Electronic Health Records (EHRs) have transformed healthcare by digitally consolidating patient medical history, encompassing structured data (e.g., demographic data, lab results), and unstructured textual data (e.g., clinical notes). These data hold significant potential for predictive modelling, and recent studies have dedicated efforts to leverage the different modalities in a cohesive and effective manner to improve predictive accuracy. This Systematic Literature Review (SLR) addresses the application of Multimodal Deep Learning (MDL) methods in EHR-based prediction modelling, specifically through the fusion of structured and textual data. Following PRISMA guidelines, we conducted a comprehensive literature search across six article databases, using a carefully designed search string. After applying inclusion and exclusion criteria, we selected 77 primary studies. Data extraction was standardized using a structured form based on the CHARMS checklist. We categorized and analysed the fusion strategies employed across the studies. By combining structured and textual data at the input level, early fusion enabled models to learn joint feature representations from the beginning, whether in vectorized representations or data textualization. Intermediate fusion, which delays integration, was particularly useful for tasks where each modality provides unique insights that need to be processed independently before being combined. Late fusion enabled modularity by integrating outputs from unimodal models, which is suitable when EHR structured and textual data have varying quality or reliability. We also identified trends and open issues that need attention. This review contributes a comprehensive understanding of EHR data fusion practices using MDL, highlighting potential pathways for future research and development in health informatics.

Manual decisions regarding the timing of surveillance endoscopy for premalignant Barrett’s oesophagus (BO) is error-prone. This leads to inefficient resource usage and safety risks. To automate decision-making, we fine-tuned Bidirectional Encoder Representations from Transformers (BERT) models to categorize BO length (EndoBERT) and worst histopathological grade (PathBERT) on 4,831 endoscopy and 4,581 pathology reports from Guy’s and St Thomas’ Hospital (GSTT). The accuracies for EndoBERT test sets from GSTT, King’s College Hospital (KCH), and Sandwell and West Birmingham Hospitals (SWB) were 0.95, 0.86, and 0.99, respectively. Average accuracies for PathBERT were 0.93, 0.91, and 0.92, respectively. A retrospective analysis of 1640 GSTT reports revealed a 27% discrepancy between endoscopists’ decisions and model recommendations. This study underscores the development and deployment of NLP-based software in BO surveillance, demonstrating high performance at multiple sites. The analysis emphasizes the potential efficiency of automation in enhancing precision and guideline adherence in clinical decision-making.

BackgroundIndividualising mental healthcare at times when a patient is most at risk of suicide involves shifting research emphasis from static risk factors to those that may be modifiable with interventions. Currently, risk assessment is based on a range of extensively reported stable risk factors, but critical to dynamic suicide risk assessment is an understanding of each individual patient’s health trajectory over time. The use of electronic health records (EHRs) and analysis using machine learning has the potential to accelerate progress in developing early warning indicators.SettingEHR data from the South London and Maudsley NHS Foundation Trust (SLaM) which provides secondary mental healthcare for 1.8 million people living in four South London boroughs.ObjectivesTo determine whether the time window proximal to a hospitalised suicide attempt can be discriminated from a distal period of lower risk by analysing the documentation and mental health clinical free text data from EHRs and (i) investigate whether the rate at which EHR documents are recorded per patient is associated with a suicide attempt; (ii) compare document-level word usage between documents proximal and distal to a suicide attempt; and (iii) compare n-gram frequency related to third-person pronoun use proximal and distal to a suicide attempt using machine learning.MethodsThe Clinical Record Interactive Search (CRIS) system allowed access to de-identified information from the EHRs. CRIS has been linked with Hospital Episode Statistics (HES) data for Admitted Patient Care. We analysed document and event data for patients who had at some point between 1 April 2006 and 31 March 2013 been hospitalised with a HES ICD-10 code related to attempted suicide (X60–X84; Y10–Y34; Y87.0/Y87.2).Findingsn = 8,247 patients were identified to have made a hospitalised suicide attempt. Of these, n = 3,167 (39.8%) of patients had at least one document available in their EHR prior to their first suicide attempt. N = 1,424 (45.0%) of these patients had been “monitored” by mental healthcare services in the past 30 days. From 60 days prior to a first suicide attempt, there was a rapid increase in the monitoring level (document recording of the past 30 days) increasing from 35.1 to 45.0%. Documents containing words related to prescribed medications/drugs/overdose/poisoning/addiction had the highest odds of being a risk indicator used proximal to a suicide attempt (OR 1.88; precision 0.91 and recall 0.93), and documents with words citing a care plan were associated with the lowest risk for a suicide attempt (OR 0.22; precision 1.00 and recall 1.00). Function words, word sequence, and pronouns were most common in all three representations (uni-, bi-, and tri-gram).ConclusionEHR documentation frequency and language use can be used to distinguish periods distal from and proximal to a suicide attempt. However, in our study 55.0% of patients with documentation, prior to their first suicide attempt, did not have a record in the preceding 30 days, meaning that there are a high number who are not seen by services at their most vulnerable point.

