10 research outputs found
Clinical Risk Factors in Haemostasis: Evaluating VTE and Bleeding through Epidemiological and NLP-based methods
No full textBlodpropper og større blødninger er hyppige og alvorlige medicinske tilstande. Blodpropper, særligt veneblodpropper (VTE), kan føre til livstruende komplikationer. Omvendt kan større blødninger være mindst lige så farlige og uforudsigelige. Håndteringen af disse tilstande kræver en omhyggelig balance mellem risikoen for blodprop og risikoen for blødning, hvilket ofte er vanskeligt, fordi disse risici ændrer sig over tid og varierer meget fra person til person. Denne afhandling kombinerer traditionel medicinsk forskning med metoder indenfor kunstig intelligens for bedre at forstå, hvem der er i risiko for VTE eller blødning.Mens traditionel medicinsk forskning har baseret sig på registre og strukturerede databaser, indeholder patientjournaler langt mere detaljeret og præcis information om VTE og blødning. En stor del af disse oplysninger er dog skrevet i ustrukturerede kliniske notater. Det betyder, at informationen ikke nemt kan tilgås eller analyseres med traditionelle computerprogrammer, og manuel gennemgang af store mængder klinisk tekst er ikke realistisk. For at løse dette anvendes en metode inden for kunstig intelligens (AI) kaldet naturlig sprogbehandling (natural language processing, NLP), som gør det muligt for computere at læse og fortolke de fritekstnotater, sundhedsprofessionelle skriver. Dette gør det muligt at identificere og anvende vigtig klinisk information hurtigt og præcist.Afhandlingen består af fire studier. Studie I viser, at ikke alle former for arvelig trombofili medfører den samme risiko for VTE. Arvelig trombofili er en fællesbetegnelse for genetiske mutationer, der øger risikoen for VTE. Studie I udfordrer tidligere antagelser og understreger behovet for mere præcise risikovurderinger, der inddrager både genetiske og kliniske forhold i en samlet vurdering. Studie II præsenterer en ny metode til kvantificering af blodtab under operationer ved at identificere oplysninger om blødning direkte i operationsbeskrivelserne. Dette hjælper med at definere, hvad der kan betragtes som en uforholdsmæssig blødning ved forskellige typer af operationer, og er derigennem et vigtigt skridt i retningen mod bedre retningslinjer for håndtering af blødning under og efter kirurgiske indgreb. Studie III fokuserer på større blødninger blandt indlagte medicinske patienter og viser, at større blødninger ofte overses, hvis man udelukkende kigger på diagnosekoder. Ved at bruge NLP kunne studiet identificere blødningsepisoder mere præcist og udvikle modeller til at forudsige, hvilke patienter der var i størst risiko for at få større blødninger. Disse modeller viste, at rutineblodprøver og patientens generelle helbredstilstand var de mest nyttige indikatorer. Studie IV tager skridtet videre og viser, hvordan NLP kan implementeres i den kliniske hverdag. Et system blev udviklet og integreret i det daglige arbejde for at støtte læger i udredningen af patienter for trombofili. Dette værktøj fremhæver relevant information i patientens journal og bidrager til øget nøjagtighed og effektivitet i udredningen. Studie IV identificerede dog også udfordringer ved klinisk anvendelse af kunstig intelligens, herunder tendensen til at stole for meget på systemet, hvilket understreger nødvendigheden af omhyggeligt design og løbende overvågning af de værktøjer, der implementeres i den kliniske hverdag.Samlet set viser denne afhandling, at NLP er et værdifuldt værktøj i forhold til blødning og VTE både i forskningen og den kliniske hverdag. Ved at identificere relevante risikofaktorer og kliniske oplysninger fra ustruktureret tekst i patientjournaler kan NLP omgå nogle af de begrænsninger, der er forbundet med registre og strukturerede databaser. Dette muliggør en mere nøjagtig