GHGeoffrey Holmes

Geoffrey Holmes

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Geoffrey Holmes has a total of 84 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Clinical Concept-Based Evaluation of Automatic Speech Recognition and Large Language Model Pipelines in Trauma Care (Preprint)

    by Renjun Cai, Geoffrey Holmes, Judy Bowen

    BACKGROUND Trauma workflows involve rapid and frequent verbal communication, multidisciplinary handovers, and extensive clinical documentation. Recent advances in automatic speech recognition (ASR) and large language models (LLMs) have shown potential for supporting clinical documentation, but many studies focus only on generic transcription accuracy rather than usefulness within clinically meaningful workflows, such as trauma handover. OBJECTIVE This study aimed to explore the feasibility of ASR with LLM pipelines to support trauma documentation workflows in a New Zealand trauma system. METHODS A modular ASR + LLM pipeline was developed in which ASR models transcribed clinical speech and LLMs generated structured clinical outputs. Multiple ASR and LLM combinations were evaluated separately and then combined into a pipeline, using public and curated clinical speech datasets. RESULTS AssemblyAI achieved the strongest overall ASR performance, with WER values of approximately 0.13 and consistently high concept-level F1 scores across datasets. However, clinically important transcription errors remained present in all ASR outputs. LLM-based refinement improved transcript readability and corrected a range of terminology and contextual errors. The greatest improvement was that GPT-5.5 increased F1 scores for clinical terms by up to 0.13. Refined transcripts were subsequently used to generate structured checklist items and support reference transcript curation, demonstrating potential applications within trauma documentation workflows. CONCLUSIONS The ASR + LLM pipeline demonstrated potential to support trauma documentation workflows within a New Zealand trauma system across diverse clinical speech datasets and multiple AI models evaluated in this study. By converting handover speech into clearer transcripts and generating structured documentation outputs, the pipeline may help clinicians with various documentation tasks. Rather than operating autonomously or replacing clinical decision-making, the pipeline is intended to serve as an assistant that supports clinician review, reduces documentation burden, and improves clarity in communication and handover between care teams.

  • Open Access

    Rethinking the Concept of Pixel Intensity Contrast From a Machine Learning Perspective

    by Sanush Abeysekera, Melanie Po-Leen Ooi, Ye-Chow Kuang, Shah Faisal, Yaminn Thawdar, Geoffrey Holmes, Dale Fletcher, Peter Reutemann

    Image contrast is a critical factor for machine vision tasks. A promising approach for enhancing contrast involves the use of algorithmically optimized, spectrally tunable illumination. However, the very definition of “contrast” is often rooted in principles of human perception, which may not be optimal for a machine observer. For an algorithm, contrast is an objective, task-driven metric that can be mathematically defined. To investigate the impact of this definition, we first use eigenvalue-based optimization algorithms to compute optimal illumination spectra. We then systematically evaluate these spectra using four distinct, physically realizable contrast formulations. Our analysis reveals that the performance of a given optimization algorithm is entirely dependent on the subsequent choice of evaluation metric. An illumination spectrum considered optimal under one metric can be significantly suboptimal when measured by another. This demonstrates that the choice of contrast metric is not a passive measurement, but an active design parameter with tangible physical consequences. From a machine learning perspective, the choice of this “loss function” should be codesigned with the physical hardware and the ultimate downstream task to achieve true system-level optimization.

  • Chemometric models for high-throughput biomass grading of the kelp Ecklonia radiata, using mid-infrared (MIR) and near-infrared (NIR) spectroscopy

    by Jacob Nepper-Davidsen, Marie Magnusson, Rebecca J. Lawton, Dale Fletcher, Geoffrey Holmes, Christopher R.K. Glasson


  • Open Access

    Feature Extractor Stacking for Cross-domain Few-shot Learning

    by Hongyu Wang, Eibe Frank, Bernhard Pfahringer, Michael Mayo, Geoffrey Holmes

    Cross-domain few-shot learning (CDFSL) addresses learning problems where knowledge needs to be transferred from one or more source domains into an instance-scarce target domain with an explicitly different distribution. Recently published CDFSL methods generally construct a universal model that combines knowledge of multiple source domains into one feature extractor. This enables efficient inference but necessitates re-computation of the extractor whenever a new source domain is added. Some of these methods are also incompatible with heterogeneous source domain extractor architectures. We propose feature extractor stacking (FES), a new CDFSL method for combining information from a collection of extractors, that can utilise heterogeneous pretrained extractors out of the box and does not maintain a universal model that needs to be re-computed when its extractor collection is updated. We present the basic FES algorithm, which is inspired by the classic stacked generalisation approach, and also introduce two variants: convolutional FES (ConFES) and regularised FES (ReFES). Given a target-domain task, these algorithms fine-tune each extractor independently, use cross-validation to extract training data for stacked generalisation from the support set, and learn a simple linear stacking classifier from this data. We evaluate our FES methods on the well-known Meta-Dataset benchmark, targeting image classification with convolutional neural networks, and show that they can achieve state-of-the-art performance.

