BMJ Health & Care Informatics最新文献

Background: Low-density lipoprotein cholesterol (LDL-C) and non-high-density lipoprotein cholesterol (non-HDL-C) levels are paramount in atherosclerotic cardiovascular disease risk management. However, 94.4% of Thai young adult are unaware of their condition. A diagnostic prediction model may assist in screening and alleviating underdiagnosis.

Objectives: Development and internal validation of diagnostic prediction models on elevated LDL-C (≥160 mg/dL) and non-HDL-C (≥160 mg/dL).

Methods: Retrospective, single-centre, tertiary-care hospital annual health examination data from 29 March 2018 to 30 August 2023 was analysed. Two models with 11 predictors from anthropometry and bioimpedance are fitted with multivariable binary logistic regression predicting elevated LDL-C and non-HDL-C. Predictor selection used the backward stepwise elimination. Four performance metrics were quantified: discrimination using area under the receiver-operating characteristic curve (AuROC); calibration by calibration plot; utility by decision curve analysis and instability by performance instability plots. Internal validation was carried out using 500 repetitions of bootstrap-resampling.

Results: Dataset included 2222 LDL-C and 5149 non-HDL-C investigations, 303 were classed as elevated LDL-C (13.6%) and 1013 as elevated non-HDL-C cases (19.7%). Two predictors, gender and metabolic age, were identified in the LDL-C model with AuROC 0.639 (95% CI 0.617 to 0.661), poor calibration, and utility in the 7%-25% probability range. Three predictors-gender, diastolic blood pressure and metabolic age-were identified in the non-HDL-C model with AuROC 0.722 (95% CI 0.705 to 0.738), good calibration and utility in 9%-55% probability range.

Discussion and conclusion: Overall results demonstrated acceptable discrimination for non-HDL-C model but inadequate performance of LDL-C model for clinical practice. An external validation study should be planned for non-HDL-C model.

Objectives: Barcode medication administration (BCMA) systems may improve patient safety with successful integration and use. This study aimed to explore the barriers and enablers for the successful use of a BCMA system by examining the patterns of medication and patient scanning over time and potential safety implications.

Methods: Retrospective longitudinal study informed by prospective clinical observations using data extracted from five hospital wards over the first 16 months after implementation to determine trends in medication and patient scanning rates, reasons for non-compliance and scanning mismatch alerts. Regression models were applied to explore factors influencing medication scanning rates across wards of different specialties.

Results: Electronic data on 613 868 medication administrations showed overall medication scanning rates per ward ranged from 5.6% to 67% and patient scanning rates from 4.6% to 89%. Reported reasons for not scanning medications were 'barcode not readable' and 'unavailability of scanners'. Scanning rates declined over time and the pattern of reason codes for not scanning also changed. Factors associated with higher scanning rates included a locally led quality improvement (QI) initiative, the medication administration time and the medication formulation, for example, tablets and liquids. Overall, 37% of scanning alerts resulted in a change in user action. Staff tried to comply with the BCMA system workflow, but workarounds were observed.

Discussion: Compliance with BCMA systems varied across wards and changed over time. QI initiatives hold promise to ensure sustained use of BCMA systems.

Conclusions: BCMA systems may help to improve medication safety, but further research is needed to confirm sustained safety benefits.

Background: Loneliness and insomnia are mutually occurring conditions. This paper investigates whether keywords depicting loneliness and insomnia are expressed together on social media. Understanding loneliness through data fills the gaps or validates the literature on loneliness from sociological and psychological perspectives. Loneliness is associated with various physical and mental health conditions but there are opportunities to understand it from the perspectives and lens of health informatics through social media data. Because loneliness is a subjective phenomenon, therefore, the self-reporting nature of social media data can provide an intimate glimpse into the feelings associated with loneliness.

Methods: This study uses sentiment analysis of collected tweets on loneliness and insomnia to filter out the tweets that have negative connotations. Those tweets are then further analysed to find out categories and themes associated with loneliness and insomnia.