identificering og prædiktion af større blødning og VTE. Studierne i afhandlingen illustrerer, hvordan NLP både kan anvendes i store epidemiologiske forskningsprojekter og i den kliniske hverdag. Det peger på, at NLP kan forbedre vores forståelse af VTE og større blødning og samtidig styrke kvaliteten af patientvaretagelsen.Blood clots and major bleeding are common and serious medical problems. Clots, specifically venous thromboembolisms (VTEs), can lead to life-threatening complications. On the other hand, major bleeding can be equally dangerous and unpredictable. Managing these conditions requires carefully balancing the risk of clotting against the risk of bleeding, which is often difficult because these risks change over time and vary greatly between individuals. This thesis combines traditional medical research with advanced computer methods to better understand who is at risk of VTE or bleeding.While traditional medical research has relied on registries and structured databases, electronic health records (EHRs) contain much more detailed and accurate information on VTE and bleeding, but much of it is written in unstructured clinical notes. This means the information cannot be easily accessed or analysed using traditional computer programs, and manually reviewing large volumes of clinical text is not realistic. To address this, an artificial intelligence (AI) method called natural language processing (NLP) enables computers to read and interpret free-text notes written by healthcare professionals, allowing relevant clinical information to be identified and extracted efficiently. The thesis includes four studies. Study I shows that different types of hereditary thrombophilia, which are genetic mutations that increase the risk of VTE, do not carry the same level of risk. The finding challenges long-standing assumptions and highlights the need for more precise risk assessments that consider both genetic and clinical context. Study II introduces a novel method to quantify blood loss during surgery by extracting information directly from the clinical notes, helping to define what can be considered excessive bleeding for different types of operations. This is an important step toward creating better guidelines for managing bleeding during and after surgery. Study III focuses on major bleeding in hospitalised medical patients, demonstrating that major bleeding is underrecognised if relying on diagnostic codes. By using NLP, the study identified bleeding events more accurately and developed models to predict which patients were most at risk. These models showed that routine blood tests and general health status were the most useful predictors. Study IV goes beyond research and shows how NLP can be implemented in real healthcare settings. A system was developed and integrated into daily workflow to support doctors when assessing patients for thrombophilia. This tool automatically highlights relevant information from the patient’s medical history, improving both accuracy and efficiency. However, Study IV also identified challenges related to clinical use, including relying too heavily on the AI system, which underlines the necessity of careful design and oversights of the tools implemented into clinical practice.In summary, this thesis demonstrates that NLP is a valuable tool for both research and clinical practice in the context of bleeding and VTE. By extracting relevant risk factors and clinical insights from unstructured EHR text, NLP helps overcome limitations of traditional data sources and enables more accurate identification and prediction of these complications. The studies included in the thesis illustrate how NLP can support largescale epidemiological research as well as clinical decision-making. This shows that NLP can help us better understand VTE and bleeding while also improving patient care.
