OBJECTIVE:Inequalities in the antiretroviral therapy (ART) cascade across subpopulations remain an ongoing challenge in the global HIV response. Eswatini achieved the UNAIDS 95-95-95 ART cascade targets by 2020, with differentiated programs to minimize inequalities across subpopulations, including for female sex workers (FSWs) and their clients. We sought to estimate the impacts of this achievement, through a retrospective impact evaluation of ART scale-up in Eswatini. DESIGN:Drawing on population-level and FSW-specific surveys, we developed a compartmental model of heterosexual HIV transmission, and calibrated it to observed HIV prevalence, incidence, and ART cascade scale-up in Eswatini. METHODS:We defined four counterfactual scenarios in which the population overall reached only 80-80-90 by 2020, but where FSW, clients, both, or neither were disproportionately left behind, reaching only 60-40-80. We estimated additional HIV infections by 2020 in counterfactual vs. observed scenarios, and identified epidemic conditions which maximized differences. RESULTS:Compared with observed cascade scale-up in Eswatini, leaving behind neither FSW nor their clients led to median (95% confidence interval, 95% CI) 8.8 (6.3-10.9) additional infections by 2020 vs. 14.3 (10.8-18.6) if both were left behind, a 63 (31-128) increase. The impact of leaving behind FSW and/or clients was largely determined by their population sizes and HIV incidence ratio among clients vs. men overall. CONCLUSION:Inequalities in the ART cascade across subpopulations can undermine the anticipated prevention impacts of cascade scale-up. As Eswatini has shown, addressing inequalities in the ART cascade that intersect with transmission risk can maximize incidence reductions from cascade scale-up.

Researchers studying hard-to-reach or minority populations are increasingly implementing nonprobability sampling strategies that are often prone to selection bias. To address this problem, existing statistical methods suggest integrating data from external probability sample, often collected by government agencies, with the nonprobability sample from the hard-to-reach population. These methods assume that all information collected in the probability sample is recorded without errors. This may not be the case if participants are unwilling to report their minority status, such as sexual orientation, truthfully in large-scale population-based surveys, leading to misclassification bias. In this paper, we propose a novel two-step approach aimed at addressing misclassification bias in the probability sample to improve the performance of the data integration methods aimed at addressing selection bias in the nonprobability sample. By applying the proposed method to simulated data, we demonstrate a significant reduction in bias and validate the proposed bootstrap variance estimator of the estimated mean (prevalence) under low, moderate, and high misclassification rates. This method is particularly beneficial when the misclassification rate is high. Finally, we illustrate the application of the two-step approach to estimate the prevalence of measures of social connectedness among sexual minority men using a real-world nonprobability sample.

Background:Inequalities in the antiretroviral therapy (ART) cascade across subpopulations remain an ongoing challenge in the global HIV response. Eswatini achieved the UNAIDS 95-95-95 targets by 2020, with differentiated programs to minimize inequalities across subpopulations, including for female sex workers (FSW) and their clients. We sought to estimate additional HIV infections expected in Eswatini if cascade scale-up had not been equal, and under which epidemic conditions these inequalities could have the largest influence. Methods:Drawing on population-level and FSW-specific surveys in Eswatini, we developed a compartmental model of heterosexual HIV transmission which included eight subpopulations and four sexual partnership types. We calibrated the model to stratified HIV prevalence, incidence, and ART cascade data. Taking observed cascade scale-up in Eswatini as the base-case-reaching 95-95-95 in the overall population by 2020-we defined four counterfactual scenarios in which the population overall reached 80-80-90 by 2020, but where FSW, clients, both, or neither were disproportionately left behind, reaching only 60-40-80. We quantified relative additional cumulative HIV infections by 2030 in counterfactual vs base-case scenarios. We further estimated linear effects of viral suppression gap among FSW and clients on additional infections by 2030, plus effect modification by FSW/client population sizes, rates of turnover, and HIV prevalence ratios. Results:Compared with the base-case scenario, leaving behind neither FSW nor their clients led to the fewest additional infections by 2030: median (95% credible interval) 14.9 (10.4, 18.4)% vs 26.3 (19.7, 33.0)% if both were left behind-a 73 (40, 149)% increase. The effect of lower cascade on additional infections was larger for clients vs FSW, and both effects increased with population size and relative HIV incidence. Conclusions:Inequalities in the ART cascade across subpopulations can undermine the anticipated prevention impacts of cascade scale-up. As Eswatini has shown, addressing inequalities in the ART cascade, particularly those that intersect with high transmission risk, could maximize incidence reductions from cascade scale-up.

