Thrift, Amanda G.

AbstractINTRODUCTIONIdentifying individuals with dementia is crucial for prevalence estimation and service planning, but reliable, scalable methods are lacking. We developed novel set algorithms using both structured and unstructured electronic health record (EHR) data, applying Diagnostic and Statistical Manual of Mental Disorders criteria for dementia case identification.METHODSOur cohort (n = 1082) included individuals aged ≥ 60 with dementia identified through specialist clinics and a comparison group without dementia. Clinicians from Australia and the United States informed predictor selection. We developed algorithms through a biostatistics stream for structured data and a natural language processing (NLP) stream for text, synthesizing results via logistic regression.RESULTSThe final structured model retained 16 variables (area under the receiver operating characteristic curve [AUC] 0.853, specificity 72.2%, sensitivity 80.6%). NLP classifiers (logistic regression, support vector machine, and random forest models) performed comparably. The final, combined model outperformed all others (AUC = 0.951, P < 0.001 for comparison to structured model).DISCUSSIONEmbedding text‐derived insights within algorithms trained on structured medical data significantly enhances dementia identification capacity.Highlights Algorithmic tools for detection of individuals with dementia are available; however, previous work has used heterogeneous case definitions which are not clinically meaningful, and has relied on proxies such as diagnostic codes or medications for case ascertainment. We used a novel, dual‐stream algorithmic development approach, simultaneously and separately modeling a clinically meaningful outcome (diagnosis of dementia according to specialized clinical impression) using structured and unstructured electronic health record datasets. Our clinically grounded case definition supported the inclusion of key structured variables (such as dementia International Classification of Disease codes and medications) as modeling predictors rather than outcomes. Our algorithms, published in detail to support validation and replication, represent a major step forward in the use of routinely collected data for detection of diagnosed dementia.

AbstractBackgroundPopulation dementia prevalence is traditionally estimated using cohort studies, surveys, routinely‐collected administrative data, and registries. Hospital Electronic Health Records (EHRs) are comprised of rich structured and unstructured (text) clinical data that are underutilised for this purpose.We aimed to develop a suite of algorithms using routinely‐collected EHR data to reliably identify cases of dementia, as a key step towards incorporating such data in prevalence estimation. Towards this, we developed a novel predictive framework integrating data‐science and biostatistical methods.MethodTraining data were sourced via the National Centre for Healthy Ageing (NCHA) Data Platform, a linked, curated, EHR‐derived data warehouse. Individuals within the platform catchment aged >60 years with confirmed dementia were identified through hospital specialist dementia clinics. A comparison group of individuals aged >60years with EHR records without dementia was recruited from the community.A panel of clinical experts (Neurology, Geriatric Medicine) informed variable and concept selection and guided data cleaning efforts within both streams. Algorithms were developed via two work‐streams; a traditional biostatistical approach to fit logistic regression models using structured data elements, and a data science stream used Natural Language Processing (NLP) to fit models to the unstructured (text) parts of the EHR, for the same individuals.ResultOf 568 individuals (362 with dementia), 434 had clinical notes available. In the data science stream using unstructured data, among a range of NLP derived models, the Random Forest classifier performed best in assigning dementia status, with Area Under the Curve (AUC) 0.95, specificity 90.2% and sensitivity 88.4%. In the biostatistics stream, 15 structured variables were included in the final model, covering demographics, health service attendance, medications, and ICD‐10 Codes, with AUC 0.94, specificity 85.9% and sensitivity 85.6%.ConclusionArtificial intelligence techniques applied to unstructured electronic health data and guided by human clinical expertise may be powerful tools in capturing the presence of dementia, at least comparable to traditional techniques using structured data, and conferring practical and scientific advantages for dementia prevalence estimation. Future validation is required in less crisply delineated real‐world settings.

AbstractObjectiveWe performed a pilot stroke incidence study, focused on feasibility and inclusion of the CONSIDER reporting guidelines, to model the design of a future population‐based study aiming to definitively determine stroke incidence, antecedents, treatment, and outcomes.Study designProspective stroke incidence study (pilot study).Setting, participantsAll people aged 15 years or older who lived in postcode‐defined areas of South Australia and Northern Territory (885 472 people, including 45 127 Aboriginal people [5.1%]) diagnosed with stroke for the first time during 1 October – 31 December 2015 and admitted to public hospitals or stroke and transient ischaemic attack clinics.Main outcome measuresFeasibility of a prospective population‐based stroke incidence study.ResultsOf the 123 participants with first strokes, ten were Aboriginal (8%); the median age of Aboriginal people was 45 years (interquartile range [IQR], 33–55 years), of non‐Indigenous people 73 years (IQR, 62–84 years). For Aboriginal people, the age‐standardised incidence of stroke was 104 (95% confidence interval [CI], 84–124) per 100 000 person‐years, for non‐Indigenous people 33 (95% CI, 22–44) per 100 000 person‐years. We found that a prospective population‐based stroke incidence study in Aboriginal people was feasible, including with respect to establishing an adequate sample size, diagnostic confirmation, identification of incident stroke, confirming stroke subtypes, establishing a stable statistical population, standardising data reporting for comparison with other stroke incidence studies, and ethical research reporting that conforms to CONSIDER guidelines.ConclusionsA larger, population‐based study of the incidence of stroke in Aboriginal people is both feasible and needed to provide robust estimates of stroke incidence, antecedents, treatments and outcomes to help guide strategies for reducing the risk of and outcomes of stroke in Aboriginal people.

