Artificial Intelligence for Diabetes Technology

Macintosh HD:Users:mmayo:Desktop:11 .png

Diabetes is prevalent in New Zealand, affecting an estimated 10.2% (7.8%-12.6%) of the population with one quarter of those cases undiagnosed (IDF Diabetes Atlas, 2017 figures). Prevalence is forecast to rise by 2045. Technological advances such as wearable glucose monitors and subcutaneous insulin infusion pumps (collectively known as closed loop artificial pancreas systems) are useful tools for improving the management of diabetes and have demonstrated positive patient outcomes. However, these devices (consisting of complex hardware and software) generate significant volumes of data requiring analysis and also introduce patient security, safety and privacy issues. In this project we use AI techniques such as neural networks, machine learning and heuristic optimisation to tackle these diabetes-related problems.


  • Jade Myers (Computer Science)
  • Kriti Narsapur (Computer Science)
  • Hongyu Wang (Computer Science)
  • Abagail Koay (Computer Science)
  • Panos Patros (Computer Science)
  • Lynne Chepulis (Waikato Medical Research Centre)
  • Ryan Paul (Waikato Medical Research Centre/Waikato District Health Board)
  • Michael Mayo (Computer Science)


  • Insulin infusion rate protocol optimisation using AI techniques
  • Outlier/anomaly detection in continuous glucose monitor/insulin pump data for identifying security and safety events
  • Blood glucose level (BGL) prediction from past BGLs and other physiological data
  • Estimating the risk of hypoglycemia and hyperglycemia, as well as general glycemic variability estimation techniques using machine learning
  • Mortality risk prediction from electronic health records using neural nets
  • Intelligent self-adaptation techniques for wearable devices to account for patient differences
  • Privacy issues, such as patient re-identification risk from glucose monitor data


Mayo, M. (2019). Improving the robustness of the glycemic variability percentage metric to sensor dropouts in continuous glucose monitor data. In N. T. Nguyen, F. L. Gaol, T. P. Hong, & B. Trawinski (Eds.), Intelligent Information and Database Systems. ACIIDS 2019. Lecture Notes in Computer Science Vol. 11432 (pp. 373-384). Cham: Springer. doi:10.1007/978-3-030-14802-7_32

Mayo, M., & Yogarajan, V. (2019). A nearest neighbour-based analysis to identify patients from continuous glucose monitor data. In N. T. Nguyen, F. L. Gaol, T. P. Hong, & B. Trawinski (Eds.), Intelligent Information and Database Systems. ACIIDS 2019. Lecture Notes in Computer Science Vol. 11432 (pp. 349-360). Cham: Springer. doi:10.1007/978-3-030-14802-7_30

Detecting Non-obvious Neuroimaging Abnormalities using Deep Learning-based Generative Models

To help computers understand the real world, computer scientists invented generative models to to produce a representation or abstraction of observed phenomena or target variables that can be calculated from observations. The intuition behind this approach follows a famous quote from Richard Feynman: “What I cannot create, I do not understand.” The neuroimaging is a complex challenge by its own nature. The hypothesis of my research is that using reconstruction error of generative models will provide a means to detect neuro-degenerative abnormalities in neuroimages that are difficult for humans to recognise, and may be indicative of neurodegenerative diseases in the very early stage or even presymptomatic stage.


The general objective of this research is to show that reconstruction loss of deep learning based generative models can be as a measurement to detect non-obvious anomalies in neuroimages. There are specific objectives that have been identified:

  • Design and implement an anomaly detector based on a deep learning-based
    generative model, in which the model is only trained on healthy images with minimum reconstruction error. Then, verify whether if this trained AE will reconstruct unseen AD sMRI images with higher reconstruction error.
  • Improve the anomaly detector so that it is capable of identifying different types of
    anomalies from the major diagnostic stages of Alzheimer’s Disease including normal, early and late stages.
  • Improve the anomaly detector to be able to highlight regions that have higher
    reconstruction error.


  • Chen Zheng
  • Bernhard Pfahringer
  • Michael Mayo

Neural networks for gene expression data analysis

Neural networks are powerful tools widely used for  extracting useful information from complex data. Microarrays are a widely used technology for analyzing genetic diseases by  measuring the level of activity of genes within a given tissue. Microarray datasets typically consist of thousands of gene expressions but only a few hundred samples.  These characteristics pose a challenge to machine learning algorithms and adversely impact their prediction accuracy and running time. This project aims at building neural network  models which are able to efficiently learn representations of gene expression data for classification, clustering, and other statistical analysis.


This project started with investigating the latest approaches in building predictive models from gene expression data. The study revealed that researchers try to deal with the class imbalance and the class overlap problems  as part of the other problems that gene expression datasets have. This project focuses on developing techniques that particularly suit high dimensional data, more specifically gene expressions, to improve the prediction performance of the models under investigation. The project aims at:

  • Developing techniques that engineers the gene expressions  e.g. solves the class imbalance problem.
  • Developing  supervised/ unsupervised neural network models that are able to learn efficiently and effectively from gene expression data.


  • Michael Mayo
  • Maisa Daoud


  • Maisa Daoud,  Michael Mayo, and Sally Jo Cunningham.  “RBFA: Radial Basis Function Autoencoders”,  IEEE Congress on Evolutionary Computation 2019 (accepted)
  • Maisa Daoud, and Michael Mayo. “A Novel Synthetic Over-Sampling Technique for Imbalanced Classification of Gene Expressions Using Autoencoders and Swarm Optimization.” Australasian Joint Conference on Artificial Intelligence. Springer, Cham, 2018.
  • Maisa Daoud, and Michael Mayo. “A Survey Of Neural Network-based Cancer Prediction Models From Microarray Data.” Artificial Intelligence in Medicine (2019),