Joshua Au Yeung

Zeljko Kraljevic, Anthony Shek, Joshua Au Yeung et al.

Protecting patient privacy in healthcare records is a top priority, and redaction is a commonly used method for obscuring directly identifiable information in text. Rule-based methods have been widely used, but their precision is often low causing over-redaction of text and frequently not being adaptable enough for non-standardised or unconventional structures of personal health information. Deep learning techniques have emerged as a promising solution, but implementing them in real-world environments poses challenges due to the differences in patient record structure and language across different departments, hospitals, and countries. In this study, we present AnonCAT, a transformer-based model and a blueprint on how deidentification models can be deployed in real-world healthcare. AnonCAT was trained through a process involving manually annotated redactions of real-world documents from three UK hospitals with different electronic health record systems and 3116 documents. The model achieved high performance in all three hospitals with a Recall of 0.99, 0.99 and 0.96. Our findings demonstrate the potential of deep learning techniques for improving the efficiency and accuracy of redaction in global healthcare data and highlight the importance of building workflows which not just use these models but are also able to continually fine-tune and audit the performance of these algorithms to ensure continuing effectiveness in real-world settings. This approach provides a blueprint for the real-world use of de-identifying algorithms through fine-tuning and localisation, the code together with tutorials is available on GitHub (https://github.com/CogStack/MedCAT).

Zeljko Kraljevic, Dan Bean, Anthony Shek et al.

Background: Electronic Health Records hold detailed longitudinal information about each patient's health status and general clinical history, a large portion of which is stored within the unstructured text. Existing approaches focus mostly on structured data and a subset of single-domain outcomes. We explore how temporal modelling of patients from free text and structured data, using deep generative transformers can be used to forecast a wide range of future disorders, substances, procedures or findings. Methods: We present Foresight, a novel transformer-based pipeline that uses named entity recognition and linking tools to convert document text into structured, coded concepts, followed by providing probabilistic forecasts for future medical events such as disorders, substances, procedures and findings. We processed the entire free-text portion from three different hospital datasets totalling 811336 patients covering both physical and mental health. Findings: On tests in two UK hospitals (King's College Hospital, South London and Maudsley) and the US MIMIC-III dataset precision@10 0.68, 0.76 and 0.88 was achieved for forecasting the next disorder in a patient timeline, while precision@10 of 0.80, 0.81 and 0.91 was achieved for forecasting the next biomedical concept. Foresight was also validated on 34 synthetic patient timelines by five clinicians and achieved relevancy of 97% for the top forecasted candidate disorder. As a generative model, it can forecast follow-on biomedical concepts for as many steps as required. Interpretation: Foresight is a general-purpose model for biomedical concept modelling that can be used for real-world risk forecasting, virtual trials and clinical research to study the progression of disorders, simulate interventions and counterfactuals, and educational purposes.

Zeljko Kraljevic, Joshua Au Yeung, Daniel Bean et al.

Foresight 2 (FS2) is a large language model fine-tuned on hospital data for modelling patient timelines (GitHub 'removed for anon'). It can understand patients' clinical notes and predict SNOMED codes for a wide range of biomedical use cases, including diagnosis suggestions, risk forecasting, and procedure and medication recommendations. FS2 is trained on the free text portion of the MIMIC-III dataset, firstly through extracting biomedical concepts and then creating contextualised patient timelines, upon which the model is then fine-tuned. The results show significant improvement over the previous state-of-the-art for the next new biomedical concept prediction (P/R - 0.73/0.66 vs 0.52/0.32) and a similar improvement specifically for the next new disorder prediction (P/R - 0.69/0.62 vs 0.46/0.25). Finally, on the task of risk forecast, we compare our model to GPT-4-turbo (and a range of open-source biomedical LLMs) and show that FS2 performs significantly better on such tasks (P@5 - 0.90 vs 0.65). This highlights the need to incorporate hospital data into LLMs and shows that small models outperform much larger ones when fine-tuned on high-quality, specialised data.

Joshua Au Yeung, Jacopo Dalmasso, Luca Foschini et al.

Background: Emerging reports of "AI psychosis" are on the rise, where user-LLM interactions may exacerbate or induce psychosis or adverse psychological symptoms. Whilst the sycophantic and agreeable nature of LLMs can be beneficial, it becomes a vector for harm by reinforcing delusional beliefs in vulnerable users. Methods: Psychosis-bench is a novel benchmark designed to systematically evaluate the psychogenicity of LLMs comprises 16 structured, 12-turn conversational scenarios simulating the progression of delusional themes(Erotic Delusions, Grandiose/Messianic Delusions, Referential Delusions) and potential harms. We evaluated eight prominent LLMs for Delusion Confirmation (DCS), Harm Enablement (HES), and Safety Intervention(SIS) across explicit and implicit conversational contexts. Findings: Across 1,536 simulated conversation turns, all LLMs demonstrated psychogenic potential, showing a strong tendency to perpetuate rather than challenge delusions (mean DCS of 0.91 $\pm$0.88). Models frequently enabled harmful user requests (mean HES of 0.69 $\pm$0.84) and offered safety interventions in only roughly a third of applicable turns (mean SIS of 0.37 $\pm$0.48). 51 / 128 (39.8%) of scenarios had no safety interventions offered. Performance was significantly worse in implicit scenarios, models were more likely to confirm delusions and enable harm while offering fewer interventions (p < .001). A strong correlation was found between DCS and HES (rs = .77). Model performance varied widely, indicating that safety is not an emergent property of scale alone. Conclusion: This study establishes LLM psychogenicity as a quantifiable risk and underscores the urgent need for re-thinking how we train LLMs. We frame this issue not merely as a technical challenge but as a public health imperative requiring collaboration between developers, policymakers, and healthcare professionals.