Paediatric reference intervals are heterogeneous and differ considerably in the classification of healthy paediatric blood samples
No full textThe aim was to elude differences in published paediatric reference intervals (RIs) and the implementations hereof in terms of classification of samples. Predicaments associated with transferring RIs published elsewhere are addressed. A local paediatric (aged 0 days to < 18 years) population of platelet count, haemoglobin level and white blood cell count, based on first draw samples from general practitioners was established. PubMed was used to identify studies with transferable RIs. The classification of local samples by the individual RIs was evaluated. Transference was done in accordance with the Clinical and Laboratory Standards Institute EP28-A3C guideline. Validation of transference was done using a quality demand based on biological variance. Twelve studies with a combined 28 RIs were transferred onto the local population, which was derived from 20,597 children. Studies varied considerably in methodology and results. In terms of classification, up to 63% of the samples would change classification from normal to diseased, depending on which RI was applied. When validating the transferred RIs, one RI was implementable in the local population. Conclusion: Published paediatric RIs are heterogeneous, making assessment of transferability problematic and resulting in marked differences in classification of paediatric samples, thereby potentially affecting diagnosis and treatment of children.What is Known:• Reference intervals (RIs) are fundamental for the interpretation of paediatric samples and thus correct diagnosis and treatment of the individual child.• Guidelines for the establishment of adult RIs exist, but there are no specific recommendations for establishing paediatric RIs, which is problematic, and laboratories often implement RIs published elsewhere as a consequence.What is New:• Paediatric RIs published in peer-reviewed scientific journals differ considerably in methodology applied for the establishment of the RI.• The RIs show marked divergence in the classification of local samples from healthy children.</p
Application of adult reference intervals in children
No full textThe aim of this study was to evaluate to which extend adult reference intervals (RIs) could be applied in children. A local paediatric population (aged 1 to < 20 years), based on first draw samples from general practitioners (GPs), was established. Children with samples taken at a hospital or > 3 samples from GPs were excluded. Analytes evaluated included haematological, liver and pancreatic function, kidney function, electrolytes, and metabolism parameters. Applicability of adult RIs in children aged 1-17 years was evaluated using individuals aged 18-19 years as reference groups for the adult RIs. The local population consisted of 31,024 children with 282,721 analyses in total. For each analyte, 17 age strata and two gender strata were established. Partitioning was not warranted in 51% of the male strata and in 69% of the female strata. Adult RIs could be applied in 42% for children aged 1-< 10 years, 57% for children aged 10-< 15 years, and 85% for children aged 15-<18 years.Conclusion: for certain analytes, there is no need to partition between adult and paediatric RIs, but a need for age- and gender-specific RIs remains for several clinical laboratory tests.What is Known:• Establishing paediatric reference intervals (RIs) is time consuming, costly, and not feasible for many laboratories. Transference of RIs established elsewhere often leads to misclassification of paediatric laboratory results.• Adult RIs are often more easily established and validated.What is New:• Adult RIs can be applied to children as young as 2 years for some analytes. Conversely, for some analytes, adult RIs cannot be applied in children aged 1-17 years.• Laboratory data can be applied in evaluating the need for partitioning in reference intervals.</p
Natural language processing for identifying major bleeding risk in hospitalised medical patients
Get PDFBackground: Major bleeding is a severe complication in critically ill medical patients, resulting in significant morbidity, mortality, and healthcare costs. This study aims to assess the incidence and risk factors for major bleeding in hospitalised medical patients using a Natural Language Processing (NLP) model. Methods: We conducted a retrospective, cross-sectional observational study using electronic health records of adult patients admitted through the Emergency Department at Odense University Hospital from January 2017 to December 2022. Major bleeding during admission was identified and validated using a natural language model, with events classified according to current guidelines. Risk factors, including demographics, comorbidities, and biochemical values at admission, were evaluated. Two risk assessment models (RAMs) were developed using Cox proportional hazards regression. Validation included, bootstrapping, K-fold cross validation, and cluster analyses. Results: Of the 46,439 eligible patients, 1246 (2.7 %) experienced major bleeding. Risk factors for major bleeding included older age, male sex, alcohol consumption, higher systolic blood pressure, lower haemoglobin, and higher creatinine. RAM 1, which included biochemical data and comorbidities, demonstrated robust predictive performance (Harrell's C-statistic = 0.726). RAM 2, a simplified model without comorbidities, maintained similar predictive accuracy (C-statistic = 0.721), indicating its potential utility in clinical settings with limited resources for detailed patient histories. Results were consistent throughout validation. Conclusion: This study highlights the incidence and risk factors of major bleeding in medical patients, emphasizing the predictive value of routinely measured biochemical markers. Furthermore, it shows the applicability of NLP models in identifying bleeding episodes in EHR text.</p
