Prediction of cardiovascular risk factors from retinal fundus photographs: Validation of a deep learning algorithm in a prospective non-interventional study in Kenya.

AIM: Hypertension and diabetes mellitus (DM) are major causes of morbidity and mortality, with growing burdens in low-income countries where they are underdiagnosed and undertreated. Advances in machine learning may provide opportunities to enhance diagnostics in settings with limited medical infrastructure. MATERIALS AND METHODS: A non-interventional study was conducted to develop and validate a machine learning algorithm to estimate cardiovascular clinical and laboratory parameters. At two sites in Kenya, digital retinal fundus photographs were collected alongside blood pressure (BP), laboratory measures and medical history. The performance of machine learning models, originally trained using data from the UK Biobank, were evaluated for their ability to estimate BP, glycated haemoglobin, estimated glomerular filtration rate and diagnoses from fundus images. RESULTS: In total, 301 participants were enrolled. Compared with the UK Biobank population used for algorithm development, participants from Kenya were younger and would probably report Black/African ethnicity, with a higher body mass index and prevalence of DM and hypertension. The mean absolute error was comparable or slightly greater for systolic BP, diastolic BP, glycated haemoglobin and estimated glomerular filtration rate. The model trained to identify DM had an area under the receiver operating curve of 0.762 (0.818 in the UK Biobank) and the hypertension model had an area under the receiver operating curve of 0.765 (0.738 in the UK Biobank). CONCLUSIONS: In a Kenyan population, machine learning models estimated cardiovascular parameters with comparable or slightly lower accuracy than in the population where they were trained, suggesting model recalibration may be appropriate. This study represents an incremental step toward leveraging machine learning to make early cardiovascular screening more accessible, particularly in resource-limited settings.

Deep Learning for Epidemiologists: An Introduction to Neural Networks.

Deep learning methods are increasingly being applied to problems in medicine and health care. However, few epidemiologists have received formal training in these methods. To bridge this gap, this article introduces the fundamentals of deep learning from an epidemiologic perspective. Specifically, this article reviews core concepts in machine learning (e.g., overfitting, regularization, and hyperparameters); explains several fundamental deep learning architectures (convolutional neural networks, recurrent neural networks); and summarizes training, evaluation, and deployment of models. Conceptual understanding of supervised learning algorithms is the focus of the article; instructions on the training of deep learning models and applications of deep learning to causal learning are out of this article's scope. We aim to provide an accessible first step towards enabling the reader to read and assess research on the medical applications of deep learning and to familiarize readers with deep learning terminology and concepts to facilitate communication with computer scientists and machine learning engineers.

Predicting overdose among individuals prescribed opioids using routinely collected healthcare utilization data.

INTRODUCTION: With increasing rates of opioid overdoses in the US, a surveillance tool to identify high-risk patients may help facilitate early intervention. OBJECTIVE: To develop an algorithm to predict overdose using routinely-collected healthcare databases. METHODS: Within a US commercial claims database (2011-2015), patients with ≥1 opioid prescription were identified. Patients were randomly allocated into the training (50%), validation (25%), or test set (25%). For each month of follow-up, pooled logistic regression was used to predict the odds of incident overdose in the next month based on patient history from the preceding 3-6 months (time-updated), using elastic net for variable selection. As secondary analyses, we explored whether using simpler models (few predictors, baseline only) or different analytic methods (random forest, traditional regression) influenced performance. RESULTS: We identified 5,293,880 individuals prescribed opioids; 2,682 patients (0.05%) had an overdose during follow-up (mean: 17.1 months). On average, patients who overdosed were younger and had more diagnoses and prescriptions. The elastic net model achieved good performance (c-statistic 0.887, 95% CI 0.872-0.902; sensitivity 80.2, specificity 80.1, PPV 0.21, NPV 99.9 at optimal cutpoint). It outperformed simpler models based on few predictors (c-statistic 0.825, 95% CI 0.808-0.843) and baseline predictors only (c-statistic 0.806, 95% CI 0.787-0.26). Different analytic techniques did not substantially influence performance. In the final algorithm based on elastic net, the strongest predictors were age 18-25 years (OR: 2.21), prior suicide attempt (OR: 3.68), opioid dependence (OR: 3.14). CONCLUSIONS: We demonstrate that sophisticated algorithms using healthcare databases can be predictive of overdose, creating opportunities for active monitoring and early intervention.

Predicting Inpatient Medication Orders From Electronic Health Record Data.

