Stephanie L. Hyland

Stephanie L. Hyland, Shruthi Bannur, Kenza Bouzid et al. · cambridge, microsoft-research

We present a radiology-specific multimodal model for the task for generating radiological reports from chest X-rays (CXRs). Our work builds on the idea that large language model(s) can be equipped with multimodal capabilities through alignment with pre-trained vision encoders. On natural images, this has been shown to allow multimodal models to gain image understanding and description capabilities. Our proposed model (MAIRA-1) leverages a CXR-specific image encoder in conjunction with a fine-tuned large language model based on Vicuna-7B, and text-based data augmentation, to produce reports with state-of-the-art quality. In particular, MAIRA-1 significantly improves on the radiologist-aligned RadCliQ metric and across all lexical metrics considered. Manual review of model outputs demonstrates promising fluency and accuracy of generated reports while uncovering failure modes not captured by existing evaluation practices. More information and resources can be found on the project website: https://aka.ms/maira.

Fangyu Liu, Qianchu Liu, Shruthi Bannur et al. · cambridge, deepmind

Label scarcity is a bottleneck for improving task performance in specialised domains. We propose a novel compositional transfer learning framework (DoT5 - domain compositional zero-shot T5) for zero-shot domain transfer. Without access to in-domain labels, DoT5 jointly learns domain knowledge (from MLM of unlabelled in-domain free text) and task knowledge (from task training on more readily available general-domain data) in a multi-task manner. To improve the transferability of task training, we design a strategy named NLGU: we simultaneously train NLG for in-domain label-to-data generation which enables data augmentation for self-finetuning and NLU for label prediction. We evaluate DoT5 on the biomedical domain and the resource-lean subdomain of radiology, focusing on NLI, text summarisation and embedding learning. DoT5 demonstrates the effectiveness of compositional transfer learning through multi-task learning. In particular, DoT5 outperforms the current SOTA in zero-shot transfer by over 7 absolute points in accuracy on RadNLI. We validate DoT5 with ablations and a case study demonstrating its ability to solve challenging NLI examples requiring in-domain expertise.

Dimitris Spathis, Stephanie L. Hyland · cambridge, microsoft-research

Clinical machine learning models show a significant performance drop when tested in settings not seen during training. Domain generalisation models promise to alleviate this problem, however, there is still scepticism about whether they improve over traditional training. In this work, we take a principled approach to identifying Out of Distribution (OoD) environments, motivated by the problem of cross-hospital generalization in critical care. We propose model-based and heuristic approaches to identify OoD environments and systematically compare models with different levels of held-out information. We find that access to OoD data does not translate to increased performance, pointing to inherent limitations in defining potential OoD environments potentially due to data harmonisation and sampling. Echoing similar results with other popular clinical benchmarks in the literature, new approaches are required to evaluate robust models on health records.

Anurag J. Vaidya, Felix Meissen, Daniel C. Castro et al.

Digitized histopathology analysis involves complex, time-intensive workflows and specialized expertise, limiting its accessibility. We introduce NOVA, an agentic framework that translates scientific queries into executable analysis pipelines by iteratively generating and running Python code. NOVA integrates 49 domain-specific tools (e.g., nuclei segmentation, whole-slide encoding) built on open-source software, and can also create new tools ad hoc. To evaluate such systems, we present SlideQuest, a 90-question benchmark -- verified by pathologists and biomedical scientists -- spanning data processing, quantitative analysis, and hypothesis testing. Unlike prior biomedical benchmarks focused on knowledge recall or diagnostic QA, SlideQuest demands multi-step reasoning, iterative coding, and computational problem solving. Quantitative evaluation shows NOVA outperforms coding-agent baselines, and a pathologist-verified case study links morphology to prognostically relevant PAM50 subtypes, demonstrating its scalable discovery potential.

Kenza Bouzid, Shruthi Bannur, Felix Meissen et al. · cambridge, microsoft-research

Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.