Objectives An association between type 2 diabetes (T2DM) and schizophrenia has long been observed, and recent research revealed presence of shared genetic factors. However, epidemiological evidence was inconsistent, some reported insignificant contribution of genetic factors to T2DM-schizophrenia comorbidity. Prior works studied people with schizophrenia, particularly, antipsychotic-naive patients, or those during the first psychotic experience to limit schizophrenia-related environmental factors. In contrast, we controlled such factors by utilizing a general population sample of individuals undiagnosed with schizophrenia. We hypothesized that if schizophrenia genetics impact T2DM development and such impact is not fully mediated by schizophrenia-related environment, people with high polygenic schizophrenia risk would exhibit elevated T2DM incidence. Methods Using a population-representative sample of adults aged ≥50 from English Longitudinal Study of Ageing ( n = 5968, 493 T2DM cases, average follow-up 8.7 years), we investigated if schizophrenia polygenic risk score (PGS-SZ) is associated with T2DM onset. A proportional hazards model with interval censoring was adjusted for age and sex (Model 1), and age, sex, BMI, hypertension, cardiovascular diseases, exercise, smoking, depressive symptoms and T2DM polygenic risk score (Model 2). According to the power calculations, hazard rates > 1.14 per standard deviation in PGS-SZ could be detected. Results We did not observe a significant association between PGS-SZ and T2DM incidence (hazard ratio 1.04; 95% CI 0.93–1.15; and 1.01, 95% CI 0.94–1.09). Conclusion Our results suggest low contribution of the intrinsic biological mechanisms driven by the polygenic risk of schizophrenia on future T2DM onset. Further research is needed.

Background:Encephalopathy is a severe co-morbid condition in critically ill patients that includes different clinical constellation of neurological symptoms. However, even for the most recognised form, delirium, this medical condition is rarely recorded in structured fields of electronic health records precluding large and unbiased retrospective studies. We aimed to identify patients with encephalopathy using a machine learning-based approach over clinical notes in electronic health records.Methods:We used a list of ICD-9 codes and clinical concepts related to encephalopathy to define a cohort of patients from the MIMIC-III dataset. Clinical notes were annotated with MedCAT and vectorized with a bag-of-word approach or word embedding using clinical concepts normalised to standard nomenclatures as features. Machine learning algorithms (support vector machines and random forest) trained with clinical notes from patients who had a diagnosis of encephalopathy (defined by ICD-9 codes) were used to classify patients with clinical concepts related to encephalopathy in their clinical notes but without any ICD-9 relevant code. A random selection of 50 patients were reviewed by a clinical expert for model validation.Results:Among 46,520 different patients, 7.5% had encephalopathy related ICD-9 codes in all their admissions (group 1, definite encephalopathy), 45% clinical concepts related to encephalopathy only in their clinical notes (group 2, possible encephalopathy) and 38% did not have encephalopathy related concepts neither in structured nor in clinical notes (group 3, non-encephalopathy). Length of stay, mortality rate or number of co-morbid conditions were higher in groups 1 and 2 compared to group 3. The best model to classify patients from group 2 as patients with encephalopathy (SVM using embeddings) had F1 of 85% and predicted 31% patients from group 2 as having encephalopathy with a probability >90%. Validation on new cases found a precision ranging from 92% to 98% depending on the criteria considered.Conclusions:Natural language processing techniques can leverage relevant clinical information that might help to identify patients with under-recognised clinical disorders such as encephalopathy. In the MIMIC dataset, this approach identifies with high probability thousands of patients that did not have a formal diagnosis in the structured information of the EHR.