The prevalence and relative disparities of mental health outcomes and well-being indicators are often inconsistent across studies of sexual minority men (SMM) due to selection biases in community-based surveys (nonprobability sample), as well as misclassification biases in population-based surveys where some SMM often conceal their sexual orientation identities. The present study estimated the prevalence of mental health related outcomes (depressive symptoms, mental health service use, anxiety) and well-being indicators (loneliness and self-rated mental health) among SMM, broken down by sexual orientation using the adjusted logistic propensity score (ALP) weighting. We applied the ALP to correct for selection biases in the 2019 Sex Now data (a community-based survey of SMMs in Canada) by reweighting it to the 2015-2018 Canadian Community Health Survey (a population survey from Statistics Canada). For all SMMs, the ALP-weighted prevalence of depressive symptoms was 15.96% (95% CI, 11.36%-23.83%), while for mental health service use, it was 32.13% (95% CI, 26.09%-41.20%). The ALP estimates lie in between the crude estimates from the two surveys. This method was successful in providing a more accurate estimate than relying on results from one survey alone. We recommend to the use of ALP on other minority populations under certain assumptions.This article is part of a Special Collection on Mental Health.

This work aimed to identify pre-existing health conditions of patients with traumatic brain injury (TBI) and develop predictive models for the first TBI event and its external causes by employing a combination of unsupervised and supervised learning algorithms. We acquired up to five years of pre-injury diagnoses for 488,107 patients with TBI and 488,107 matched control patients who entered the emergency department or acute care hospitals between April 1st, 2002, and March 31st, 2020. Diagnoses were obtained from the Ontario Health Insurance Plan (OHIP) database which contains province-wide claims data by physicians in Ontario, Canada for inpatient and outpatient services. A screening process was conducted on the OHIP diagnostic codes to limit the subsequent analysis to codes that were predictive of TBI, which concluded that 314 codes were significantly associated with TBI. The Latent Dirichlet Allocation (LDA) model was applied to the diagnostic codes and generated an optimal number of 19 topics that concur with published literature but also suggest other unexplored areas. Estimated word-topic probabilities from the LDA model helped us detect pre-morbid conditions among patients with TBI by uncovering the underlying patterns of diagnoses, meanwhile estimated document-topic probabilities were utilized in variable creation as form of a dimension reduction. We created 19 topic scores for each patient in the cohort which were utilized along with socio-demographic factors for Random Forest binary classifier models. Test set performances evaluated using area under the receiver operating characteristic curve (AUC) were: TBI event (AUC = 0.85), external cause of injury: falls (AUC = 0.85), struck by/against (AUC = 0.83), cyclist collision (AUC = 0.76), motor vehicle collision (AUC = 0.83). Our analysis successfully demonstrated the feasibility of using machine learning to predict TBI due to various external causes and identified the most important factors that contribute to this prediction.