Fernando Pérez-García, Harshita Sharma, Sam Bond-Taylor et al.

Language-supervised pre-training has proven to be a valuable method for extracting semantically meaningful features from images, serving as a foundational element in multimodal systems within the computer vision and medical imaging domains. However, the computed features are limited by the information contained in the text, which is particularly problematic in medical imaging, where the findings described by radiologists focus on specific observations. This challenge is compounded by the scarcity of paired imaging-text data due to concerns over leakage of personal health information. In this work, we fundamentally challenge the prevailing reliance on language supervision for learning general-purpose biomedical imaging encoders. We introduce RAD-DINO, a biomedical image encoder pre-trained solely on unimodal biomedical imaging data that obtains similar or greater performance than state-of-the-art biomedical language-supervised models on a diverse range of benchmarks. Specifically, the quality of learned representations is evaluated on standard imaging tasks (classification and semantic segmentation), and a vision-language alignment task (text report generation from images). To further demonstrate the drawback of language supervision, we show that features from RAD-DINO correlate with other medical records (e.g., sex or age) better than language-supervised models, which are generally not mentioned in radiology reports. Finally, we conduct a series of ablations determining the factors in RAD-DINO's performance; notably, we observe that RAD-DINO's downstream performance scales well with the quantity and diversity of training data, demonstrating that image-only supervision is a scalable approach for training a foundational biomedical image encoder. Model weights of RAD-DINO trained on publicly available datasets are available at https://huggingface.co/microsoft/rad-dino.

Daniel C. Castro, Aurelia Bustos, Shruthi Bannur et al. · cambridge, microsoft-research

Radiology report generation (RRG) aims to create free-text radiology reports from clinical imaging. Grounded radiology report generation (GRRG) extends RRG by including the localisation of individual findings on the image. Currently, there are no manually annotated chest X-ray (CXR) datasets to train GRRG models. In this work, we present a dataset called PadChest-GR (Grounded-Reporting) derived from PadChest aimed at training GRRG models for CXR images. We curate a public bi-lingual dataset of 4,555 CXR studies with grounded reports (3,099 abnormal and 1,456 normal), each containing complete lists of sentences describing individual present (positive) and absent (negative) findings in English and Spanish. In total, PadChest-GR contains 7,037 positive and 3,422 negative finding sentences. Every positive finding sentence is associated with up to two independent sets of bounding boxes labelled by different readers and has categorical labels for finding type, locations, and progression. To the best of our knowledge, PadChest-GR is the first manually curated dataset designed to train GRRG models for understanding and interpreting radiological images and generated text. By including detailed localization and comprehensive annotations of all clinically relevant findings, it provides a valuable resource for developing and evaluating GRRG models from CXR images. PadChest-GR can be downloaded under request from https://bimcv.cipf.es/bimcv-projects/padchest-gr/

Anja Thieme, Abhijith Rajamohan, Benjamin Cooper et al. · cambridge, microsoft-research

Nasogastric tubes (NGTs) are feeding tubes that are inserted through the nose into the stomach to deliver nutrition or medication. If not placed correctly, they can cause serious harm, even death to patients. Recent AI developments demonstrate the feasibility of robustly detecting NGT placement from Chest X-ray images to reduce risks of sub-optimally or critically placed NGTs being missed or delayed in their detection, but gaps remain in clinical practice integration. In this study, we present a human-centered approach to the problem and describe insights derived following contextual inquiry and in-depth interviews with 15 clinical stakeholders. The interviews helped understand challenges in existing workflows, and how best to align technical capabilities with user needs and expectations. We discovered the trade-offs and complexities that need consideration when choosing suitable workflow stages, target users, and design configurations for different AI proposals. We explored how to balance AI benefits and risks for healthcare staff and patients within broader organizational and medical-legal constraints. We also identified data issues related to edge cases and data biases that affect model training and evaluation; how data documentation practices influence data preparation and labelling; and how to measure relevant AI outcomes reliably in future evaluations. We discuss how our work informs design and development of AI applications that are clinically useful, ethical, and acceptable in real-world healthcare services.

