Sheng Wong

Isabel Tsintsiper, Sheng Wong, Beth Albert et al.

Large language models (LLMs) are increasingly embedded in healthcare workflows for documentation, education, and clinical decision support. However, these systems are trained on large text corpora that encode existing biases, including sex disparities in diagnosis and treatment, raising concerns that such patterns may be reproduced or amplified. We systematically examined whether contemporary LLMs exhibit sex-specific biases in clinical reasoning and how model configuration influences these behaviours. We conducted three experiments using 50 clinician-authored vignettes spanning 44 specialties in which sex was non-informative to the initial diagnostic pathway. Four general-purpose LLMs (ChatGPT (gpt-4o-mini), Claude 3.7 Sonnet, Gemini 2.0 Flash and DeepSeekchat). All models demonstrated significant sex-assignment skew, with predicted sex differing by model. At temperature 0.5, ChatGPT assigned female sex in 70% of cases (95% CI 0.66-0.75), DeepSeek in 61% (0.57-0.65) and Claude in 59% (0.55-0.63), whereas Gemini showed a male skew, assigning a female sex in 36% of cases (0.32-0.41). Contemporary LLMs exhibit stable, model-specific sex biases in clinical reasoning. Permitting abstention reduces explicit labelling but does not eliminate downstream diagnostic differences. Safe clinical integration requires conservative and documented configuration, specialty-level clinical data auditing, and continued human oversight when deploying general-purpose models in healthcare settings.

Sheng Wong, Ravi Shankar, Beth Albert et al.

Supervised deep learning models for automated CTG analysis are typically constrained by narrowly curated labelled datasets and limited patient cohorts, leaving substantial volumes of physiologically informative clinical recordings untapped. To address this limitation, we propose Physiology-aware Representation Learning via Integrated Self-supervision and Metadata for CTG (PRISM-CTG), a clinically grounded self-supervised foundation model (FM) for CTG that leverages large-scale unlabelled recordings to learn transferable domain-level representations. PRISM-CTG is pretrained using a multi-view self-supervised framework that jointly optimises 3 complementary pretext objectives: random-projected guided masked signal reconstruction, clinical variable prediction, and feature classification. Each objective is associated with a dedicated task-specific token, enabling specialised representation learning, while controlled cross-attention facilitates information exchange across clinical context. By reframing patient metadata and domain knowledge, which are often underutilised in conventional training as prediction targets, Prism-CTG transforms readily available clinical information into additional supervisory targets that guide clinically meaningful representation learning. Extensive experiments across 7 downstream CTG tasks in both antepartum and intrapartum domains demonstrated that PRISM-CTG consistently outperforms in-domain and SSL baselines. Notably, PRISM-CTG demonstrated strong generalisation under external validation on 2 datasets, while achieving comparable performance to studies trained on substantially larger, privately labelled datasets. To our knowledge, this is the first study to introduce large-scale FM for CTG that learns domain-level representations.

Sheng Wong, Beth Albert, Gabriel Davis Jones

Cardiotocography (CTG) is essential for fetal monitoring but is frequently compromised by diverse artefacts which obscure true fetal heart rate (FHR) patterns and can lead to misdiagnosis or delayed intervention. Current deep-learning approaches typically bypass comprehensive noise handling, applying minimal preprocessing or focusing solely on downstream classification, while traditional methods rely on simple interpolation or rule-based filtering that addresses only missing samples and fail to correct complex artefact types. We present CleanCTG, an end-to-end dual-stage model that first identifies multiple artefact types via multi-scale convolution and context-aware cross-attention, then reconstructs corrupted segments through artefact-specific correction branches. Training utilised over 800,000 minutes of physiologically realistic, synthetically corrupted CTGs derived from expert-verified "clean" recordings. On synthetic data, CleanCTG achieved perfect artefact detection (AU-ROC = 1.00) and reduced mean squared error (MSE) on corrupted segments to 2.74 x 10^-4 (clean-segment MSE = 2.40 x 10^-6), outperforming the next best method by more than 60%. External validation on 10,190 minutes of clinician-annotated segments yielded AU-ROC = 0.95 (sensitivity = 83.44%, specificity 94.22%), surpassing six comparator classifiers. Finally, when integrated with the Dawes-Redman system on 933 clinical CTG recordings, denoised traces increased specificity (from 80.70% to 82.70%) and shortened median time to decision by 33%. These findings suggest that explicit artefact removal and signal reconstruction can both maintain diagnostic accuracy and enable shorter monitoring sessions, offering a practical route to more reliable CTG interpretation.