Clinical information collected in electronic health records (EHRs) is becoming an essential source to emulate randomized experiments. Since patients do not interact with the healthcare system at random, the longitudinal information in large observational databases must account for irregular visits. Moreover, we need to also account for subject-specific unmeasured confounders which may act as a common cause for treatment assignment mechanism (e.g. glucose-lowering medications) while also influencing the outcome (e.g. Hemoglobin A1c). We used the calibration of longitudinal weights to improve the finite sample properties and to account for subject-specific unmeasured confounders. A Monte Carlo simulation study is conducted to evaluate the performance of calibrated inverse probability estimators using time-dependent treatment assignment and irregular visits with subject-specific unmeasured confounders. The simulation study showed that the longitudinal weights with calibrated restrictions improved the finite sample bias when compared to the stabilized weights. The application of the calibrated weights is demonstrated using the exposure of glucose lowering medications and the longitudinal outcome of Hemoglobin A1c. Our results support the effectiveness of glucose lowering medications in reducing Hemoglobin A1c among type II diabetes patients with elevated glycemic index ( $$\ge 8.5 \%$$ ) using stabilized and calibrated weights.

IntroductionWe investigated the association between sleep disorders (SDs) and incident dementia in adults with traumatic brain injury (TBI). MethodsAdults with a TBI between 2003 and 2013 were followed until incident dementia. Sleep disorders at TBI were predictors in Cox regression models, controlling for other dementia risks. ResultsOver 52 months, 4.6% of the 712,708 adults (59% male, median age 44, <1% with SD) developed dementia. An SD was associated with a 26% and a 23% of increased risk of dementia in male and female participants (hazard ratio [HR] 1.26, 95% confidence interval [CI] 1.11-1.42 and HR 1.23, 95% CI 1.09-1.40, respectively). In male participants, SD was associated with a 93% increased risk of early-onset dementia (HR 1.93, 95% CI 1.29-2.87); this did not hold in female participants (HR 1.38, 95% CI 0.78-2.44). DiscussionIn a province-wide cohort, SDs at TBI were independently associated with incident dementia. Clinical trials testing sex-specific SD care after TBI for dementia prevention are timely. HighlightsTBI and sleep disorders are linked to each other, and to dementia.It is unclear if sleep disorders pose a sex-specific dementia risk in brain injury.In this study, presence of a sleep disorder increased dementia risk in both sexes.The risk differed by type of sleep disorder, which differed between the sexes.Sleep disorder awareness and care in persons with brain injury is vital for dementia prevention.

Topic models are a class of unsupervised machine learning models, which facilitate summarization, browsing and retrieval from large unstructured document collections. This study reviews several methods for assessing the quality of unsupervised topic models estimated using non-negative matrix factorization. Techniques for topic model validation have been developed across disparate fields. We synthesize this literature, discuss the advantages and disadvantages of different techniques for topic model validation, and illustrate their usefulness for guiding model selection on a large clinical text corpus. Using a retrospective cohort design, we curated a text corpus containing 382,666 clinical notes collected between 01/01/2017 through 12/31/2020 from primary care electronic medical records in Toronto Canada. Several topic model quality metrics have been proposed to assess different aspects of model fit. We explored the following metrics: reconstruction error, topic coherence, rank biased overlap, Kendall’s weighted tau, partition coefficient, partition entropy and the Xie-Beni statistic. Depending on context, cross-validation and/or bootstrap stability analysis were used to estimate these metrics on our corpus. Cross-validated reconstruction error favored large topic models (K ≥ 100 topics) on our corpus. Stability analysis using topic coherence and the Xie-Beni statistic also favored large models (K = 100 topics). Rank biased overlap and Kendall’s weighted tau favored small models (K = 5 topics). Few model evaluation metrics suggested mid-sized topic models (25 ≤ K ≤ 75) as being optimal. However, human judgement suggested that mid-sized topic models produced expressive low-dimensional summarizations of the corpus. Topic model quality indices are transparent quantitative tools for guiding model selection and evaluation. Our empirical illustration demonstrated that different topic model quality indices favor models of different complexity; and may not select models aligning with human judgment. This suggests that different metrics capture different aspects of model goodness of fit. A combination of topic model quality indices, coupled with human validation, may be useful in appraising unsupervised topic models.

