Foresight-a generative pretrained transformer for modelling of patient timelines using electronic health records: a retrospective modelling study.

BACKGROUND: An electronic health record (EHR) holds detailed longitudinal information about a patient's health status and general clinical history, a large portion of which is stored as unstructured, free text. Existing approaches to model a patient's trajectory focus mostly on structured data and a subset of single-domain outcomes. This study aims to evaluate the effectiveness of Foresight, a generative transformer in temporal modelling of patient data, integrating both free text and structured formats, to predict a diverse array of future medical outcomes, such as disorders, substances (eg, to do with medicines, allergies, or poisonings), procedures, and findings (eg, relating to observations, judgements, or assessments). METHODS: Foresight is a novel transformer-based pipeline that uses named entity recognition and linking tools to convert EHR document text into structured, coded concepts, followed by providing probabilistic forecasts for future medical events, such as disorders, substances, procedures, and findings. The Foresight pipeline has four main components: (1) CogStack (data retrieval and preprocessing); (2) the Medical Concept Annotation Toolkit (structuring of the free-text information from EHRs); (3) Foresight Core (deep-learning model for biomedical concept modelling); and (4) the Foresight web application. We processed the entire free-text portion from three different hospital datasets (King's College Hospital [KCH], South London and Maudsley [SLaM], and the US Medical Information Mart for Intensive Care III [MIMIC-III]), resulting in information from 811 336 patients and covering both physical and mental health institutions. We measured the performance of models using custom metrics derived from precision and recall. FINDINGS: Foresight achieved a precision@10 (ie, of 10 forecasted candidates, at least one is correct) of 0·68 (SD 0·0027) for the KCH dataset, 0·76 (0·0032) for the SLaM dataset, and 0·88 (0·0018) for the MIMIC-III dataset, for forecasting the next new disorder in a patient timeline. Foresight also achieved a precision@10 value of 0·80 (0·0013) for the KCH dataset, 0·81 (0·0026) for the SLaM dataset, and 0·91 (0·0011) for the MIMIC-III dataset, for forecasting the next new biomedical concept. In addition, Foresight was validated on 34 synthetic patient timelines by five clinicians and achieved a relevancy of 33 (97% [95% CI 91-100]) of 34 for the top forecasted candidate disorder. As a generative model, Foresight 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, to simulate interventions and counterfactuals, and for educational purposes. FUNDING: National Health Service Artificial Intelligence Laboratory, National Institute for Health and Care Research Biomedical Research Centre, and Health Data Research UK.

Artificial intelligence (AI) for neurologists: do digital neurones dream of electric sheep?

Artificial intelligence (AI) is routinely mentioned in journals and newspapers, and non-technical outsiders may have difficulty in distinguishing hyperbole from reality. We present a practical guide to help non-technical neurologists to understand healthcare AI. AI is being used to support clinical decisions in treating neurological disorders. We introduce basic concepts of AI, such as machine learning and natural language processing, and explain how AI is being used in healthcare, giving examples its benefits and challenges. We also cover how AI performance is measured, and its regulatory aspects in healthcare. An important theme is that AI is a general-purpose technology like medical statistics, with broad utility applicable in various scenarios, such that niche approaches are outpaced by approaches that are broadly applicable in many disease areas and specialties. By understanding AI basics and its potential applications, neurologists can make informed decisions when evaluating AI used in their clinical practice. This article was written by four humans, with generative AI helping with formatting and image generation.

Federated learning for medical imaging radiology.

Federated learning (FL) is gaining wide acceptance across the medical AI domains. FL promises to provide a fairly acceptable clinical-grade accuracy, privacy, and generalisability of machine learning models across multiple institutions. However, the research on FL for medical imaging AI is still in its early stages. This paper presents a review of recent research to outline the difference between state-of-the-art [SOTA] (published literature) and state-of-the-practice [SOTP] (applied research in realistic clinical environments). Furthermore, the review outlines the future research directions considering various factors such as data, learning models, system design, governance, and human-in-loop to translate the SOTA into SOTP and effectively collaborate across multiple institutions.

Latent Transformer Models for out-of-distribution detection.