Harshita Sharma, Valentina Salvatelli, Shaury Srivastav et al.

There is growing interest in applying AI to radiology report generation, particularly for chest X-rays (CXRs). This paper investigates whether incorporating pixel-level information through segmentation masks can improve fine-grained image interpretation of multimodal large language models (MLLMs) for radiology report generation. We introduce MAIRA-Seg, a segmentation-aware MLLM framework designed to utilize semantic segmentation masks alongside CXRs for generating radiology reports. We train expert segmentation models to obtain mask pseudolabels for radiology-specific structures in CXRs. Subsequently, building on the architectures of MAIRA, a CXR-specialised model for report generation, we integrate a trainable segmentation tokens extractor that leverages these mask pseudolabels, and employ mask-aware prompting to generate draft radiology reports. Our experiments on the publicly available MIMIC-CXR dataset show that MAIRA-Seg outperforms non-segmentation baselines. We also investigate set-of-marks prompting with MAIRA and find that MAIRA-Seg consistently demonstrates comparable or superior performance. The results confirm that using segmentation masks enhances the nuanced reasoning of MLLMs, potentially contributing to better clinical outcomes.

Shruthi Bannur, Kenza Bouzid, Daniel C. Castro et al.

Radiology reporting is a complex task requiring detailed medical image understanding and precise language generation, for which generative multimodal models offer a promising solution. However, to impact clinical practice, models must achieve a high level of both verifiable performance and utility. We augment the utility of automated report generation by incorporating localisation of individual findings on the image - a task we call grounded report generation - and enhance performance by incorporating realistic reporting context as inputs. We design a novel evaluation framework (RadFact) leveraging the logical inference capabilities of large language models (LLMs) to quantify report correctness and completeness at the level of individual sentences, while supporting the new task of grounded reporting. We develop MAIRA-2, a large radiology-specific multimodal model designed to generate chest X-ray reports with and without grounding. MAIRA-2 achieves state of the art on existing report generation benchmarks and establishes the novel task of grounded report generation.

Matthias Hüser, Martin Faltys, Xinrui Lyu et al.

The development of respiratory failure is common among patients in intensive care units (ICU). Large data quantities from ICU patient monitoring systems make timely and comprehensive analysis by clinicians difficult but are ideal for automatic processing by machine learning algorithms. Early prediction of respiratory system failure could alert clinicians to patients at risk of respiratory failure and allow for early patient reassessment and treatment adjustment. We propose an early warning system that predicts moderate/severe respiratory failure up to 8 hours in advance. Our system was trained on HiRID-II, a data-set containing more than 60,000 admissions to a tertiary care ICU. An alarm is typically triggered several hours before the beginning of respiratory failure. Our system outperforms a clinical baseline mimicking traditional clinical decision-making based on pulse-oximetric oxygen saturation and the fraction of inspired oxygen. To provide model introspection and diagnostics, we developed an easy-to-use web browser-based system to explore model input data and predictions visually.

Emily Alsentzer, Matthew B. A. McDermott, Fabian Falck et al.

A collection of the accepted abstracts for the Machine Learning for Health (ML4H) workshop at NeurIPS 2020. This index is not complete, as some accepted abstracts chose to opt-out of inclusion.