Understanding the factors associated with elevated risks and adverse consequences of traumatic brain injury (TBI) is an integral part of developing preventive measures for TBI. Brain injury outcomes differ based on one’s sex (biological characteristics) and gender (social characteristics reflecting norms and relationships), however, whether it is sex or gender that drives differences in early (30-day) mortality and discharge location post-TBI event are unknown. In the absence of gender variable in existing data, we developed a method for “measuring gender” in 276,812 residents of Ontario, Canada who entered the emergency department and acute care hospitals with a TBI diagnostic code between April 1 st , 2002 and March 31 st , 2020. We analysed differences in diagnostic codes between the sexes to derive gender score that reflected social dimensions. Sex had a significant effect on early mortality after severe TBI with a rate ratio (95% confidence interval (CI)) of 1.54 (1.24-1.91). Gender had a more significant effect than sex on discharge location. A person expressing more female-like characteristics have lower odds of being discharged to rehabilitation versus home with odds ratio (95% CI) of 0.54 (0.32-0.88). The method we propose offers an opportunity to measure gender effect independently of sex on TBI outcomes.

Background/Objectives: Unsupervised topic models are often used to facilitate improved understanding of large unstructured clinical text datasets. In this study we investigated how ICD-9 diagnostic codes, collected alongside clinical text data, could be used to establish concurrent-, convergent- and discriminant-validity of learned topic models. Design/Setting: Retrospective open cohort design. Data were collected from primary care clinics located in Toronto, Canada between 01/01/2017 through 12/31/2020. Methods: We fit a non-negative matrix factorization topic model, with K = 50 latent topics/themes, to our input document term matrix (DTM). We estimated the magnitude of association between each Boolean-valued ICD-9 diagnostic code and each continuous latent topical vector. We identified ICD-9 diagnostic codes most strongly associated with each latent topical vector; and qualitatively interpreted how these codes could be used for external validation of the learned topic model. Results: The DTM consisted of 382,666 documents and 2210 words/tokens. We correlated concurrently assigned ICD-9 diagnostic codes with learned topical vectors, and observed semantic agreement for a subset of latent constructs (e.g. conditions of the breast, disorders of the female genital tract, respiratory disease, viral infection, eye/ear/nose/throat conditions, conditions of the urinary system, and dermatological conditions, etc.). Conclusions: When fitting topic models to clinical text corpora, researchers can leverage contemporaneously collected electronic medical record data to investigate the external validity of fitted latent variable models.

Background & Objective: Biomedical text data are increasingly available for research. Tokenization is an initial step in many biomedical text mining pipelines. Tokenization is the process of parsing an input biomedical sentence (represented as a digital character sequence) into a discrete set of word/token symbols, which convey focused semantic/syntactic meaning. The objective of this study is to explore variation in tokenizer outputs when applied across a series of challenging biomedical sentences. Method: Diaz [2015] introduce 24 challenging example biomedical sentences for comparing tokenizer performance. In this study, we descriptively explore variation in outputs of eight tokenizers applied to each example biomedical sentence. The tokenizers compared in this study are the NLTK white space tokenizer, the NLTK Penn Tree Bank tokenizer, Spacy and SciSpacy tokenizers, Stanza/Stanza-Craft tokenizers, the UDPipe tokenizer, and R-tokenizers. Results: For many examples, tokenizers performed similarly effectively; however, for certain examples, there were meaningful variation in returned outputs. The white space tokenizer often performed differently than other tokenizers. We observed performance similarities for tokenizers implementing rule-based systems (e.g. pattern matching and regular expressions) and tokenizers implementing neural architectures for token classification. Oftentimes, the challenging tokens resulting in the greatest variation in outputs, are those words which convey substantive and focused biomedical/clinical meaning (e.g. x-ray, IL-10, TCR/CD3, CD4+ CD8+, and (Ca2+)-regulated). Conclusion: When state-of-the-art, open-source tokenizers from Python and R were applied to a series of challenging biomedical example sentences, we observed subtle variation in the returned outputs.