Results: Through the frequency of word occurrence analysis, it was seen that the association between loneliness and insomnia can be found. The association, in the tweets analysed, is mediated by words denoting depressive symptoms. Moreover, the themes and categories which are associated with the expression of both loneliness and insomnia are those of personal feelings and time.

Background: Telecare may provide an alternative to maintaining post-acute stroke care services in making benefit to both the providers and the stroke survivors, although study is needed to investigate its feasibility and effectiveness in integrating this innovative delivery mode into a routine.

Objectives: The objectives of this study are to assess the feasibility and effectiveness of telecare consultations in a nurse-led post-acute stroke clinic.

Methods: A pre- and post-test one group quasi-experimental design was adopted. Subjects were recruited in the clinic and received three secondary stroke care consultations in 3 months via telecare from stroke nurses. Data were collected at pre- and post-intervention. A Wilcoxon signed-rank test was used to compare the two time-points for differences in effectiveness.

Results: Ninety-two stroke survivors participated. The drop-out rate was 27%. The majority perceived the programme as time-friendly and cost-saving and as alleviating their health-related worries. At the 3-month follow-up, notable improvements were observed in the activities of daily living and the strength domain of stroke-specific quality of life.

Conclusions: Integrating telecare consultations within nurse-led stroke clinics is a feasible and acceptable strategy for monitoring the health and fostering the self-care abilities of individuals following their discharge from hospital after an acute stroke episode.

Objectives: We aimed to evaluate the performance of multiple large language models (LLMs) in data extraction from unstructured and semi-structured electronic health records.

Methods: 50 synthetic medical notes in English, containing a structured and an unstructured part, were drafted and evaluated by domain experts, and subsequently used for LLM-prompting. 18 LLMs were evaluated against a baseline transformer-based model. Performance assessment comprised four entity extraction and five binary classification tasks with a total of 450 predictions for each LLM. LLM-response consistency assessment was performed over three same-prompt iterations.

Results: Claude 3.0 Opus, Claude 3.0 Sonnet, Claude 2.0, GPT 4, Claude 2.1, Gemini Advanced, PaLM 2 chat-bison and Llama 3-70b exhibited an excellent overall accuracy >0.98 (0.995, 0.988, 0.988, 0.988, 0.986, 0.982, 0.982, and 0.982, respectively), significantly higher than the baseline RoBERTa model (0.742). Claude 2.0, Claude 2.1, Claude 3.0 Opus, PaLM 2 chat-bison, GPT 4, Claude 3.0 Sonnet and Llama 3-70b showed a marginally higher and Gemini Advanced a marginally lower multiple-run consistency than the baseline model RoBERTa (Krippendorff's alpha value 1, 0.998, 0.996, 0.996, 0.992, 0.991, 0.989, 0.988, and 0.985, respectively).

Discussion: Claude 3.0 Opus, Claude 3.0 Sonnet, Claude 2.0, GPT 4, Claude 2.1, Gemini Advanced, PaLM 2 chat bison and Llama 3-70b performed the best, exhibiting outstanding performance in both entity extraction and binary classification, with highly consistent responses over multiple same-prompt iterations. Their use could leverage data for research and unburden healthcare professionals. Real-data analyses are warranted to confirm their performance in a real-world setting.

Conclusion: Claude 3.0 Opus, Claude 3.0 Sonnet, Claude 2.0, GPT 4, Claude 2.1, Gemini Advanced, PaLM 2 chat-bison and Llama 3-70b seem to be able to reliably extract data from unstructured and semi-structured electronic health records. Further analyses using real data are warranted to confirm their performance in a real-world setting.

In this perspective article, we consider the use of predictive models in healthcare and associated challenges. We will argue that patients can play a valuable role in supporting the safe and practicable embedding of such tools and provide some examples.

Background: Prognostic models help aid medical decision-making. Various prognostic models are available via websites such as MDCalc, but these models typically predict one outcome, for example, stroke risk. Each model requires individual predictors, for example, age, lab results and comorbidities. There is no clinical tool available to predict multiple outcomes from a list of common medical predictors.