BACKGROUND AND AIMS:Prevalence of type two diabetes mellitus (T2DM) in people with severe mental illness (SMI) is 2-3 times higher than in general population. Predictive modelling has advanced greatly in the past decade, and it is important to apply cutting-edge methods to vulnerable groups. However, few T2DM prediction models account for the presence of mental illness, and none seemed to have been developed specifically for people with SMI. Therefore, we aimed to develop and internally validate a T2DM prevalence model for people with SMI. METHODS:We utilised a large cross-sectional sample representative of a multi-ethnic population from London (674,000 adults); 10,159 people with SMI formed our analytical sample (1,513 T2DM cases). We fitted a linear logistic regression and XGBoost as stand-alone models and as a stacked ensemble. Age, sex, body mass index, ethnicity, area-based deprivation, past hypertension, cardiovascular diseases, prescribed antipsychotics, and SMI illness were the predictors. RESULTS:Logistic regression performed well while detecting T2DM presence for people with SMI: area under the receiver operator curve (ROC-AUC) was 0.83 (95 % CI 0.79-0.87). XGBoost and LR-XGBoost ensemble performed equally well, ROC-AUC 0.83 (95 % CI 0.79-0.87), indicating a negligible contribution of non-linear terms to predictive power. Ethnicity was the most important predictor after age. We demonstrated how the derived models can be utilised and estimated a 2.14 % (95 %CI 2.03 %-2.24 %) increase in T2DM prevalence in East London SMI population in 20 years' time, driven by the projected demographic changes. CONCLUSIONS:Primary care data, the setting where prediction models could be most fruitfully used, provide enough information for well-performing T2DM prevalence models for people with SMI. We demonstrated how thorough internal cross-validation of an ensemble of a linear and machine-learning model can quantify the predictive value of non-linearity in the data.

Objective This paper evaluates the application of a natural language processing (NLP) model for extracting clinical text referring to interpersonal violence using electronic health records (EHRs) from a large mental healthcare provider. Design A multidisciplinary team iteratively developed guidelines for annotating clinical text referring to violence. Keywords were used to generate a dataset which was annotated (ie, classified as affirmed, negated or irrelevant) for: presence of violence, patient status (ie, as perpetrator, witness and/or victim of violence) and violence type (domestic, physical and/or sexual). An NLP approach using a pretrained transformer model, BioBERT (Bidirectional Encoder Representations from Transformers for Biomedical Text Mining) was fine-tuned on the annotated dataset and evaluated using 10-fold cross-validation. Setting We used the Clinical Records Interactive Search (CRIS) database, comprising over 500 000 de-identified EHRs of patients within the South London and Maudsley NHS Foundation Trust, a specialist mental healthcare provider serving an urban catchment area. Participants Searches of CRIS were carried out based on 17 predefined keywords. Randomly selected text fragments were taken from the results for each keyword, amounting to 3771 text fragments from the records of 2832 patients. Outcome measures We estimated precision, recall and F1 score for each NLP model. We examined sociodemographic and clinical variables in patients giving rise to the text data, and frequencies for each annotated violence characteristic. Results Binary classification models were developed for six labels (violence presence, perpetrator, victim, domestic, physical and sexual). Among annotations affirmed for the presence of any violence, 78% (1724) referred to physical violence, 61% (1350) referred to patients as perpetrator and 33% (731) to domestic violence. NLP models’ precision ranged from 89% (perpetrator) to 98% (sexual); recall ranged from 89% (victim, perpetrator) to 97% (sexual). Conclusions State of the art NLP models can extract and classify clinical text on violence from EHRs at acceptable levels of scale, efficiency and accuracy.

Pain has been an area of growing interest in the past decade and is known to be associated with mental health issues. Due to the ambiguous nature of how pain is described in text, it presents a unique natural language processing (NLP) challenge. Understanding how pain is described in text and utilizing this knowledge to improve NLP tasks would be of substantial clinical importance. Not much work has previously been done in this space. For this reason, and in order to develop an English lexicon for use in NLP applications, an exploration of pain concepts within free text was conducted. The exploratory text sources included two hospital databases, a social media platform (Twitter), and an online community (Reddit). This exploration helped select appropriate sources and inform the construction of a pain lexicon. The terms within the final lexicon were derived from three sources-literature, ontologies, and word embedding models. This lexicon was validated by two clinicians as well as compared to an existing 26-term pain sub-ontology and MeSH (Medical Subject Headings) terms. The final validated lexicon consists of 382 terms and will be used in downstream NLP tasks by helping select appropriate pain-related documents from electronic health record (EHR) databases, as well as pre-annotating these words to help in development of an NLP application for classification of mentions of pain within the documents. The lexicon and the code used to generate the embedding models have been made publicly available.

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 customising 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.