Exploration of complement split products in plasma and urine as biomarkers of kidney graft rejection
Get PDFIntroduction: The complement system, consisting of more than thirty different soluble and cell-bound proteins, exerts essential functions both in the innate and adaptive immune systems and is believed to be an important contributor to allograft injury in kidney transplantation. The anaphylatoxins C3a and C5a are powerful chemoattractants, recruiting immune effector cells toward the site of complement activation and enhance T-cell response, while C3dg binding to CR2 on B-cells, enhances B-cell immunity at several stages of the B-cell differentiation. Complement split products in plasma and urine could reflect ongoing inflammation and tissue injury. We, therefore, investigated if complement split products increase in plasma and urine in kidney transplant recipients with rejection. Method: In this case-control feasibility study, complement factors C3a, C3dg, C4a, and C5a were measured in plasma and C3dg and sC5b-9 associated C9 neoantigen in urine in 15 kidney transplant recipients with rejection (cases) and 15 kidney transplant recipients without (controls). The groups were matched on the type of transplantation and the time from transplantation to sampling. The complement split products were compared (i) between cases and controls and (ii) within the rejection group over time, comparing the measurements at rejection with measurements where the kidney transplant recipients were clinically stable. Possible moderators were explored, and results adjusted accordingly. P values < 0.05 were considered significant. Plasma C3dg was analyzed by immune-electrophoresis, plasma C3a, plasma C4a, and plasma C5a by flow cytometry, and urine C3dg and urine C9neo by ELISA. Results: In plasma, there were no significant differences between the rejection and the control group. However, steroids and pretransplant C3dg levels significantly influenced C3dg. Within the rejection group, plasma C3a and C3dg were significantly higher at the time of rejection compared to the stable phase (p < 0.01). In urine, C3dg/creatinine and C9 neoantigen/creatinine ratios were not different between the rejection and the control group. Urine C3dg/creatinine and urine C9 neoantigen/creatinine ratios correlated to urine albumin and significantly increased after the transplantation (p < 0.001). Conclusion: This study shows increased plasma C3a and C3dg in kidney transplant recipients, primarily with T cell mediated rejection. This finding suggests that consecutive measurements of C3a and C3dg in plasma could be applicable to monitor alloreactivity in kidney transplant recipients. Urine complement split products are unsuitable as rejection biomarkers since the permeability of the glomerular filtration barrier strongly influences them. Prospective longitudinal studies on plasma C3a and C3dg dynamics will be needed to validate present findings.</p
Only one-third of referrals for fatty liver disease are on time: real-world study reveals opportunities to avoid unnecessary and delayed referrals
No full textBACKGROUND AND AIMS: Fatty liver disease is a global health concern, but in the absence of specific guidelines, current referral patterns differ according to the preferences of the general practitioners. Outpatient Gastroenterology clinics spend futile resources on liver-healthy patients while diagnosing decompensated patients delayed. We aimed to describe referral patterns to a regional outpatient Gastroenterology clinic.METHODS: We reviewed 9684 referrals from primary care for suspected liver disease in the years 2016-2017, during two years. Data were extracted from the patients' hospital records to assess the clinical workup and patient outcomes until a mean of 43 months after the time of referral. Referrals were categorized as unnecessary (no signs of liver disease), timely (significant fibrosis/compensated cirrhosis), or delayed (decompensated cirrhosis).RESULTS: We included 375 patient referrals from primary care. The main reason for referral was elevated transaminases. More than half (54%) of patients had no signs of liver disease, being unnecessarily referred for evaluation, while 17% had decompensated liver disease and were thus referred too late.CONCLUSIONS: Only one-third of patients referred on suspicion of liver disease were referred on time, either before presenting with decompensated liver cirrhosis or with some evidence of significant liver disease, e.g., liver fibrosis. There is a huge unmet need for clinical referral pathways in