Any clinically-deployed image-processing pipeline must be robust to the full range of inputs it may be presented with. One popular approach to this challenge is to develop predictive models that can provide a measure of their uncertainty. Another approach is to use generative modelling to quantify the likelihood of inputs. Inputs with a low enough likelihood are deemed to be out-of-distribution and are not presented to the downstream predictive model. In this work, we evaluate several approaches to segmentation with uncertainty for the task of segmenting bleeds in 3D CT of the head. We show that these models can fail catastrophically when operating in the far out-of-distribution domain, often providing predictions that are both highly confident and wrong. We propose to instead perform out-of-distribution detection using the Latent Transformer Model: a VQ-GAN is used to provide a highly compressed latent representation of the input volume, and a transformer is then used to estimate the likelihood of this compressed representation of the input. We demonstrate this approach can identify images that are both far- and near- out-of-distribution, as well as provide spatial maps that highlight the regions considered to be out-of-distribution. Furthermore, we find a strong relationship between an image's likelihood and the quality of a model's segmentation on it, demonstrating that this approach is viable for filtering out unsuitable images.

Mapping and evaluating national data flows: transparency, privacy, and guiding infrastructural transformation.

The importance of big health data is recognised worldwide. Most UK National Health Service (NHS) care interactions are recorded in electronic health records, resulting in an unmatched potential for population-level datasets. However, policy reviews have highlighted challenges from a complex data-sharing landscape relating to transparency, privacy, and analysis capabilities. In response, we used public information sources to map all electronic patient data flows across England, from providers to more than 460 subsequent academic, commercial, and public data consumers. Although NHS data support a global research ecosystem, we found that multistage data flow chains limit transparency and risk public trust, most data interactions do not fulfil recommended best practices for safe data access, and existing infrastructure produces aggregation of duplicate data assets, thus limiting diversity of data and added value to end users. We provide recommendations to support data infrastructure transformation and have produced a website (https://DataInsights.uk) to promote transparency and showcase NHS data assets.

AI chatbots not yet ready for clinical use.

As large language models (LLMs) expand and become more advanced, so do the natural language processing capabilities of conversational AI, or "chatbots". OpenAI's recent release, ChatGPT, uses a transformer-based model to enable human-like text generation and question-answering on general domain knowledge, while a healthcare-specific Large Language Model (LLM) such as GatorTron has focused on the real-world healthcare domain knowledge. As LLMs advance to achieve near human-level performances on medical question and answering benchmarks, it is probable that Conversational AI will soon be developed for use in healthcare. In this article we discuss the potential and compare the performance of two different approaches to generative pretrained transformers-ChatGPT, the most widely used general conversational LLM, and Foresight, a GPT (generative pretrained transformer) based model focused on modelling patients and disorders. The comparison is conducted on the task of forecasting relevant diagnoses based on clinical vignettes. We also discuss important considerations and limitations of transformer-based chatbots for clinical use.

Diagnostic signature for heart failure with preserved ejection fraction (HFpEF): a machine learning approach using multi-modality electronic health record data.

BACKGROUND: Heart failure with preserved ejection fraction (HFpEF) is thought to be highly prevalent yet remains underdiagnosed. Evidence-based treatments are available that increase quality of life and decrease hospitalization. We sought to develop a data-driven diagnostic model to predict from electronic health records (EHR) the likelihood of HFpEF among patients with unexplained dyspnea and preserved left ventricular EF. METHODS AND RESULTS: The derivation cohort comprised patients with dyspnea and echocardiography results. Structured and unstructured data were extracted using an automated informatics pipeline. Patients were retrospectively diagnosed as HFpEF (cases), non-HF (control cohort I), or HF with reduced EF (HFrEF; control cohort II). The ability of clinical parameters and investigations to discriminate cases from controls was evaluated by extreme gradient boosting. A likelihood scoring system was developed and validated in a separate test cohort. The derivation cohort included 1585 consecutive patients: 133 cases of HFpEF (9%), 194 non-HF cases (Control cohort I) and 1258 HFrEF cases (Control cohort II). Two HFpEF diagnostic signatures were derived, comprising symptoms, diagnoses and investigation results. A final prediction model was generated based on the averaged likelihood scores from these two models. In a validation cohort consisting of 269 consecutive patients [with 66 HFpEF cases (24.5%)], the diagnostic power of detecting HFpEF had an AUROC of 90% (P < 0.001) and average precision of 74%. CONCLUSION: This diagnostic signature enables discrimination of HFpEF from non-cardiac dyspnea or HFrEF from EHR and can assist in the diagnostic evaluation in patients with unexplained dyspnea. This approach will enable identification of HFpEF patients who may then benefit from new evidence-based therapies.

