Machine learning with routine electronic medical record data to identify people at high risk of disengagement from HIV care in Tanzania
Ontology highlight
ABSTRACT:
SUBMITTER: Fahey C
PROVIDER: S-EPMC10021592 | biostudies-literature | 2022 Jan
REPOSITORIES: biostudies-literature
Similar Datasets
Project description:ObjectivesThis study aimed to develop a machine learning (ML) model to predict disengagement from HIV care, high viral load or death among people living with HIV (PLHIV) with the goal of enabling proactive support interventions in Tanzania. The algorithm addressed common challenges when applying ML to electronic medical record (EMR) data: (1) imbalanced outcome distribution; (2) heterogeneity across multisite EMR data and (3) evolving virological suppression thresholds.DesignObservational study using a national EMR database.SettingConducted in two regions in Tanzania, using data from the National HIV Care database.ParticipantsThe study included over 6 million HIV care visit records from 295 961 PLHIV in two regions in Tanzania's National HIV Care database from January 2015 to May 2023.ResultsOur ML model effectively identified PLHIV at increased risk of adverse outcomes. Key predictors included past disengagement from care, antiretroviral therapy (ART) status (which tracks a patient's engagement with ART across visits), age and time on ART. The downsampling approach we implemented effectively managed imbalanced data to reduce prediction bias. Site-specific algorithms performed better compared with a universal approach, highlighting the importance of tailoring ML models to local contexts. A sensitivity analysis confirmed the model's robustness to changes in viral load suppression thresholds.ConclusionsML models leveraging large-scale databases of patient data offer significant potential to identify PLHIV for interventions to enhance engagement in HIV care in resource-limited settings. Tailoring algorithms to local contexts and flexibility towards evolving clinical guidelines are essential for maximising their impact.
| S-EPMC11423721 | biostudies-literature
Project description:BackgroundIdentification of people with HIV from electronic health record (EHR) data is an essential first step in the study of important HIV outcomes, such as risk assessment. This task has been historically performed via manual chart review, but the increased availability of large clinical data sets has led to the emergence of phenotyping algorithms to automate this process. Existing algorithms for identifying people with HIV rely on a combination of International Classification of Disease codes and laboratory tests or closely mimic clinical testing guidelines for HIV diagnosis. However, we found that existing algorithms in the literature missed a significant proportion of people with HIV in our data.ObjectiveThe aim of this study is to develop and evaluate HIV-Phen, an updated criteria-based HIV phenotyping algorithm.MethodsWe developed an algorithm using HIV-specific laboratory tests and medications and compared it with previously published algorithms in national and local data sets to identify cohorts of people with HIV. Cohort demographics were compared with those reported in the national and local surveillance data. Chart reviews were performed on a subsample of patients from the local database to calculate the sensitivity, specificity, positive predictive value, negative predictive value, and accuracy of the algorithm.ResultsOur new algorithm identified substantially more people with HIV in both national (up to an 85.75% increase) and local (up to an 83.20% increase) EHR databases than the previously published algorithms. The demographic characteristics of people with HIV identified using our algorithm were similar to those reported in national and local HIV surveillance data. Our algorithm demonstrated improved sensitivity over existing algorithms (98% vs 56%-92%) while maintaining a similar overall accuracy (96% vs 80%-96%).ConclusionsWe developed and evaluated an updated criteria-based phenotyping algorithm for identifying people with HIV in EHR data that demonstrates improved sensitivity over existing algorithms.
| S-EPMC8727048 | biostudies-literature
Project description:ObjectivesAn understanding of the predictors of comorbidity among people living with HIV (PLWH) is critical for effective HIV care management. In this study, we identified predictors of comorbidity burden among PLWH based on machine learning models with electronic health record (EHR) data.MethodsThe study population are individuals with a HIV diagnosis between January 2005 and December 2016 in South Carolina (SC). The change of comorbidity burden, represented by the Charlson Comorbidity Index (CCI) score, was measured by the score difference between pre- and post-HIV diagnosis, and dichotomized into a binary outcome variable. Thirty-five risk predictors from multiple domains were used to predict the increase in comorbidity burden based on the logistic least absolute shrinkage and selection operator (Lasso) regression analysis using 80% data for model development and 20% data for validation.ResultsOf 8253 PLWH, the mean value of the CCI score difference was 0.8 ± 1.9 (range from 0 to 21) with 2328 (28.2%) patients showing an increase in CCI score after HIV diagnosis. Top predictors for an increase in CCI score using the LASSO model included older age at HIV diagnosis, positive family history of chronic conditions, tobacco use, longer duration with retention in care, having PEBA insurance, having low recent CD4+ cell count and duration of viral suppression.ConclusionThe application of machine learning methods to EHR data could identify important predictors of increased comorbidity burden among PLWH with high accuracy. Results may enhance the understanding of comorbidities and provide the evidence based data for integrated HIV and comorbidity care management of PLWH.