Sheng Wong, Ravi Shankar, Beth Albert et al.

Foundation models (FMs) and large language models (LLMs) have demonstrated promising generalization across diverse domains for time-series analysis, yet their potential for electronic fetal monitoring (EFM) and cardiotocography (CTG) analysis remains underexplored. Most existing CTG studies relied on domain-specific models and lack systematic comparisons with modern foundation or language models, limiting our understanding of whether these models can outperform specialized systems in fetal health assessment. In this study, we present the first comprehensive benchmark of state-of-the-art architectures for automated antepartum CTG classification. Over 2,500 20-minutes recordings were used to evaluate over 15 models spanning domain-specific, time-series, foundation, and language-model categories under a unified framework. Fine-tuned LLMs consistently outperformed both foundation and domain-specific models across data-availability scenarios, except when uterine-activity signals were absent, where domain-specific models showed greater robustness. These performance gains, however, required substantially higher computational resources. Our results highlight that while fine-tuned LLMs achieved state-of-the-art performance for CTG classification, practical deployment must balance performance with computational efficiency.

Ravi Shankar, Sheng Wong, Lin Li et al.

Reliable abstention is critical for retrieval-augmented generation (RAG) systems, particularly in safety-critical domains such as women's health, where incorrect answers can lead to harm. We present an energy-based model (EBM) that learns a smooth energy landscape over a dense semantic corpus of 2.6M guideline-derived questions, enabling the system to decide when to generate or abstain. We benchmark the EBM against a calibrated softmax baseline and a k-nearest neighbour (kNN) density heuristic across both easy and hard abstention splits, where hard cases are semantically challenging near-distribution queries. The EBM achieves superior abstention performance abstention on semantically hard cases, reaching AUROC 0.961 versus 0.950 for softmax, while also reducing FPR@95 (0.235 vs 0.331). On easy negatives, performance is comparable across methods, but the EBM's advantage becomes most pronounced in safety-critical hard distributions. A comprehensive ablation with controlled negative sampling and fair data exposure shows that robustness stems primarily from the energy scoring head, while the inclusion or exclusion of specific negative types (hard, easy, mixed) sharpens decision boundaries but is not essential for generalisation to hard cases. These results demonstrate that energy-based abstention scoring offers a more reliable confidence signal than probability-based softmax confidence, providing a scalable and interpretable foundation for safe RAG systems.

Scott Barnett, Zac Brannelly, Stefanus Kurniawan et al.

Large Language Models (LLMs) have the unique capability to understand and generate human-like text from input queries. When fine-tuned, these models show enhanced performance on domain-specific queries. OpenAI highlights the process of fine-tuning, stating: "To fine-tune a model, you are required to provide at least 10 examples. We typically see clear improvements from fine-tuning on 50 to 100 training examples, but the right number varies greatly based on the exact use case." This study extends this concept to the integration of LLMs within Retrieval-Augmented Generation (RAG) pipelines, which aim to improve accuracy and relevance by leveraging external corpus data for information retrieval. However, RAG's promise of delivering optimal responses often falls short in complex query scenarios. This study aims to specifically examine the effects of fine-tuning LLMs on their ability to extract and integrate contextual data to enhance the performance of RAG systems across multiple domains. We evaluate the impact of fine-tuning on the LLMs' capacity for data extraction and contextual understanding by comparing the accuracy and completeness of fine-tuned models against baseline performances across datasets from multiple domains. Our findings indicate that fine-tuning resulted in a decline in performance compared to the baseline models, contrary to the improvements observed in standalone LLM applications as suggested by OpenAI. This study highlights the need for vigorous investigation and validation of fine-tuned models for domain-specific tasks.