With the emergence of AI systems to assist clinical decision-making, several ethical dilemmas are brought to the general attention. AI systems are claimed to be the solution for many high-skilled medical tasks where machines can potentially surpass human ability as for example in identifying normal and abnormal chest X-rays. However, there are also warns that AI tools could be the basis for a human replacement that can risk dehumanisation in medicine. In recent years, important proposals in the domain of AI ethics in healthcare have identified main ethical issues, as for example fairness, autonomy, transparency, and responsibility. The human warranty, which implies human evaluation of the AI procedures, has been described to lower the ethical risks. However, as relevant these works have been, translating principles into action has proved challenging as existing codes were mostly a description of principles. There is a great need to produce how-to proposals that are specific enough to be action-guiding. We present five human-focussed facts designed into a framework of human action for an ethical AI in healthcare. Through the factors, we examine the role of medical practitioners, patients, and developers in designing, implementing, and using AI in a responsible manner that preserves human dignity. The facts encompass a range of ethical concerns that were commonly found in relevant literature. Given that it is crucial to bring as many perspectives as possible to the field, this work contributes to translate principles into human action to guarantee an ethical AI in health.

Related paper: Iniesta, Raquel. "The human role to guarantee an ethical AI in healthcare: a five-facts approach." AI and Ethics 5, no. 1 (2025): 385-397. https://link.springer.com/article/10.1007/s43681-023-00353-x

The study introduces a dynamic, transdiagnostic risk calculator for psychosis that updates individual risk estimates over time by combining electronic health record data with natural language processing and a Cox landmark modelling framework. Unlike previous static risk calculators that rely on a single baseline assessment, this approach incorporates evolving symptom information at multiple time points, improving discrimination, calibration and clinical utility across follow-up. The internal–external validation across geographically distinct NHS boroughs strengthens confidence in the model’s robustness and transportability. Clinically, the work demonstrates a feasible route to earlier and more accurate detection of individuals at increased risk of psychosis within routine secondary mental health care, with clear potential to support timely intervention and improve patient outcomes. 

Related paper:  Krakowski K, Oliver D, Arribas M, Stahl D, Fusar-Poli P. Dynamic and Transdiagnostic Risk Calculator Based on Natural Language Processing for the Prediction of Psychosis in Secondary Mental Health Care: Development and Internal-External Validation Cohort Study. Biol Psychiatry. 2024 Oct 1;96(7):604-614. doi: 10.1016/j.biopsych.2024.05.022. Epub 2024 Jun 7. PMID: 38852896.  https://pubmed.ncbi.nlm.nih.gov/38852896/

We tested a transdiagnostic risk calculator for psychosis to see if it could work to detect people at risk for psychosis earlier across mental health services in the real-world. This was the first study of its kind in psychiatry. After refining the system and integrating it into the electronic health records (EHR) at South London and Maudsley NHS Trust, we automatically screened all patients receiving their first mental health diagnosis (excluding psychosis and organic disorders) over the course of one year. When someone’s two-year risk of psychosis was estimated at 5% or higher, we notified their clinician and recommended a specialised assessment. Out of 3,722 people screened, 115 were identified as at-risk. Most clinicians (77%) responded to the system’s alerts, showing strong engagement and support for this approach. Our findings show that combining precision psychiatry with EHR data can feasibly help detect individuals who may be developing psychosis earlier, allowing for more timely intervention and care.

Related paper:  Oliver, Dominic, Giulia Spada, Craig Colling, Matthew Broadbent, Helen Baldwin, Rashmi Patel, Robert Stewart et al. "Real-world implementation of precision psychiatry: transdiagnostic risk calculator for the automatic detection of individuals at-risk of psychosis." Schizophrenia Research 227 (2021): 52-60.  https://doi.org/10.1016/j.schres.2020.05.007

Temporal drift—changes in patient populations, clinical practices, and data systems over time—can degrade model accuracy, leading to miscalibration and harmful decisions. Our research explores strategies to detect and mitigate this drift, including static and dynamic updating methods. Findings show that dynamic approaches, which continuously adapt to new data, offer the greatest potential for maintaining clinical benefit, though they require robust infrastructure. Beyond accuracy, we examine algorithmic fairness, as drift can disproportionately affect demographic subgroups and exacerbate health disparities. We also investigate challenges arising post-implementation, where treatment decisions alter predictor-outcome relationships, complicating updates. To address these, we explore drift detection systems and counterfactual causal frameworks. Our work emphasizes integrating drift monitoring, fairness assessment, and adaptive updating throughout the model lifecycle to ensure safe, equitable, and effective clinical use. 

Related commentary:  Logeswaran, Yanakan, and Dominic Oliver. "It’s About Time: Why We Need to Consider Temporal Drift When Developing and Implementing Clinical Prediction Models." Biological Psychiatry: Cognitive Neuroscience and Neuroimaging 10, no. 7 (2025): 677-678. https://www.sciencedirect.com/science/article/pii/S2451902225001624

Highlights

• Type 2 diabetes mellitus (T2DM) is common in people with severe mental illness.

• Few T2DM prediction models include severe mental illness (SMI); none seem to target SMI.

• We derived T2DM prevalence model for people with SMI living in East London.

• UK primary care data can be used to develop well-performing models for SMI.

• Ensemble of a linear and a machine-learning model can help quantify data non-linearity.

Related paper:  Shamsutdinova, D., Das-Munshi, J., Ashworth, M., Roberts, A., & Stahl, D. (2023). Predicting type 2 diabetes prevalence for people with severe mental illness in a multi-ethnic East London population. International Journal of Medical Informatics, 172, 105019. https://www.sciencedirect.com/science/article/pii/S1386505623000369