  • Image Classification Using Class-Agnostic Object Detection

    by Geoffrey Holmes, Eibe Frank, Dale Fletcher

  • Open Access

    Sampling Permutations for Shapley Value Estimation

    by Rory Mitchell, Joshua Cooper, Eibe Frank, Geoffrey Holmes

    Game-theoretic attribution techniques based on Shapley values are used to interpret black-box machine learning models, but their exact calculation is generally NP-hard, requiring approximation methods for non-trivial models. As the computation of Shapley values can be expressed as a summation over a set of permutations, a common approach is to sample a subset of these permutations for approximation. Unfortunately, standard Monte Carlo sampling methods can exhibit slow convergence, and more sophisticated quasi-Monte Carlo methods have not yet been applied to the space of permutations. To address this, we investigate new approaches based on two classes of approximation methods and compare them empirically. First, we demonstrate quadrature techniques in a RKHS containing functions of permutations, using the Mallows kernel in combination with kernel herding and sequential Bayesian quadrature. The RKHS perspective also leads to quasi-Monte Carlo type error bounds, with a tractable discrepancy measure defined on permutations. Second, we exploit connections between the hypersphere $\mathbb{S}^{d-2}$ and permutations to create practical algorithms for generating permutation samples with good properties. Experiments show the above techniques provide significant improvements for Shapley value estimates over existing methods, converging to a smaller RMSE in the same number of model evaluations.


  • Open Access

    Bandwidth-Optimal Random Shuffling for GPUs

    by Rory Mitchell, Daniel Stokes, Eibe Frank, Geoffrey Holmes

    Linear-time algorithms that are traditionally used to shuffle data on CPUs, such as the method of Fisher-Yates, are not well suited to implementation on GPUs due to inherent sequential dependencies, and existing parallel shuffling algorithms are unsuitable for GPU architectures because they incur a large number of read/write operations to high latency global memory. To address this, we provide a method of generating pseudo-random permutations in parallel by fusing suitable pseudo-random bijective functions with stream compaction operations. Our algorithm, termed “bijective shuffle” trades increased per-thread arithmetic operations for reduced global memory transactions. It is work-efficient, deterministic, and only requires a single global memory read and write per shuffle input, thus maximising use of global memory bandwidth. To empirically demonstrate the correctness of the algorithm, we develop a statistical test for the quality of pseudo-random permutations based on kernel space embeddings. Experimental results show that the bijective shuffle algorithm outperforms competing algorithms on GPUs, showing improvements of between one and two orders of magnitude and approaching peak device bandwidth.


  • Overcoming Channel Uncertainties in Touchable Molecular Communication for Direct-Drug-Targeting-Assisted Immuno-Chemotherapy

    by Neda Sharifi, Yu Zhou, Geoffrey Holmes, Yifan Chen

    Objective: The performance of targeted immuno-chemotherapy of tumor is highly exposed to drug absorption in systemic circulation, which reduces its efficiency and increases side-effects. Direct drug targeting (DDT) combined with immuno-chemotherapy has the potential to mitigate the undesired systemic exposure, by using drug-loaded nanorobots to target cancer cells with the shortest possible physiological routes. This process can be modeled by the “touchable” (i.e., externally controllable and trackable) molecular communication system. However, such a complex process still suffers from various pharmacokinetic uncertainties caused by diffusion, degeneration, and branching of nanorobots (DDT pharmacokinetic uncertainties), as well as tumor/immune system modeling errors. The current work aims at identifying optimal drug administration plans by overcoming such challenges. Methods: A revisited tumor-immune interaction model is proposed to incorporate randomness of the drug concentration in the tumor site. Then, a robust multiple model predictive control (MMPC) scheme for the proposed tumor-immune interaction model is designed that uses multiple system models and an adaptive switch to identify the optimal plans for mixed drug administration via drug-loaded nanorobots. Furthermore, a wide range of prediction horizons under different loss scenarios of drug-loaded nanorobots and system model mismatches have been investigated in order to identify safe operating regions. From the molecular communications paradigm, this can be considered as a more robust information transmission system with feedback of channel state information to the transmitter implemented in the control unit. Results: The efficacy of the proposed MMPC is illustrated through identification of globally optimized drug administration schedules subject to various controller operation imperfections, which lead to successful cancer treatment as demonstrated through computational experiments. Conclusion: By combining DDT with conventional mixed immunotherapy and chemotherapy, the proposed robust MMPC offers promising performance in controlling tumor growth while keeping the immune cell density higher than a specific level in the presence of both DDT pharmacokinetic uncertainties and system model mismatches. Significance: We believe that the proposed design framework represents an important first step towards clinically relevant DDT in the combined immunotherapy and chemotherapy of tumor given its robust performance.


  • Open Access

    Probability Calibration Trees

    by Tim Leathart, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer

    Obtaining accurate and well calibrated probability estimates from classifiers is useful in many applications, for example, when minimising the expected cost of classifications. Existing methods of calibrating probability estimates are applied globally, ignoring the potential for improvements by applying a more fine-grained model. We propose probability calibration trees, a modification of logistic model trees that identifies regions of the input space in which different probability calibration models are learned to improve performance. We compare probability calibration trees to two widely used calibration methods---isotonic regression and Platt scaling---and show that our method results in lower root mean squared error on average than both methods, for estimates produced by a variety of base learners.

  • Environmental Data Science

    by Karina Gibert, Jeffery S. Horsburgh, Ioannis N. Athanasiadis, Geoff Holmes