Stephanie L. Hyland, Shruti Tople

Introducing noise in the training of machine learning systems is a powerful way to protect individual privacy via differential privacy guarantees, but comes at a cost to utility. This work looks at whether the inherent randomness of stochastic gradient descent (SGD) could contribute to privacy, effectively reducing the amount of \emph{additional} noise required to achieve a given privacy guarantee. We conduct a large-scale empirical study to examine this question. Training a grid of over 120,000 models across four datasets (tabular and images) on convex and non-convex objectives, we demonstrate that the random seed has a larger impact on model weights than any individual training example. We test the distribution over weights induced by the seed, finding that the simple convex case can be modelled with a multivariate Gaussian posterior, while neural networks exhibit multi-modal and non-Gaussian weight distributions. By casting convex SGD as a Gaussian mechanism, we then estimate an `intrinsic' data-dependent $ε_i(\mathcal{D})$, finding values as low as 6.3, dropping to 1.9 using empirical estimates. We use a membership inference attack to estimate $ε$ for non-convex SGD and demonstrate that hiding the random seed from the adversary results in a statistically significant reduction in attack performance, corresponding to a reduction in the effective $ε$. These results provide empirical evidence that SGD exhibits appreciable variability relative to its dataset sensitivity, and this `intrinsic noise' has the potential to be leveraged to improve the utility of privacy-preserving machine learning.

Stefan G. Stark, Stephanie L. Hyland, Melanie F. Pradier et al.

The recent adoption of Electronic Health Records (EHRs) by health care providers has introduced an important source of data that provides detailed and highly specific insights into patient phenotypes over large cohorts. These datasets, in combination with machine learning and statistical approaches, generate new opportunities for research and clinical care. However, many methods require the patient representations to be in structured formats, while the information in the EHR is often locked in unstructured texts designed for human readability. In this work, we develop the methodology to automatically extract clinical features from clinical narratives from large EHR corpora without the need for prior knowledge. We consider medical terms and sentences appearing in clinical narratives as atomic information units. We propose an efficient clustering strategy suitable for the analysis of large text corpora and to utilize the clusters to represent information about the patient compactly. To demonstrate the utility of our approach, we perform an association study of clinical features with somatic mutation profiles from 4,007 cancer patients and their tumors. We apply the proposed algorithm to a dataset consisting of about 65 thousand documents with a total of about 3.2 million sentences. We identify 341 significant statistical associations between the presence of somatic mutations and clinical features. We annotated these associations according to their novelty, and report several known associations. We also propose 32 testable hypotheses where the underlying biological mechanism does not appear to be known but plausible. These results illustrate that the automated discovery of clinical features is possible and the joint analysis of clinical and genetic datasets can generate appealing new hypotheses.

Stephanie L. Hyland, Martin Faltys, Matthias Hüser et al.

Intensive care clinicians are presented with large quantities of patient information and measurements from a multitude of monitoring systems. The limited ability of humans to process such complex information hinders physicians to readily recognize and act on early signs of patient deterioration. We used machine learning to develop an early warning system for circulatory failure based on a high-resolution ICU database with 240 patient years of data. This automatic system predicts 90.0% of circulatory failure events (prevalence 3.1%), with 81.8% identified more than two hours in advance, resulting in an area under the receiver operating characteristic curve of 94.0% and area under the precision-recall curve of 63.0%. The model was externally validated in a large independent patient cohort.

Xinrui Lyu, Matthias Hueser, Stephanie L. Hyland et al.

In this work, we investigate unsupervised representation learning on medical time series, which bears the promise of leveraging copious amounts of existing unlabeled data in order to eventually assist clinical decision making. By evaluating on the prediction of clinically relevant outcomes, we show that in a practical setting, unsupervised representation learning can offer clear performance benefits over end-to-end supervised architectures. We experiment with using sequence-to-sequence (Seq2Seq) models in two different ways, as an autoencoder and as a forecaster, and show that the best performance is achieved by a forecasting Seq2Seq model with an integrated attention mechanism, proposed here for the first time in the setting of unsupervised learning for medical time series.