| S-EPMC8058944 | biostudies-literature
Project description:BackgroundElectronic medical record (EMR) chronic disease measurement can help direct primary care prevention and treatment strategies and plan health services resource management. Incomplete data and poor consistency of coded disease values within EMR problem lists are widespread issues that limit primary and secondary uses of these data. These issues were shared by the McMaster University Sentinel and Information Collaboration (MUSIC), a primary care practice-based research network (PBRN) located in Hamilton, Ontario, Canada.ObjectiveWe sought to develop and evaluate the effectiveness of new EMR interface tools aimed at improving the quantity and the consistency of disease codes recorded within the disease registry across the MUSIC PBRN.MethodsWe used a single-arm prospective trial design with preintervention and postintervention data analysis to assess the effect of the intervention on disease recording volume and quality. The MUSIC network holds data on over 75,080 patients, 37,212 currently rostered. There were 4 MUSIC network clinician champions involved in gap analysis of the disease coding process and in the iterative design of new interface tools. We leveraged terminology standards and factored EMR workflow and usability into a new interface solution that aimed to optimize code selection volume and quality while minimizing physician time burden. The intervention was integrated as part of usual clinical workflow during routine billing activities.ResultsAfter implementation of the new interface (June 25, 2017), we assessed the disease registry codes at 3 and 6 months (intervention period) to compare their volume and quality to preintervention levels (baseline period). A total of 17,496 International Classification of Diseases, 9th Revision (ICD9) code values were recorded in the disease registry during the 11.5-year (2006 to mid-2017) baseline period. A large gain in disease recording occurred in the intervention period (8516/17,496, 48.67% over baseline), resulting in a total of 26,774 codes. The coding rate increased by a factor of 11.2, averaging 1419 codes per month over the baseline average rate of 127 codes per month. The proportion of preferred ICD9 codes increased by 17.03% in the intervention period (11,007/17,496, 62.91% vs 7417/9278, 79.94%; χ21=819.4; P<.001). A total of 45.03% (4178/9278) of disease codes were entered by way of the new screen prompt tools, with significant increases between quarters (Jul-Sep: 2507/6140, 40.83% vs Oct-Dec: 1671/3148, 53.08%; χ21=126.2; P<.001).ConclusionsThe introduction of clinician co-designed, workflow-embedded disease coding tools is a very effective solution to the issues of poor disease coding and quality in EMRs. The substantial effectiveness in a routine care environment demonstrates usability, and the intervention detail described here should be generalizable to any setting. Significant improvements in problem list coding within primary care EMRs can be realized with minimal disruption to routine clinical workflow.
| S-EPMC7418012 | biostudies-literature
Project description:BackgroundThe limitations of existing HIV risk prediction tools are a barrier to implementation of pre-exposure prophylaxis (PrEP). We developed and validated an HIV prediction model to identify potential PrEP candidates in a large health-care system.MethodsOur study population was HIV-uninfected adult members of Kaiser Permanente Northern California, a large integrated health-care system, who were not yet using PrEP and had at least 2 years of previous health plan enrolment with at least one outpatient visit from Jan 1, 2007, to Dec 31, 2017. Using 81 electronic health record (EHR) variables, we applied least absolute shrinkage and selection operator (LASSO) regression to predict incident HIV diagnosis within 3 years on a subset of patients who entered the cohort in 2007-14 (development dataset), assessing ten-fold cross-validated area under the receiver operating characteristic curve (AUC) and 95% CIs. We compared the full model to simpler models including only men who have sex with men (MSM) status and sexually transmitted infection (STI) positivity, testing, and treatment. Models were validated prospectively with data from an independent set of patients who entered the cohort in 2015-17. We computed predicted probabilities of incident HIV diagnosis within 3 years (risk scores), categorised as low risk (<0·05%), moderate risk (0·05% to <0·20%), high risk (0·20% to <1·0%), and very high risk (≥1·0%), for all patients in the validation dataset.FindingsOf 3 750 664 patients in 2007-17 (3 143 963 in the development dataset and 606 701 in the validation dataset), there were 784 incident HIV cases within 3 years of baseline. The LASSO procedure retained 44 predictors in the full model, with an AUC of 0·86 (95% CI 0·85-0·87) for incident HIV cases in 2007-14. Model performance remained high in the validation dataset (AUC 0·84, 0·80-0·89). The full model outperformed simpler models including only MSM status and STI positivity. For the full model, flagging 13 463 (2·2%) patients with high or very high HIV risk scores in the validation dataset identified 32 (38·6%) of the 83 incident HIV cases, including 32 (46·4%) of 69 male cases and none of the 14 female cases. The full model had equivalent sensitivity by race whereas simpler models identified fewer black than white HIV cases.InterpretationPrediction models using EHR data can identify patients at high risk of HIV acquisition who could benefit from PrEP. Future studies should optimise EHR-based HIV risk prediction tools and evaluate their effect on prescription of PrEP.FundingKaiser Permanente Community Benefit Research Program and the US National Institutes of Health.