Background: Most strokes and cardiovascular diseases (CVDs) are potentially preventable if their risk factors are identified and well controlled. Digital platforms, such as the PreventS-MD web app (PreventS-MD) may aid health care professionals (HCPs) in assessing and managing risk factors and promoting lifestyle changes for their patients. Methods: This is a mixed-methods cross-sectional two-phase survey using a largely positivist (quantitative and qualitative) framework. During Phase 1, a prototype of PreventS-MD was tested internationally by 59 of 69 consenting HCPs of different backgrounds, age, sex, working experience, and specialties using hypothetical data. Collected comments/suggestions from the study HCPs in Phase 1 were reviewed and implemented. In Phase 2, a near-final version of PreventS-MD was developed and tested by 58 of 72 consenting HCPs using both hypothetical and real patient (n = 10) data. Qualitative semi-structured interviews with real patients (n = 10) were conducted, and 1 month adherence to the preventive recommendations was assessed by self-reporting. The four System Usability Scale (SUS) groups of scores (0–50 unacceptable; 51–68 poor; 68–80.3 good; >80.3 excellent) were used to determine usability of PreventS-MD. Findings: Ninety-nine HCPs from 27 countries (45% from low- to middle-income countries) participated in the study, and out of them, 10 HCPs were involved in the development of PreventS before the study, and therefore were not involved in the survey. Of the remaining 89 HCPs, 69 consented to the first phase of the survey, and 59 of them completed the first phase of the survey (response rate 86%), and 58 completed the second phase of the survey (response rate 84%). The SUS scores supported good usability of the prototype (mean score = 80.2; 95% CI [77.0–84.0]) and excellent usability of the final version of PreventS-MD (mean score = 81.7; 95% CI [79.1–84.3]) in the field. Scores were not affected by the age, sex, working experience, or specialty of the HCPs. One-month follow-up of the patients confirmed the high level of satisfaction/acceptability of PreventS-MD and (100%) adherence to the recommendations. Interpretation: The PreventS-MD web app has a high level of usability, feasibility, and satisfaction by HCPs and individuals at risk of stroke/CVD. Individuals at risk of stroke/CVD demonstrated a high level of confidence and motivation in following and adhering to preventive recommendations generated by PreventS-MD.

Background The greatest potential to reduce the burden of stroke is by primary prevention of first-ever stroke, which constitutes three quarters of all stroke. In addition to population-wide prevention strategies (the ‘mass’ approach), the ‘high risk’ approach aims to identify individuals at risk of stroke and to modify their risk factors, and risk, accordingly. Current methods of assessing and modifying stroke risk are difficult to access and implement by the general population, amongst whom most future strokes will arise. To help reduce the burden of stroke on individuals and the population a new app, the Stroke Riskometer™, has been developed. We aim to explore the validity of the app for predicting the risk of stroke compared with current best methods. Methods 752 stroke outcomes from a sample of 9501 individuals across three countries (New Zealand, Russia and the Netherlands) were utilized to investigate the performance of a novel stroke risk prediction tool algorithm (Stroke Riskometer™) compared with two established stroke risk score prediction algorithms (Framingham Stroke Risk Score [FSRS] and QStroke). We calculated the receiver operating characteristics (ROC) curves and area under the ROC curve (AUROC) with 95% confidence intervals, Harrels C-statistic and D-statistics for measure of discrimination, R2 statistics to indicate level of variability accounted for by each prediction algorithm, the Hosmer-Lemeshow statistic for calibration, and the sensitivity and specificity of each algorithm. Results The Stroke Riskometer™ performed well against the FSRS five-year AUROC for both males (FSRS = 75·0% (95% CI 72·3%–77·6%), Stroke Riskometer™ = 74·0(95% CI 71·3%–76·7%) and females [FSRS = 70·3% (95% CI 67·9%–72·8%, Stroke Riskometer™ = 71·5% (95% CI 69·0%–73·9%)], and better than QStroke [males–59·7% (95% CI 57·3%–62·0%) and comparable to females = 71·1% (95% CI 69·0%–73·1%)]. Discriminative ability of all algorithms was low (C-statistic ranging from 0·51–0·56, D-statistic ranging from 0·01–0·12). Hosmer-Lemeshow illustrated that all of the predicted risk scores were not well calibrated with the observed event data ( P < 0·006). Conclusions The Stroke Riskometer™ is comparable in performance for stroke prediction with FSRS and QStroke. All three algorithms performed equally poorly in predicting stroke events. The Stroke Riskometer™ will be continually developed and validated to address the need to improve the current stroke risk scoring systems to more accurately predict stroke, particularly by identifying robust ethnic/race ethnicity group and country specific risk factors.