Moving towards vertically integrated artificial intelligence development.

Substantial interest and investment in clinical artificial intelligence (AI) research has not resulted in widespread translation to deployed AI solutions. Current attention has focused on bias and explainability in AI algorithm development, external validity and model generalisability, and lack of equity and representation in existing data. While of great importance, these considerations also reflect a model-centric approach seen in published clinical AI research, which focuses on optimising architecture and performance of an AI model on best available datasets. However, even robustly built models using state-of-the-art algorithms may fail once tested in realistic environments due to unpredictability of real-world conditions, out-of-dataset scenarios, characteristics of deployment infrastructure, and lack of added value to clinical workflows relative to cost and potential clinical risks. In this perspective, we define a vertically integrated approach to AI development that incorporates early, cross-disciplinary, consideration of impact evaluation, data lifecycles, and AI production, and explore its implementation in two contrasting AI development pipelines: a scalable "AI factory" (Mayo Clinic, Rochester, United States), and an end-to-end cervical cancer screening platform for resource poor settings (Paps AI, Mbarara, Uganda). We provide practical recommendations for implementers, and discuss future challenges and novel approaches (including a decentralised federated architecture being developed in the NHS (AI4VBH, London, UK)). Growth in global clinical AI research continues unabated, and introduction of vertically integrated teams and development practices can increase the translational potential of future clinical AI projects.

Anticoagulation for atrial fibrillation in people with serious mental illness in the general hospital setting.

OBJECTIVE: People with serious mental illnesses (SMI) have an increased risk of stroke compared to the general population. This study aims to evaluate oral anticoagulation prescription trends in atrial fibrillation (AF) patients with and without a comorbid SMI. METHODS: An open-source retrieval system for clinical data (CogStack) was used to identify a cohort of AF patients with SMI who ever had an inpatient admission to King's College Hospital from 2011 to 2020. A Natural Language Processing pipeline was used to calculate CHA2DS2-VASc and HASBLED risk scores from Electronic Health Records free text. Antithrombotic prescriptions of warfarin and Direct acting oral anti-coagulants (DOACs) (apixaban, rivaroxaban, dabigatran, edoxaban) were extracted from discharge summaries. RESULTS: Among patients included in the study (n = 16 916), 2.7% had a recorded co-morbid SMI diagnosis. Compared to non-SMI patients, those with SMI had significantly higher CHA2DS2-VASc (mean (SD): 5.3 (1.96) vs 4.7 (2.08), p < 0.001) and HASBLED scores (mean (SD): 3.2 (1.27) vs 2.5 (1.29), p < 0.001). Among AF patients having a CHA2DS2-VASc ≥2, those with co-morbid SMI were less likely than non-SMI patients to be prescribed an OAC (44% vs 54%, p < 0.001). However, there was no evidence of a significant difference between the two groups since 2019. CONCLUSION: Over recent years, DOAC prescription rates have increased among AF patients with SMI in acute hospitals. More research is needed to confirm whether the introduction of DOACs has reduced OAC treatment gaps in people with serious mental illness and to assess whether the use of DOACs has improved health outcomes in this population.

COVID-related hospitalization, intensive care treatment, and all-cause mortality in patients with psychosis and treated with clozapine.