Multimorbidity research in mental health services requires data from physical health conditions which is traditionally limited in mental health care electronic health records. In this study, we aimed to extract data from physical health conditions from clinical notes using SemEHR. Data was extracted from Clinical Record Interactive Search (CRIS) system at South London and Maudsley Biomedical Research Centre (SLaM BRC) and the cohort consisted of all individuals who had received a primary or secondary diagnosis of severe mental illness between 2007 and 2018. Three pairs of annotators annotated 2403 documents with an average Cohen's Kappa of 0.757. Results show that the NLP performance varies across different diseases areas (F1 0.601 - 0.954) suggesting that the language patterns or terminologies of different condition groups entail different technical challenges to the same NLP task.

Biomedical documents such as Electronic Health Records (EHRs) contain a large amount of information in an unstructured format. The data in EHRs is a hugely valuable resource documenting clinical narratives and decisions, but whilst the text can be easily understood by human doctors it is challenging to use in research and clinical applications. To uncover the potential of biomedical documents we need to extract and structure the information they contain. The task at hand is Named Entity Recognition and Linking (NER+L). The number of entities, ambiguity of words, overlapping and nesting make the biomedical area significantly more difficult than many others. To overcome these difficulties, we have developed the Medical Concept Annotation Tool (MedCAT), an open-source unsupervised approach to NER+L. MedCAT uses unsupervised machine learning to disambiguate entities. It was validated on MIMIC-III (a freely accessible critical care database) and MedMentions (Biomedical papers annotated with mentions from the Unified Medical Language System). In case of NER+L, the comparison with existing tools shows that MedCAT improves the previous best with only unsupervised learning (F1=0.848 vs 0.691 for disease detection; F1=0.710 vs. 0.222 for general concept detection). A qualitative analysis of the vector embeddings learnt by MedCAT shows that it captures latent medical knowledge available in EHRs (MIMIC-III). Unsupervised learning can improve the performance of large scale entity extraction, but it has some limitations when working with only a couple of entities and a small dataset. In that case options are supervised learning or active learning, both of which are supported in MedCAT via the MedCATtrainer extension. Our approach can detect and link millions of different biomedical concepts with state-of-the-art performance, whilst being lightweight, fast and easy to use.

Discovering the low-energy conformations of a molecule is of great interest to computational chemists, with applications in in silico materials design and drug discovery. In this paper, we propose a variable neighbourhood search heuristic for the conformational search problem. Using the structure of a molecule, neighbourhoods are chosen to allow for the efficient use of a binary quadratic optimizer for conformational search. The method is flexible with respect to the choice of molecular force field and the number of discretization levels in the search space, and can be further generalized to take advantage of higher-order binary polynomial optimizers. It is well-suited for the use of devices such as quantum annealers. After carefully defining neighbourhoods, the method easily adapts to the size and topology of these devices, allowing for seamless scaling alongside their future improvements.

Objective Routinely collected healthcare data are a powerful research resource but often lack detailed disease-specific information that is collected in clinical free text, for example, clinic letters. We aim to use natural language processing techniques to extract detailed clinical information from epilepsy clinic letters to enrich routinely collected data. Design We used the general architecture for text engineering (GATE) framework to build an information extraction system, ExECT (extraction of epilepsy clinical text), combining rule-based and statistical techniques. We extracted nine categories of epilepsy information in addition to clinic date and date of birth across 200 clinic letters. We compared the results of our algorithm with a manual review of the letters by an epilepsy clinician. Setting De-identified and pseudonymised epilepsy clinic letters from a Health Board serving half a million residents in Wales, UK. Results We identified 1925 items of information with overall precision, recall and F1 score of 91.4%, 81.4% and 86.1%, respectively. Precision and recall for epilepsy-specific categories were: epilepsy diagnosis (88.1%, 89.0%), epilepsy type (89.8%, 79.8%), focal seizures (96.2%, 69.7%), generalised seizures (88.8%, 52.3%), seizure frequency (86.3%–53.6%), medication (96.1%, 94.0%), CT (55.6%, 58.8%), MRI (82.4%, 68.8%) and electroencephalogram (81.5%, 75.3%). Conclusions We have built an automated clinical text extraction system that can accurately extract epilepsy information from free text in clinic letters. This can enhance routinely collected data for research in the UK. The information extracted with ExECT such as epilepsy type, seizure frequency and neurological investigations are often missing from routinely collected data. We propose that our algorithm can bridge this data gap enabling further epilepsy research opportunities. While many of the rules in our pipeline were tailored to extract epilepsy specific information, our methods can be applied to other diseases and also can be used in clinical practice to record patient information in a structured manner.