Objective: Identify a constrained set of outcome-agnostic predictors.

Methods: We proposed a novel technique aggregating the standardised mean difference across hundreds of outcomes to learn a constrained set of predictors that appear to be predictive of many outcomes. Model performance was evaluated using the constrained set of predictors across eight prediction tasks. We compared against existing models, models using only age/sex predictors and models without any predictor constraints.

Results: We identified 67 predictors in our constrained set, plus age/sex. Our predictors included illnesses in the following categories: cardiovascular, kidney/liver, mental health, gastrointestinal, infectious and oncologic. Models developed using the constrained set of predictors achieved comparable discrimination compared with models using hundreds or thousands of predictors for five of the eight prediction tasks and slightly lower discrimination for three of the eight tasks. The constrained predictor models performed as good or better than all existing clinical models.

Conclusions: It is possible to develop models for hundreds or thousands of outcomes that use the same small set of predictors. This makes it feasible to implement many prediction models via a single website form. Our set of predictors can also be used for future models and prognostic model research.

Aim: We aimed to identify enablers and barriers of using primary care routine data for healthcare research, to formulate recommendations for improving efficiency in knowledge discovery.

Background: Data recorded routinely in primary care can be used for estimating the impact of interventions provided within routine care for all people who are clinically eligible. Despite official promotion of 'efficient trial designs', anecdotally researchers in the Asthma UK Centre for Applied Research (AUKCAR) have encountered multiple barriers to accessing and using routine data.

Methods: Using studies within the AUKCAR portfolio as exemplars, we captured limitations, barriers, successes, and strengths through correspondence and discussions with the principal investigators and project managers of the case studies.

Results: We identified 14 studies (8 trials, 2 developmental studies and 4 observational studies). Investigators agreed that using routine primary care data potentially offered a convenient collection of data for effectiveness outcomes, health economic assessment and process evaluation in one data extraction. However, this advantage was overshadowed by time-consuming processes that were major barriers to conducting efficient research. Common themes were multiple layers of information governance approvals in addition to the ethics and local governance approvals required by all health service research; lack of standardisation so that local approvals required diverse paperwork and reached conflicting conclusions as to whether a study should be approved. Practical consequences included a trial that over-recruited by 20% in order to randomise 144 practices with all required permissions, and a 5-year delay in reporting a trial while retrospectively applied regulations were satisfied to allow data linkage.

Conclusions: Overcoming the substantial barriers of using routine primary care data will require a streamlined governance process, standardised understanding/application of regulations and adequate National Health Service IT (Information Technology) capability. Without policy-driven prioritisation of these changes, the potential of this valuable resource will not be leveraged.

Background: In Australia, with the recent introduction of electronic health records (EHRs) into hospitals, the use of hospital-based EHRs for research is a relatively new concept. The aim of this study was to explore the attitudes of older healthcare consumers on sharing their health data with an emerging EHR-based Research Data Platform within the National Centre for Healthy Ageing.

Methods: This was a qualitative study. Two workshops were conducted in March 2022 with consumer representatives across Peninsula Health, Victoria, Australia. The workshops comprised three parts: (1) an ice-breaker (2) an introduction to EHR-based research through the presentation of 'use case' scenarios and (3) focus group discussions. Qualitative data were analysed using reflexive thematic analysis.

Results: Consumer participants (n=16) were aged between 62 and 83 years and were of mixed gender. The overarching theme was related to trust in the use of EHR data for research; themes included: (1) benefits of sharing data, (2) uncertainty around data collection processes and (3) data sharing fears. The three themes within the overarching theme all reflect participants' levels of trust.

Conclusion: Our study identified fundamental issues related to trust in the use of EHR data for research, with both healthcare and broader societal factors contributing to consumer attitudes. Processes to support transparent and clear communication with consumers are essential to support the responsible use of EHR data for research.