There is currently a lack of an efficient, objective and systemic approach towards the classification of Alzheimer’s disease (AD), due to its complex etiology and pathogenesis. As AD is inherently dynamic, it is also not clear how the relationships among AD indicators vary over time. To address these issues, we propose a hybrid computational approach for AD classification and evaluate it on the heterogeneous longitudinal AIBL dataset. Specifically, using clinical dementia rating as an index of AD severity, the most important indicators (mini-mental state examination, logical memory recall, grey matter and cerebrospinal volumes from MRI and active voxels from PiB-PET brain scans, ApoE, and age) can be automatically identified from parallel data mining algorithms. In this work, Bayesian network modelling across different time points is used to identify and visualize time-varying relationships among the significant features, and importantly, in an efficient way using only coarse-grained data. Crucially, our approach suggests key data features and their appropriate combinations that are relevant for AD severity classification with high accuracy. Overall, our study provides insights into AD developments and demonstrates the potential of our approach in supporting efficient AD diagnosis.

Accurate computational models for clinical decision support systems require clean and reliable data but, in clinical practice, data are often incomplete. Hence, missing data could arise not only from training datasets but also test datasets which could consist of a single undiagnosed case, an individual. This work addresses the problem of extreme missingness in both training and test data by evaluating multiple imputation and classification workflows based on both diagnostic classification accuracy and computational cost. Extreme missingness is defined as having ∼50% of the total data missing in more than half the data features. In particular, we focus on dementia diagnosis due to long time delays, high variability, high attrition rates and lack of practical data imputation strategies in its diagnostic pathway. We identified and replicated the extreme missingness structure of data from a real-world memory clinic on a larger open dataset, with the original complete data acting as ground truth. Overall, we found that computational cost, but not accuracy, varies widely for various imputation and classification approaches. Particularly, we found that iterative imputation on the training dataset combined with a reduced-feature classification model provides the best approach, in terms of speed and accuracy. Taken together, this work has elucidated important factors to be considered when developing a predictive model for a dementia diagnostic support system.

Limited training data, high dimensionality, image complexity, and similarity between classes are challenges confronting hyperspectral image (HSI) classification often resulting in suboptimal classification performance. The capsule network (CapsNet) preserves the hierarchy between different parts of the entity in an image by replacing scalar representations with vectors and can address these aforementioned issues. Motivated by CapsNet, this article presents a novel end-to-end deep learning (DL) architecture, the hybrid capsule network (HCapsNet), for HSI classification. HCapsNet employs 2-D and 3-D convolutional neural networks (CNNs) to extract higher level spatial and spectral features. In order to establish a route between capsules in the lower layers to the most-related capsule in the higher layer, dynamic routing (DR) is used to identify several overlapped objects during training sessions. Hyperparameter optimization is performed using nested cross-validation (nested-CV) to ensure thorough generalization evaluation. The proposed HCapsNet significantly outperformed the state-of-the-art methods in terms of overall classification accuracy on three widely used hyperspectral datasets, Indian Pines dataset achieving ( $>3\%$ , $p< {1}\times 1\times 10^{-11}$ ), the University of Pavia dataset ( $>4\%$ , $p< {1}\times 1\times 10^{-9}$ ), the Salinas Valley dataset ( $>3\%$ , $p< {1}\times 1\times 10^{-10}$ ) when using only 1% of the data for training. The performance of all CNN-based approaches degraded significantly with smaller training sample sizes. HCapsNet, therefore, is demonstrated to offer significant advantages in HSI classification problems with low sample sizes.

Objective: Despite the potential of machine learning techniques to improve dementia diagnostic processes, research outcomes are often not readily translated to or adopted in clinical practice. Importantly, the time taken to administer diagnostic assessment has yet to be taken into account in feature-selection based optimisation for dementia diagnosis. We address these issues by considering the impact of assessment time as a practical constraint for feature selection of cognitive and functional assessments in Alzheimer’s disease diagnosis. Methods: We use three different feature selection algorithms to select informative subsets of dementia assessment items from a large open-source dementia dataset. We use cost-sensitive feature selection to optimise our feature selection results for assessment time as well as diagnostic accuracy. To encourage clinical adoption and further evaluation of our proposed accuracy-vs-cost optimisation algorithms, we also implement a sandbox-like toolbox with graphical user interface to evaluate user-chosen subsets of assessment items. Results: We find that there are subsets of accuracy-cost optimised assessment items that can perform better in terms of diagnostic accuracy and/or total assessment time than most other standard assessments. Discussion: Overall, our analysis and accompanying sandbox tool can facilitate clinical users and other stakeholders to apply their own domain knowledge to analyse and decide which dementia diagnostic assessment items are useful, and aid the redesigning of dementia diagnostic assessments. Clinical Impact (Clinical Research): By optimising diagnostic accuracy and assessment time, we redesign predictive and efficient dementia diagnostic assessments and develop a sandbox interface to facilitate evaluation and testing by clinicians and non-specialists.

Multivariate batch time-series data sets within Semiconductor manufacturing processes present a difficult environment for effective Anomaly Detection (AD). The challenge is amplified by the limited availability of ground truth labelled data. In scenarios where AD is possible, black box modelling approaches constrain model interpretability. These challenges obstruct the widespread adoption of Deep Learning solutions. The objective of the study is to demonstrate an AD approach which employs 1-Dimensional Convolutional AutoEncoders (1d-CAE) and Localised Reconstruction Error (LRE) to improve AD performance and interpretability. Using LRE to identify sensors and data that result in the anomaly, the explainability of the Deep Learning solution is enhanced. The Tennessee Eastman Process (TEP) and LAM 9600 Metal Etcher datasets have been utilised to validate the proposed framework. The results show that the proposed LRE approach outperforms global reconstruction errors for similar model architectures achieving an AUC of 1.00. The proposed unsupervised learning approach with AE and LRE improves model explainability which is expected to be beneficial for deployment in semiconductor manufacturing where interpretable and trustworthy results are critical for process engineering teams.