive feature selection
    • ADNI
    • Healthcare costs
    • Cognitive and functional assessments

    Cite this