In a general inpatient population, we predicted patient-specific medication orders based on structured information in the electronic health record (EHR). Data on over three million medication orders from an academic medical center were used to train two machine-learning models: A deep learning sequence model and a logistic regression model. Both were compared with a baseline that ranked the most frequently ordered medications based on a patient's discharge hospital service and amount of time since admission. Models were trained to predict from 990 possible medications at the time of order entry. Fifty-five percent of medications ordered by physicians were ranked in the sequence model's top-10 predictions (logistic model: 49%) and 75% ranked in the top-25 (logistic model: 69%). Ninety-three percent of the sequence model's top-10 prediction sets contained at least one medication that physicians ordered within the next day. These findings demonstrate that medication orders can be predicted from information present in the EHR.

Using prescription claims to detect aberrant behaviors with opioids: comparison and validation of 5 algorithms.

OBJECTIVE: Compare and validate 5 algorithms to detect aberrant behavior with opioids: Opioid Misuse Score, Controlled Substance-Patterns of Utilization Requiring Evaluation (CS-PURE), Overutilization Monitoring System, Katz, and Cepeda algorithms. STUDY DESIGN AND SETTING: We identified new prescription opioid users from 2 insurance databases: Medicaid (2000-2006) and Clinformatics Data Mart (CDM; 2004-2013). Patients were followed 1 year, and aberrant opioid behavior was defined according to each algorithm, using Cohen's kappa to assess agreement. Risk differences were calculated comparing risk of opioid-related adverse events for identified aberrant and nonaberrant users. RESULTS: About 3.8 million Medicaid and 4.3 million CDM patients initiated prescription opioid use. Algorithms flagged potential aberrant behavior in 0.02% to 12.8% of initiators in Medicaid and 0.01% to 7.9% of initiators in CDM. Cohen's kappa values were poor to moderate (0.00 to 0.50 in Medicaid; 0.00 to 0.30 in CDM). Algorithms varied substantially in their ability to predict opioid-related adverse events; the Overutilization Monitoring System had the highest risk differences between aberrant and nonaberrant users (14.0% in Medicaid; 13.4% in CDM), and the Katz algorithm had the lowest (0.96% in Medicaid; 0.47% in CDM). CONCLUSIONS: In 2 large databases, algorithms applied to prescription data had varying accuracy in identifying increased risk of adverse opioid-related events.

Trends in Neonatal Prophylaxis and Predictors of Combination Antiretroviral Prophylaxis in US Infants from 1990 to 2015.

Postnatal antiretroviral (ARV) prophylaxis for infants born to women with HIV is a critical component of perinatal HIV transmission prevention. However, variability in prophylaxis regimens remains and consistency with guidelines has not been evaluated in the United States. We evaluated trends over time in prophylaxis regimens among 6386 HIV-exposed uninfected (HEU) infants using pooled data spanning two decades from three US-based cohorts: the Women and Infants Transmission Study (WITS, 1990-2007), Pediatric AIDS Clinical Trials Group (PACTG) 219C (1993-2007), and the PHACS Surveillance Monitoring of ART Toxicities (SMARTT) study (2007-2015). We also identified maternal and infant risk factors for use of combination prophylaxis regimens (≥2 ARVs) and examined consistency with US perinatal guidelines. We found that receipt of combination prophylaxis between 1996 and 2015 ranged from 2% to 15%, with a consistent median duration of 6 weeks. Infants whose mothers had lower CD4 T-cell counts, higher viral load (VL), no antepartum ARVs, age <20 years at delivery, and Cesarean delivery had significantly higher rates of combination prophylaxis, while infants born 2006-2010 (vs. 2011-2015), who were Hispanic or with lower maternal education levels, had significantly lower rates. Predictors for combination prophylaxis varied over time, with the strongest associations of maternal VL in later birth cohorts. While use of combination prophylaxis increased over time, only 50% of high-risk infants received such regimens in 2011-2015. In conclusion, HEU infants at higher risk of HIV acquisition are more likely to receive combination neonatal prophylaxis, consistent with US guidelines. However, substantial variability remains, and infants at higher risk often fail to receive combination prophylaxis.

Effect of Gestational Age at Tenofovir-Emtricitabine-Efavirenz Initiation on Adverse Birth Outcomes in Botswana.