Cristóbal Esteban, Stephanie L. Hyland, Gunnar Rätsch

Generative Adversarial Networks (GANs) have shown remarkable success as a framework for training models to produce realistic-looking data. In this work, we propose a Recurrent GAN (RGAN) and Recurrent Conditional GAN (RCGAN) to produce realistic real-valued multi-dimensional time series, with an emphasis on their application to medical data. RGANs make use of recurrent neural networks in the generator and the discriminator. In the case of RCGANs, both of these RNNs are conditioned on auxiliary information. We demonstrate our models in a set of toy datasets, where we show visually and quantitatively (using sample likelihood and maximum mean discrepancy) that they can successfully generate realistic time-series. We also describe novel evaluation methods for GANs, where we generate a synthetic labelled training dataset, and evaluate on a real test set the performance of a model trained on the synthetic data, and vice-versa. We illustrate with these metrics that RCGANs can generate time-series data useful for supervised training, with only minor degradation in performance on real test data. This is demonstrated on digit classification from 'serialised' MNIST and by training an early warning system on a medical dataset of 17,000 patients from an intensive care unit. We further discuss and analyse the privacy concerns that may arise when using RCGANs to generate realistic synthetic medical time series data.

Paulina Grnarova, Florian Schmidt, Stephanie L. Hyland et al.

We present an automatic mortality prediction scheme based on the unstructured textual content of clinical notes. Proposing a convolutional document embedding approach, our empirical investigation using the MIMIC-III intensive care database shows significant performance gains compared to previously employed methods such as latent topic distributions or generic doc2vec embeddings. These improvements are especially pronounced for the difficult problem of post-discharge mortality prediction.

Stephanie L. Hyland, Gunnar Rätsch

A major challenge in the training of recurrent neural networks is the so-called vanishing or exploding gradient problem. The use of a norm-preserving transition operator can address this issue, but parametrization is challenging. In this work we focus on unitary operators and describe a parametrization using the Lie algebra $\mathfrak{u}(n)$ associated with the Lie group $U(n)$ of $n \times n$ unitary matrices. The exponential map provides a correspondence between these spaces, and allows us to define a unitary matrix using $n^2$ real coefficients relative to a basis of the Lie algebra. The parametrization is closed under additive updates of these coefficients, and thus provides a simple space in which to do gradient descent. We demonstrate the effectiveness of this parametrization on the problem of learning arbitrary unitary operators, comparing to several baselines and outperforming a recently-proposed lower-dimensional parametrization. We additionally use our parametrization to generalize a recently-proposed unitary recurrent neural network to arbitrary unitary matrices, using it to solve standard long-memory tasks.

Stephanie L. Hyland, Theofanis Karaletsos, Gunnar Rätsch

Identifying relationships between concepts is a key aspect of scientific knowledge synthesis. Finding these links often requires a researcher to laboriously search through scien- tific papers and databases, as the size of these resources grows ever larger. In this paper we describe how distributional semantics can be used to unify structured knowledge graphs with unstructured text to predict new relationships between medical concepts, using a probabilistic generative model. Our approach is also designed to ameliorate data sparsity and scarcity issues in the medical domain, which make language modelling more challenging. Specifically, we integrate the medical relational database (SemMedDB) with text from electronic health records (EHRs) to perform knowledge graph completion. We further demonstrate the ability of our model to predict relationships between tokens not appearing in the relational database.

Stephanie L. Hyland, Theofanis Karaletsos, Gunnar Rätsch

Neural language models are a powerful tool to embed words into semantic vector spaces. However, learning such models generally relies on the availability of abundant and diverse training examples. In highly specialised domains this requirement may not be met due to difficulties in obtaining a large corpus, or the limited range of expression in average use. Such domains may encode prior knowledge about entities in a knowledge base or ontology. We propose a generative model which integrates evidence from diverse data sources, enabling the sharing of semantic information. We achieve this by generalising the concept of co-occurrence from distributional semantics to include other relationships between entities or words, which we model as affine transformations on the embedding space. We demonstrate the effectiveness of this approach by outperforming recent models on a link prediction task and demonstrating its ability to profit from partially or fully unobserved data training labels. We further demonstrate the usefulness of learning from different data sources with overlapping vocabularies.