← RECAI Labs
RESEARCH
What we’re publishing.
Papers and datasets our team has authored, across clinical AI, AI safety, NLP, speech and causal inference. Each one opens in a new tab at the publisher.
Filter
Large Language Models for Optimizing Patient Recruitment Decisions in Voice Data Generation Projects
Clinical models inherit the imbalances of the data they were trained on, and for voice datasets that imbalance often starts at recruitment. A generative-AI pipeline that recommends recruitment decisions up front, rather than correcting the skew afterwards.
Overview of the 11th Social Media Mining for Health (#SMM4H) and Health Real-World Data (HeaRD) Shared Tasks at ACL 2026
The overview of the eleventh #SMM4H and HeaRD shared tasks: what each task asked of its entrants, which systems took part, and how they scored on health signals drawn from social media and real-world data.
SurgiDiff: A Context-Aware Diffusion Recommender for Safe Surgical Autonomy
An ensemble of diffusion models generates a range of possible trajectories, each carrying its own uncertainty, and a language model judges them on motion quality and task success before one is executed. Tested on needle reach, needle pick, gauze retrieval and peg transfer.
Parkinson’s Disease Drug Dose Optimization Using Multi-Objective Reinforcement Learning
Controlling motor symptoms and avoiding dyskinesia are competing aims, and most reinforcement learning in healthcare collapses them into one score fixed in advance. This models both at once and returns a set of policies showing the trade-off instead.
Interpreting Agentic Systems: Beyond Model Explanations to System-Level Accountability
Interpretability tools were built for static models. This assesses what they can and cannot tell you about systems that plan over many steps and act on an environment, and where the gaps are.
How to Recover Long Audio Sequences Through Gradient Inversion Attack With Dynamic Segment-based Reconstruction
Gradient inversion recovers training data from the updates a model exchanges during training. This extends the attack to long audio sequences, reconstructing them segment by segment rather than in one fixed-length pass.
MedSynth: Realistic, Synthetic Medical Dialogue-Note Pairs
Over 10,000 synthetic dialogue–note pairs spanning more than 2,000 ICD-10 codes, built so models can be trained on clinical documentation without patient data. Code and dataset are open.
Locking Open Weight Models with Spectral Deformation
Releasing weights openly also releases the ability to fine-tune them into something harmful. Manipulating a model’s spectral properties makes further training arbitrarily slow while leaving its original behaviour intact.
Comprehensive descriptive analysis of large Alzheimer’s disease patient cohorts
Prevalence, demographics, comorbidities, treatment, insurance, cost and five-year survival, estimated across Oracle EHR Real-World Data and IQVIA — the reference description other studies can be measured against.
MLHOps: Machine Learning Health Operations
Deploying a model in a hospital is the start of the work, not the end of it. A survey of the processes for deploying and maintaining clinical machine learning reliably, with guidelines aimed at developers and clinicians separately.
Stratification of surgical outcomes in trigeminal neuralgia using multimodal data
Trigeminal neuralgia has few individualised tools for predicting treatment outcome. Supervised and unsupervised learning over complete pain history, clinical data and brain imaging from 102 patients, used to stratify surgical outcomes.
Representation Noising: A Defence Mechanism Against Harmful Finetuning
Safety training can be stripped out by fine-tuning, even through an API where the attacker never sees the weights. Representation noising removes the information an attacker would need to recover, so the harmful capability is not sitting there waiting to be restored.
Immunization against harmful fine-tuning attacks
Defences in this area are usually claimed rather than compared. This sets out formal conditions for what counts as one, stated in terms of the attacker’s training budget, so the claims become testable.
The Utility and Implications of Ambient Scribes in Primary Care
Ambient scribes listen to the consultation and draft the note from it. What that offers a primary care practice, and what it implies for documentation burden, accuracy and the trust placed in the record.
Trends in heart failure costs for commercially insured patients in the United States (2006–2021)
A population-scale analysis of commercial claims from 2006 to 2021, separating hospitalization, surgery and pharmacy so the trend can be read by category rather than as one number.
Combined use of Donepezil and Memantine increases the probability of five-year survival of Alzheimer’s disease patients
A causal inference study using doubly robust estimators on Oracle EHR Real-World Data, testing formally for interaction between Donepezil and Memantine rather than assuming it.
A causal inference study: The impact of the combined administration of Donepezil and Memantine on decreasing hospital and emergency department visits of Alzheimer’s disease patients
Doubly robust estimation on one of the largest clinical databases, comparing each monotherapy and the combination against no medication for visits per year.