Among human immunodeficiency virus-positive women in Botswana on the recommended first-line antiretroviral therapy regimen, tenofovir-emtricitabine-efavirenz, initiated within the first or early second trimester, we found no increased risk of stillbirth, neonatal death, preterm/very preterm delivery, or the infant being born small or very small for gestational age. Treatment with tenofovir-emtricitabine-efavirenz <1 year before conception increased the risk of preterm delivery slightly over late-second-trimester treatment initiation (adjusted risk ratio, 1.33 [95% confidence interval, 1.04-1.70]).

Birth Outcomes for Pregnant Women with HIV Using Tenofovir-Emtricitabine.

BACKGROUND: In a previous trial of antiretroviral therapy (ART) involving pregnant women with human immunodeficiency virus (HIV) infection, those randomly assigned to receive tenofovir, emtricitabine, and ritonavir-boosted lopinavir (TDF-FTC-LPV/r) had infants at greater risk for very premature birth and death within 14 days after delivery than those assigned to receive zidovudine, lamivudine, and ritonavir-boosted lopinavir (ZDV-3TC-LPV/r). METHODS: Using data from two U.S.-based cohort studies, we compared the risk of adverse birth outcomes among infants with in utero exposure to ZDV-3TC-LPV/r, TDF-FTC-LPV/r, or TDF-FTC with ritonavir-boosted atazanavir (ATV/r). We evaluated the risk of preterm birth (<37 completed weeks of gestation), very preterm birth (<34 completed weeks), low birth weight (<2500 g), and very low birth weight (<1500 g). Risk ratios with 95% confidence intervals were estimated with the use of modified Poisson models to adjust for confounding. RESULTS: There were 4646 birth outcomes. Few infants or fetuses were exposed to TDF-FTC-LPV/r (128 [2.8%]) as the initial ART regimen during gestation, in contrast with TDF-FTC-ATV/r (539 [11.6%]) and ZDV-3TC-LPV/r (954 [20.5%]). As compared with women receiving ZDV-3TC-LPV/r, women receiving TDF-FTC-LPV/r had a similar risk of preterm birth (risk ratio, 0.90; 95% confidence interval [CI], 0.60 to 1.33) and low birth weight (risk ratio, 1.13; 95% CI, 0.78 to 1.64). As compared to women receiving TDF-FTC-ATV/r, women receiving TDF-FTC-LPV/r had a similar or slightly higher risk of preterm birth (risk ratio, 1.14; 95% CI, 0.75 to 1.72) and low birth weight (risk ratio, 1.45; 95% CI, 0.96 to 2.17). There were no significant differences between regimens in the risk of very preterm birth or very low birth weight. CONCLUSIONS: The risk of adverse birth outcomes was not higher with TDF-FTC-LPV/r than with ZDV-3TC-LPV/r or TDF-FTC-ATV/r among HIV-infected women and their infants in the United States, although power was limited for some comparisons. (Funded by the National Institutes of Health and others.).

Patterns of opioid initiation at first visits for pain in United States primary care settings.

PURPOSE: The primary objective of this study was to characterize variation in patterns of opioid prescribing within primary care settings at first visits for pain, and to describe variation by condition, geography, and patient characteristics. METHODS: 2014 healthcare utilization data from Optum's Clinformatics™ DataMart were used to evaluate individuals 18 years or older with an initial presentation to primary care for 1 of 10 common pain conditions. The main outcomes assessed were (1) the proportion of first visits for pain associated with an opioid prescription fill and (2) the proportion of opioid prescriptions with >7 days' supply. RESULTS: We identified 205 560 individuals who met inclusion criteria; 9.1% of all visits were associated with an opioid fill, ranging from 4.1% (headache) to 28.2% (dental pain). Approximately half (46%) of all opioid prescriptions supplied more than 7 days, and 10% of prescriptions supplied ≥30 days. We observed a 4-fold variation in rates of opioid initiation by state, with highest rates of prescribing in Alabama (16.6%) and lowest rates in New York (3.7%). CONCLUSIONS: In 2014, nearly half of all patients filling opioid prescriptions received more than 7 days' of opioids in an initial prescription. Policies limiting initial supplies have the potential to substantially impact opioid prescribing in the primary care setting.

Risk of neonatal drug withdrawal after intrauterine co-exposure to opioids and psychotropic medications: cohort study.