Measurements of response inhibition components of reactive inhibition and proactive inhibition within the stop-signal paradigm have been of particular interest to researchers since the 1980s. While frequentist nonparametric and Bayesian parametric methods have been proposed to precisely estimate the entire distribution of reactive inhibition, quantified by stop signal reaction times (SSRT), there is no method yet in the stop signal task literature to precisely estimate the entire distribution of proactive inhibition. We identify the proactive inhibition as the difference of go reaction times for go trials following stop trials versus those following go trials and introduce an Asymmetric Laplace Gaussian (ALG) model to describe its distribution. The proposed method is based on two assumptions of independent trial type (go/stop) reaction times and Ex-Gaussian (ExG) models. Results indicated that the four parametric ALG model uniquely describes the proactive inhibition distribution and its key shape features, and its hazard function is monotonically increasing, as are its three parametric ExG components. In conclusion, the four parametric ALG model can be used for both response inhibition components and its parameters and descriptive and shape statistics can be used to classify both components in a spectrum of clinical conditions.

For centuries, the study of traumatic brain injury (TBI) has been centred on historical observation and analyses of personal, social, and environmental processes, which have been examined separately. Today, computation implementation and vast patient data repositories can enable a concurrent analysis of personal, social, and environmental processes, providing insight into changes in health status transitions over time. We applied computational and data visualization techniques to categorize decade-long health records of 235,003 patients with TBI in Canada, from preceding injury to the injury event itself. Our results highlighted that health status transition patterns in TBI emerged along with the projection of comorbidity where many disorders, social and environmental adversities preceding injury are reflected in external causes of injury and injury severity. The strongest associations between health status preceding TBI and health status at the injury event were between multiple body system pathology and advanced age-related brain pathology networks. The interwoven aspects of health status on a time continuum can influence post-injury trajectories and should be considered in TBI risk analysis to improve prevention, diagnosis, and care.

BACKGROUND:Health care organizations are collecting increasing volumes of clinical text data. Topic models are a class of unsupervised machine learning algorithms for discovering latent thematic patterns in these large unstructured document collections.OBJECTIVE:We aimed to comparatively evaluate several methods for estimating temporal topic models using clinical notes obtained from primary care electronic medical records from Ontario, Canada.METHODS:We used a retrospective closed cohort design. The study spanned from January 01, 2011, through December 31, 2015, discretized into 20 quarterly periods. Patients were included in the study if they generated at least 1 primary care clinical note in each of the 20 quarterly periods. These patients represented a unique cohort of individuals engaging in high-frequency use of the primary care system. The following temporal topic modeling algorithms were fitted to the clinical note corpus: nonnegative matrix factorization, latent Dirichlet allocation, the structural topic model, and the BERTopic model.RESULTS:Temporal topic models consistently identified latent topical patterns in the clinical note corpus. The learned topical bases identified meaningful activities conducted by the primary health care system. Latent topics displaying near-constant temporal dynamics were consistently estimated across models (eg, pain, hypertension, diabetes, sleep, mood, anxiety, and depression). Several topics displayed predictable seasonal patterns over the study period (eg, respiratory disease and influenza immunization programs).CONCLUSIONS:Nonnegative matrix factorization, latent Dirichlet allocation, structural topic model, and BERTopic are based on different underlying statistical frameworks (eg, linear algebra and optimization, Bayesian graphical models, and neural embeddings), require tuning unique hyperparameters (optimizers, priors, etc), and have distinct computational requirements (data structures, computational hardware, etc). Despite the heterogeneity in statistical methodology, the learned latent topical summarizations and their temporal evolution over the study period were consistently estimated. Temporal topic models represent an interesting class of models for characterizing and monitoring the primary health care system.