Assessing a patient's risk of an impending suicide attempt has been hampered by limited information about dynamic factors that change rapidly in the days leading up to an attempt. The storage of patient data in electronic health records (EHRs) has facilitated population-level risk assessment studies using machine learning techniques. Until recently, most such work has used only structured EHR data and excluded the unstructured text of clinical notes. In this article, we describe our experiments on suicide risk assessment, modelling the problem as a classification task. Given the wealth of text data in mental health EHRs, we aimed to assess the impact of using this data in distinguishing periods prior to a suicide attempt from those not preceding such an attempt. We compare three different feature sets, one structured and two text-based, and show that inclusion of text features significantly improves classification accuracy in suicide risk assessment.

In order to understand complex language in the text of the Electronic Health Record (EHR), it is sometimes necessary to consider contextual information from other parts of the text. We introduce a novel neural network architecture to classify sentences from EHRs, taking in to account such distant contextual information. We test our method on detection of suicide-related information in documents from Autism Spectrum Disorder patients. Our approach improves accuracy over the state of the art. Introduction Classification of sentences (and more generally of any chunk or mention) is an important component of electronic health record (EHR) text analytics, and therefore for the reuse of EHRs for research. For example, classification of suicide-related mentions in psychiatric disorders may be used to create retrospective clinical cohorts for epidemiological research, and potentially to aid risk prediction in clinical settings [5]. Most previous classification approaches [1,2,3,4,5] consider only the sentence itself, and context from distantly surrounding text is ignored. However, this surrounding context may supply additional information to aid classification, with preceding and following sentences adding weight or nuance to a point. For example, consider these imaginary consecutive sentences: 1. Her parents thought she felt life not worth living, although in clinic she denied ever having been suicidal and said she had never self-harmed. 2. She denied any current plans or thoughts of self-harm. Sentence 1 contains several negated statements, conjunctions, and discusses people other than the patient, making it difficult for a classifier to reach the correct decision. The classifier could use sentence 2 to provide additional context to correctly classify sentence 1 as negative for suicidal thoughts. In this paper, we address this problem with a novel neural network architecture – Context LSTM-CNN (Figure 1) that is not only based on the sentence, but also takes the full context from the rest of the document into consideration, without significantly increasing the computational cost. Context LSTM-CNN first applies a bi-directional Long short-term memory (LSTM) [6] over sentence word embeddings to capture the sentence level relationship between the words, and passes these embeddings into a CNN [1] network for classification. At the same time, the left and right contexts of the sentence are encoded into fixed-size codes with a Fixed-Size Ordinally-Forgetting Encoding (FOFE) [7]. Finally we feed the left and right context FOFE codes together with the CNN output into a final output layer. The computational cost of adding the small, fixed size FOFE code to the network is negligible. Figure 1. Architecture of Context LSTM-CNN Preliminary Experiment We conducted a preliminary experiment on the corpus reported in [5]. This corpus of 4916 documents from Autism Spectrum Disorder patients’ records, labels all mentions of relevance to suicidal behaviour, in one of three categories: ‘positive for suicidality’, ‘negated for suicidality’ and ‘suicidality uncertain’. The labeled mentions are mostly of sentence size or close to sentence size, so we treat this as a sentence classification task, with all other unlabeled sentences as ‘not relevant’, giving four classes. We compared our approach against two state of the art neural net baselines: LSTM and CNN. We randomly held out 20% of the document as a test set, and used the remainder for training. We repeated this holdout process five times, and report the average F-measure for each class in Table 1. Results and Conclusions The results show that Context LSTM-CNN improves F1 by up to 10 to 20% when compared to the baselines. In addition, the Context LSTM-CNN training and application time is almost identical to LSTM. Table 1. F-1 score of classifiers over ASD suicidal ideation sentences Not relevant Suicidality Pos Suicidality Neg Uncertain CNN 98.27 58.66 66.70 35.23 LSTM 98.31 59.31 65.32 29.26 Context LSTM-CNN 99.22 74.00 78.15 48.43 Acknowledgement JD is supported by a Medical Research Council (MRC) Clinical Research Training Fellowship (MR/L017105/1), SV is supported by the Swedish Research Council (2015-00359), Marie Skłodowska Curie Actions, Cofund, Project INCA 600398, RD is funded by a Clinician Scientist Fellowship (research project e-HOST-IT) from the Health Foundation in partnership with the Academy of Medical Sciences. This poster represents independent research part funded by the National Institute for Health Research (NIHR) Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health and Social Care.