Objectives To assess the impact of in utero co-exposure to psychotropic medications and opioids on the incidence and severity of neonatal drug withdrawal.Design Observational cohort study.Setting Nationwide sample of pregnancies in publicly insured women in the US, nested in the Medicaid Analytic eXtract (2000-10).Participants 201 275 pregnant women with public insurance who were exposed to opioids around the time of delivery and their liveborn infants.Interventions In utero exposure to psychotropic medications, in particular antidepressants, atypical antipsychotics, benzodiazepines, gabapentin, and non-benzodiazepine hypnotics (Z drugs), with prescriptions filled within the same time window as prescriptions for opioids.Main outcome measure Diagnosis of neonatal drug withdrawal in infants exposed in utero to opioids and psychotropic medications compared with opioids alone.Results The absolute risk for neonatal drug withdrawal ranged from 1.0% in infants exposed in utero to prescription opioids alone to 11.4% for those exposed to opioids co-prescribed with gabapentin. Among neonates exposed in utero to prescription opioids, the relative risk adjusted for propensity score was 1.34 (95% confidence interval 1.22 to 1.47) with concomitant exposure to antidepressants, 1.49 (1.35 to 1.63) with benzodiazepines, 1.61 (1.26 to 2.06) with gabapentin, 1.20 (0.95 to 1.51) with antipsychotics, and 1.01 (0.88 to 1.15) with Z drugs. In utero exposure to two or more psychotropic medications along with opioids was associated with a twofold increased risk of withdrawal (2.05, 1.77 to 2.37). The severity of the withdrawal seemed increased in neonates exposed to both opioids and psychotropic medications compared with opioids alone.Conclusions During pregnancy, the use of psychotropic medications in addition to prescription opioids is common, despite a lack of safety data. The current findings suggest that these drugs could further increase the risk and severity of neonatal drug withdrawal.

Zidovudine use in pregnancy and congenital malformations.

OBJECTIVE: There is inconsistent evidence that zidovudine use during pregnancy increases overall, cardiac, and male genital malformations. DESIGN: We conducted a systematic review and meta-analysis of zidovudine use and malformations and, using Bayesian methods, combined it with data from a cohort study of mother-infant pairs in the nationwide Medicaid Analytic eXtract (MAX). METHODS: Using MAX data (2000-2010), we identified pregnant women with HIV treated with antiretroviral therapy (ART). Women with at least one zidovudine dispensing during the first trimester were compared to women receiving ART without zidovudine in the first trimester. Malformation outcomes were defined using diagnosis/procedure codes. To adjust for confounding, we performed 1 : 1 propensity score matching. Bayesian methods require specification of a prior, which we developed in the meta-analysis. Logistic regression models combined MAX data with the prior, estimating odds ratios (ORs) and 95% credible intervals. RESULTS: Fourteen articles contributed information on overall malformations, seven on cardiac malformations, and five on male genital malformations. In MAX, matching led to a sample of 735 women each in the zidovudine and comparator groups. When comparing first trimester zidovudine use to other ART, the Bayesian procedure yielded OR estimates slightly above the null for overall [OR = 1.11, 95% credible interval (0.80-1.55)] and cardiac [OR = 1.30 (0.63-2.71)] malformations. There were no zidovudine-exposed cases of male genital malformations in MAX, but the meta-analysis yielded elevated OR estimates [OR = 2.57 (1.26-5.24)]. CONCLUSION: For most malformations, first-trimester zidovudine was not associated with increased risk. The potential increase in male genital malformations was small in absolute terms, and should be evaluated further.

Dramatic decline in substance use by HIV-infected pregnant women in the United States from 1990 to 2012.

OBJECTIVE: We aimed to describe temporal changes in substance use among HIV-infected pregnant women in the United States from 1990 to 2012. DESIGN: Data came from two prospective cohort studies (Women and Infants Transmission Study and Surveillance Monitoring for Antiretroviral Therapy Toxicities Study). METHODS: Women were classified as using a substance during pregnancy if they self-reported use or had a positive biological sample. To account for correlation between repeated pregnancies by the same woman, generalized estimating equation models were used to test for temporal trends and evaluate predictors of substance use. RESULTS: Over the 23-year period, substance use among the 5451 HIV-infected pregnant women sharply declined; 82% of women reported substance use during pregnancy in 1990, compared with 23% in 2012. Use of each substance decreased significantly (P < 0.001 for each substance) in an approximately linear fashion, until reaching a plateau in 2006. Multivariable models showed substance use was inversely associated with receiving antiretroviral therapy. Among the subset of 824 women with multiple pregnancies under observation, women who used a substance in their previous pregnancy were at elevated risk of substance use during their next pregnancy (risk ratio, 5.71; 95% confidence interval, 4.63-7.05). CONCLUSION: A substantial decrease in substance use during pregnancy was observed between 1990 and 2012 in two large US cohorts of HIV-infected women. Substance use prevalence in these cohorts became similar to that of pregnant women in the general US population by the mid-2000s, suggesting that the observed decrease may be due to an epidemiological transition of the HIV epidemic among women in the United States.