primary care. Strengths and Limitations of this StudyA strength of this study is the complete mapping of all potential referrals to the outpatient clinic in the two-year period. Instead of retrieving the historic data by ICD-10 diagnosis codes, and reflecting only those patients where the GP clearly suspects liver disease, we have a strong reliance on our methods. We screened all potentially relevant referrals, e.g., referrals due to weight loss or fatigue, which may reflect symptoms of cirrhosis. Thereby we are confident that we have not missed any patients that originally were referred with unspecific symptoms, but after evaluation are diagnosed with liver disease.Another strength of our study is the long follow-up period, which allows us to fully evaluate the course for the individual patient, and the potential later coming diagnoses.Finally, it is a strength of the study that we were not exclusive to one liver disease etiology, both ALD and NAFLD etiology were included in the study.A limitation of this study is the use of historic data, and the fact that it is a single-center study, showing only the referral patterns in one outpatient Gastroenterology clinic.</p
Investigating anatomical bias in clinical machine learning algorithms
Get PDFClinical machine learning algorithms have shown promising results and could potentially be implemented in clinical practice to provide diagnosis support and improve patient treatment. Barriers for realisation of the algorithms’ full potential include bias which is systematic and unfair discrimination against certain individuals in favor of others. The objective of this work is to measure anatomical bias in clinical text algorithms. We define anatomical bias as unfair algorithmic outcomes against patients with medical conditions in specific anatomical locations. We measure the degree of anatomical bias across two machine learning models and two Danish clinical text classification tasks, and find that clinical text algorithms are highly prone to anatomical bias. We argue that datasets for creating clinical text algorithms should be curated carefully to isolate the effect of anatomical location in order to avoid bias against patient subgroups.</p
Correction to: Hidden Present, Visible Absent in the City of Dreams: Assembling the Collective Imagination
No full textThe original version of this article unfortunately contained a mistake. The name of “Tuuli Pern” is now corrected in the author group of this article. The original article has been corrected.</p
Hidden Present, Visible Absent in the City of Dreams: Assembling the Collective Imagination
No full textThis paper serves as a collaborative auto-ethnography of psychological researchers, engaged in a unique encounter with each other and with the streets, artefacts, history and ghosts of Vienna, the City of Dreams. This small international and interdisciplinary group engaged in four pre-planned exercises in this city, each geared towards developing the sensitivity of researchers to notions of embodied introspection. Participants were asked to recollect and diarise their internal dialogue and these voices were assembled according to the practice of bricolage. This paper aims to demonstrate how new forms of knowledge might be created, based on the material experience of place, and the assembling of the collective imagination of researchers. It also aims to demonstrate how this collective imagination might be written about in novel ways, with a decentred author capturing the atmosphere while it lasts
Deep learning detects and visualizes bleeding events in electronic health records
Get PDFBackground: Bleeding is associated with a significantly increased morbidity and mortality. Bleeding events are often described in the unstructured text of electronic health records, which makes them difficult to identify by manual inspection. Objectives: To develop a deep learning model that detects and visualizes bleeding events in electronic health records. Patients/Methods: Three hundred electronic health records with International Classification of Diseases, Tenth Revision diagnosis codes for bleeding or leukemia were extracted. Each sentence in the electronic health record was annotated as positive or negative for bleeding. The annotated sentences were used to develop a deep learning model that detects bleeding at sentence and note level. Results: On a balanced test set of 1178 sentences, the best-performing deep learning model achieved a sensitivity of 0.90, specificity of 0.90, and negative predictive value of 0.90. On a test set consisting of 700 notes, of which 49 were positive for bleeding, the model achieved a note-level sensitivity of 1.00, specificity of 0.52, and negative predictive value of 1.00. By using a sentence-level model on a note level, the model can explain its predictions by visualizing the exact sentence in a note that contains information regarding bleeding. Moreover, we found that the model performed consistently well across different types of bleedings. Conclusions: A deep learning model can be used to detect and visualize bleeding events in the free text of electronic health records. The deep learning model can thus facilitate systematic assessment of bleeding risk, and thereby optimize patient care and safety.</p