BackgroundDeriving structured data from unstructured clinical notes in electronic health records (EHRs) requires natural language processing and clinical expertise, which is often costly, and frequently a one-off investment. We implemented SemEHR, a semantic search system that reduces the expertise and effort required in this context. We aimed to use it to characterise and select patients for projects such as the UK Department of Health 100,000 Genome Project.MethodsBuilt upon the off-the-shelf toolkits, Bio-YODIE and CogStack, SemEHR integrates heterogeneous EHR documents and identifies contextualised (negation, temporality, and experiencer) mentions of a wide range of biomedical concepts including SNOMED CT, ICD-10, LOINC, and Drug Ontology. Text mining and semantics techniques are incorporated to derive a longitudinal patient panorama, combining structured profiles and unstructured records, available through semantic search interfaces.FindingsWe deployed SemEHR in various UK hospital EHRs, including the South London and Maudsley NHS Foundation Trust, where 46 million concept mentions were identified from 18 million documents. In a liver disease study, SemEHR identified 94 of 100 hepatitis C positive manually annotated patients. In a HIV study, SemEHR identified 21 of 23 true positives in a 1000-patient cohort. At King's College Hospital, SemEHR is being used to recruit patients into the 100,000 Genomes Project, where ontological associations are integrated to match recruitment criteria and populate complex phenotype models. A preliminary evaluation suggests that the tool is able to validate previously submitted cases and is very fast in searching phenotypes.InterpretationUsing SemEHR, a query such as “find patients with a family history of hepatitis C”, which previously might have required the user to have natural language processing expertise, becomes a simple search, for which SemEHR retrieves a relevant patient cohort, populates patient-level summaries, and provides a link to each mention in the original source. Results and feedback from the multiple studies have proven its efficiency: previously weeks or months of work can be done within minutes in some cases.FundingMedical Research Council (MC_PC_14089); NIHR Biomedical Research Centre for Mental Health, Biomedical Research Unit for Dementia; the European Union's Horizon 2020 (No 644753 KConnect); Wellcome Trust Seed Award in Science (109823/Z/15/Z); National Institute for Health Research University College London Hospital's Biomedical Research Centre; Arthritis Research UK; British Heart Foundation; Cancer Research UK; Chief Scientist Office; Economic and Social Research Council; Engineering and Physical Sciences Research Council; National Institute for Social Care and Health Research; Wellcome Trust (grant MR/K006584/1).

Natural language processing is being pressed into use to facilitate the selection of cases for medical research in electronic health record databases, though study inclusion criteria may be complex, and the linguistic cues indicating eligibility may be subtle.Finding cases of first episode psychosis raised a number of problems for automated approaches, providing an opportunity to explore how machine learning technologies might be used to overcome them.A system was delivered that achieved an AUC of 0.85, enabling 95% of relevant cases to be identified whilst halving the work required in manually reviewing cases.The techniques that made this possible are presented.

ObjectivesTo identify negative symptoms in the clinical records of a large sample of patients with schizophrenia using natural language processing and assess their relationship with clinical outcomes.DesignObservational study using an anonymised electronic health record case register.SettingSouth London and Maudsley NHS Trust (SLaM), a large provider of inpatient and community mental healthcare in the UK.Participants7678 patients with schizophrenia receiving care during 2011.Main outcome measuresHospital admission, readmission and duration of admission.Results10 different negative symptoms were ascertained with precision statistics above 0.80. 41% of patients had 2 or more negative symptoms. Negative symptoms were associated with younger age, male gender and single marital status, and with increased likelihood of hospital admission (OR 1.24, 95% CI 1.10 to 1.39), longer duration of admission (β-coefficient 20.5 days, 7.6–33.5), and increased likelihood of readmission following discharge (OR 1.58, 1.28 to 1.95).ConclusionsNegative symptoms were common and associated with adverse clinical outcomes, consistent with evidence that these symptoms account for much of the disability associated with schizophrenia. Natural language processing provides a means of conducting research in large representative samples of patients, using data recorded during routine clinical practice.

We present two approaches to time expression identification, as entered in to SemEval-2015 Task 6, Clinical TempEval.The first is a comprehensive rule-based approach that favoured recall, and which achieved the best recall for time expression identification in Clinical TempEval.The second is an SVM-based system built using readily available components, which was able to achieve a competitive F1 in a short development time.We discuss how the two approaches perform relative to each other, and how characteristics of the corpus affect the suitability of different approaches and their outcomes.