Understanding treatment refusal among adults presenting for HIV-testing in Soweto, South Africa: a qualitative study.

HIV treatment initiatives have focused on increasing access to antiretroviral therapy (ART). There is growing evidence, however, that treatment availability alone is insufficient to stop the epidemic. In South Africa, only one third of individuals living with HIV are actually on treatment. Treatment refusal has been identified as a phenomenon among people who are asymptomatic, however, factors driving refusal remain poorly understood. We interviewed 50 purposively sampled participants who presented for voluntary counseling and testing in Soweto to elicit a broad range of detailed perspectives on ART refusal. We then integrated our core findings into an explanatory framework. Participants described feeling "too healthy" to start treatment, despite often having a diagnosis of AIDS. This subjective view of wellness was framed within the context of treatment being reserved for the sick. Taking ART could also lead to unintended disclosure and social isolation. These data provide a novel explanatory model of treatment refusal, recognizing perceived risks and social costs incurred when disclosing one's status through treatment initiation. Our findings suggest that improving engagement in care for people living with HIV in South Africa will require optimizing social integration and connectivity for those who test positive.

Whoonga and the abuse and diversion of antiretrovirals in Soweto, South Africa.

Media reports have described recreational use of HIV antiretroviral medication in South Africa, but little has been written about this phenomenon in the scientific literature. We present original, qualitative data from eight semi-structured interviews that characterize recreational antiretroviral use in Soweto, South Africa. Participants reported that antiretrovirals, likely efavirenz, are crushed, mixed with illicit drugs (in a mixture known as whoonga), and smoked. They described medications being stolen from patients and expressed concern that antiretroviral abuse jeopardized the safety of both patients and users. Further studies are needed to understand the prevalence, patterns, and consequences of antiretroviral abuse and diversion.

Individual-based simulation models of HIV transmission: reporting quality and recommendations.

BACKGROUND: Individual-based modeling is a growing technique in the HIV transmission and prevention literature, but insufficient attention has been paid to formally evaluate the quality of reporting in this field. We present reporting recommendations for individual-based models for HIV treatment and prevention, assess the quality of reporting in the existing literature, and comment on the contribution of this model type to HIV policy and prediction. METHODS: We developed reporting recommendations for individual-based HIV transmission mathematical models, and through a systematic search, used them to evaluate the reporting in the existing literature. We identified papers that employed individual-based simulation models and were published in English prior to December 31, 2012. Articles were included if the models they employed simulated and tracked individuals, simulated HIV transmission between individuals in a particular population, and considered a particular treatment or prevention intervention. The papers were assessed with the reporting recommendations. FINDINGS: Of 214 full text articles examined, 32 were included in the evaluation, representing 20 independent individual-based HIV treatment and prevention mathematical models. Manuscripts universally reported the objectives, context, and modeling conclusions in the context of the modeling assumptions and the model's predictive capabilities, but the reporting of individual-based modeling methods, parameterization and calibration was variable. Six papers discussed the time step used and one discussed efforts to maintain internal validity in coding. CONCLUSION: Individual-based models represent detailed HIV transmission processes with the potential to contribute to inference and policy making for many different regions and populations. The rigor in reporting of assumptions, methods, and calibration of individual-based models focused on HIV transmission and prevention varies greatly. Higher standards for reporting of statistically rigorous calibration and model assumption testing need to be implemented to increase confidence in existing and future modeling results.

Gender power control, sexual experiences, safer sex practices, and potential HIV risk behaviors among young Asian-American women.

We examined the prevalence of three domains of sexual behaviors among young Asian-American women: sexual experiences, safer sex practices, and potential HIV risk behaviors. We also investigated the impact of gender power control on these domains. Among sexually experienced women, 51% reported using condoms during their most recent sex act, 63% reported inconsistent condom use, and 18% reported ever having forced sex. Multiple logistic regression analyses revealed that women's perceived lower relationship power control was not associated with vaginal sex or safer sex practices, but it was powerfully associated with forced sex and all three potential HIV risk behaviors. This study demonstrates that control within young Asian-American women's intimate relationships exerts different associations depending on the type of sexual behavior. The application of the Theory of Gender and Power should be employed with prudence when designing HIV interventions for this population.