Clozapine, an antipsychotic, is associated with increased susceptibility to infection with COVID-19, compared to other antipsychotics. Here, we investigate associations between clozapine treatment and increased risk of adverse outcomes of COVID-19, namely COVID-related hospitalisation, intensive care treatment, and death, amongst patients taking antipsychotics with schizophrenia-spectrum disorders. Using the clinical records of South London and Maudsley NHS Foundation Trust, we identified 157 individuals who had an ICD-10 diagnosis of schizophrenia-spectrum disorders, were taking antipsychotics (clozapine or other antipsychotics) at the time of COVID-19 pandemic in the UK and had a laboratory-confirmed COVID-19 infection. The following health outcomes were measured: COVID-related hospitalisation, COVID-related intensive care treatment and death. We tested associations between clozapine treatment and each outcome using logistic regression models, adjusting for gender, age, ethnicity, neighbourhood deprivation, obesity, smoking status, diabetes, asthma, bronchitis and hypertension using propensity scores. Of the 157 individuals who developed COVID-19 while on antipsychotics (clozapine or other antipsychotics), there were 28% COVID-related hospitalisations, 8% COVID-related intensive care treatments and 8% deaths of any cause during the 28 days follow-up period. amongst those taking clozapine, there were 25% COVID-related hospitalisations, 7% COVID-related intensive care treatments and 7% deaths. In both unadjusted and adjusted analyses, we found no significant association between clozapine and any of the outcomes. Thus, we found no evidence that patients with clozapine treatment at time of COVID-19 infection had increased risk of hospitalisation, intensive care treatment or death, compared to non-clozapine antipsychotic-treated patients. However, further research should be considered in larger samples to confirm this.

Best practices in the real-world data life cycle.

With increasing digitization of healthcare, real-world data (RWD) are available in greater quantity and scope than ever before. Since the 2016 United States 21st Century Cures Act, innovations in the RWD life cycle have taken tremendous strides forward, largely driven by demand for regulatory-grade real-world evidence from the biopharmaceutical sector. However, use cases for RWD continue to grow in number, moving beyond drug development, to population health and direct clinical applications pertinent to payors, providers, and health systems. Effective RWD utilization requires disparate data sources to be turned into high-quality datasets. To harness the potential of RWD for emerging use cases, providers and organizations must accelerate life cycle improvements that support this process. We build on examples obtained from the academic literature and author experience of data curation practices across a diverse range of sectors to describe a standardized RWD life cycle containing key steps in production of useful data for analysis and insights. We delineate best practices that will add value to current data pipelines. Seven themes are highlighted that ensure sustainability and scalability for RWD life cycles: data standards adherence, tailored quality assurance, data entry incentivization, deploying natural language processing, data platform solutions, RWD governance, and ensuring equity and representation in data.

A Knowledge Distillation Ensemble Framework for Predicting Short- and Long-Term Hospitalization Outcomes From Electronic Health Records Data.

The ability to perform accurate prognosis is crucial for proactive clinical decision making, informed resource management and personalised care. Existing outcome prediction models suffer from a low recall of infrequent positive outcomes. We present a highly-scalable and robust machine learning framework to automatically predict adversity represented by mortality and ICU admission and readmission from time-series of vital signs and laboratory results obtained within the first 24 hours of hospital admission. The stacked ensemble platform comprises two components: a) an unsupervised LSTM Autoencoder that learns an optimal representation of the time-series, using it to differentiate the less frequent patterns which conclude with an adverse event from the majority patterns that do not, and b) a gradient boosting model, which relies on the constructed representation to refine prediction by incorporating static features. The model is used to assess a patient's risk of adversity and provides visual justifications of its prediction. Results of three case studies show that the model outperforms existing platforms in ICU and general ward settings, achieving average Precision-Recall Areas Under the Curve (PR-AUCs) of 0.891 (95% CI: 0.878-0.939) for mortality and 0.908 (95% CI: 0.870-0.935) in predicting ICU admission and readmission.

Natural language word embeddings as a glimpse into healthcare language and associated mortality surrounding end of life.

OBJECTIVES: To clarify real-world linguistic nuances around dying in hospital as well as inaccuracy in individual-level prognostication to support advance care planning and personalised discussions on limitation of life sustaining treatment (LST). DESIGN: Retrospective cross-sectional study of real-world clinical data. SETTING: Secondary care, urban and suburban teaching hospitals. PARTICIPANTS: All inpatients in 12-month period from 1 October 2018 to 30 September 2019. METHODS: Using unsupervised natural language processing, word embedding in latent space was used to generate phrase clusters with most similar semantic embeddings to 'Ceiling of Treatment' and their prognostication value. RESULTS: Word embeddings with most similarity to 'Ceiling of Treatment' clustered around phrases describing end-of-life care, ceiling of care and LST discussions. The phrases have differing prognostic profile with the highest 7-day mortality in the phrases most explicitly referring to end of life-'Withdrawal of care' (56.7%), 'terminal care/end of life care' (57.5%) and 'un-survivable' (57.6%). CONCLUSION: Vocabulary used at end-of-life discussions are diverse and has a range of associations to 7-day mortality. This highlights the importance of correct application of terminology during LST and end-of-life discussions.

Neurological injury from virtual reality mishap.

Consumer virtual reality systems are becoming increasingly popular with the increasing availability of devices and gamified technologies. Self-sustained injury risks exist with the use of this technology in the uncontrolled home environment, however, the public awareness of these risks may not be recognised. We present a case of a low- impact virtual reality fall resulting in spinal cord injury, hypoglossal nerve injury, vertebral artery dissection and traumatic brain injury.

Pre-existing cardiovascular disease rather than cardiovascular risk factors drives mortality in COVID-19.

BACKGROUND: The relative association between cardiovascular (CV) risk factors, such as diabetes and hypertension, established CV disease (CVD), and susceptibility to CV complications or mortality in COVID-19 remains unclear. METHODS: We conducted a cohort study of consecutive adults hospitalised for severe COVID-19 between 1st March and 30th June 2020. Pre-existing CVD, CV risk factors and associations with mortality and CV complications were ascertained. RESULTS: Among 1721 patients (median age 71 years, 57% male), 349 (20.3%) had pre-existing CVD (CVD), 888 (51.6%) had CV risk factors without CVD (RF-CVD), 484 (28.1%) had neither. Patients with CVD were older with a higher burden of non-CV comorbidities. During follow-up, 438 (25.5%) patients died: 37% with CVD, 25.7% with RF-CVD and 16.5% with neither. CVD was independently associated with in-hospital mortality among patients < 70 years of age (adjusted HR 2.43 [95% CI 1.16-5.07]), but not in those ≥ 70 years (aHR 1.14 [95% CI 0.77-1.69]). RF-CVD were not independently associated with mortality in either age group (< 70 y aHR 1.21 [95% CI 0.72-2.01], ≥ 70 y aHR 1.07 [95% CI 0.76-1.52]). Most CV complications occurred in patients with CVD (66%) versus RF-CVD (17%) or neither (11%; p < 0.001). 213 [12.4%] patients developed venous thromboembolism (VTE). CVD was not an independent predictor of VTE. CONCLUSIONS: In patients hospitalised with COVID-19, pre-existing established CVD appears to be a more important contributor to mortality than CV risk factors in the absence of CVD. CVD-related hazard may be mediated, in part, by new CV complications. Optimal care and vigilance for destabilised CVD are essential in this patient group. Trial registration n/a.

Multi-domain clinical natural language processing with MedCAT: The Medical Concept Annotation Toolkit.

Electronic health records (EHR) contain large volumes of unstructured text, requiring the application of information extraction (IE) technologies to enable clinical analysis. We present the open source Medical Concept Annotation Toolkit (MedCAT) that provides: (a) a novel self-supervised machine learning algorithm for extracting concepts using any concept vocabulary including UMLS/SNOMED-CT; (b) a feature-rich annotation interface for customizing and training IE models; and (c) integrations to the broader CogStack ecosystem for vendor-agnostic health system deployment. We show improved performance in extracting UMLS concepts from open datasets (F1:0.448-0.738 vs 0.429-0.650). Further real-world validation demonstrates SNOMED-CT extraction at 3 large London hospitals with self-supervised training over ∼8.8B words from ∼17M clinical records and further fine-tuning with ∼6K clinician annotated examples. We show strong transferability (F1 > 0.94) between hospitals, datasets and concept types indicating cross-domain EHR-agnostic utility for accelerated clinical and research use cases.