| S-EPMC7152802 | biostudies-literature
Project description:BackgroundRoutine Data Quality Assessments (RDQAs) were developed to measure and improve facility-level electronic medical record (EMR) data quality. We assessed if RDQAs were associated with improvements in data quality in KenyaEMR, an HIV care and treatment EMR used at 341 facilities in Kenya.MethodsRDQAs assess data quality by comparing information recorded in paper records to KenyaEMR. RDQAs are conducted during a one-day site visit, where approximately 100 records are randomly selected and 24 data elements are reviewed to assess data completeness and concordance. Results are immediately provided to facility staff and action plans are developed for data quality improvement. For facilities that had received more than one RDQA (baseline and follow-up), we used generalized estimating equation models to determine if data completeness or concordance improved from the baseline to the follow-up RDQAs.Results27 facilities received two RDQAs and were included in the analysis, with 2369 and 2355 records reviewed from baseline and follow-up RDQAs, respectively. The frequency of missing data in KenyaEMR declined from the baseline (31% missing) to the follow-up (13% missing) RDQAs. After adjusting for facility characteristics, records from follow-up RDQAs had 0.43-times the risk (95% CI: 0.32-0.58) of having at least one missing value among nine required data elements compared to records from baseline RDQAs. Using a scale with one point awarded for each of 20 data elements with concordant values in paper records and KenyaEMR, we found that data concordance improved from baseline (11.9/20) to follow-up (13.6/20) RDQAs, with the mean concordance score increasing by 1.79 (95% CI: 0.25-3.33).ConclusionsThis manuscript demonstrates that RDQAs can be implemented on a large scale and used to identify EMR data quality problems. RDQAs were associated with meaningful improvements in data quality and could be adapted for implementation in other settings.
| S-EPMC5905951 | biostudies-literature
Project description:Background Electronic medical records (EMRs) allow identification of disease-specific patient populations, but varying electronic cohort definitions could result in different populations. We compared the characteristics of an electronic medical record-derived atrial fibrillation (AF) patient population using 5 different electronic cohort definitions. Methods and Results Adult patients with at least 1 AF billing code from January 1, 2010, to December 31, 2017, were included. Based on different electronic cohort definitions, we trained 5 different logistic regression models using a labeled training data set (n=786). Each model yielded a predicted probability; patients were classified as having AF if the probability was higher than a specified cut point. Test characteristics were calculated for each model. These models were then applied to the full cohort and resulting characteristics were compared. In the training set, the comprehensive model (including demographics, billing codes, and natural language processing results) performed best, with an area under the curve of 0.89, sensitivity of 0.90, and specificity of 0.87. Among a candidate population (n=22 000), the proportion of patients identified as having AF varied from 61% in the model using diagnosis or procedure International Classification of Diseases (ICD) billing codes to 83% in the model using natural language processing of clinical notes. Among identified AF patients, the proportion of patients with a CHA2DS2-VASc score ≥2 varied from 69% to 85%; oral anticoagulant treatment rates varied from 50% to 66% depending on the model. Conclusions Different electronic cohort definitions result in substantially different AF study samples. This difference threatens the quality and reproducibility of electronic medical record-based research and quality initiatives.
| S-EPMC7335556 | biostudies-literature
Project description:BackgroundStudy of pulmonary arterial hypertension (PAH) in claims-based (CB) cohorts may facilitate understanding of disease epidemiology, however previous CB algorithms to identify PAH have had limited test characteristics. We hypothesized that machine learning algorithms (MLA) could accurately identify PAH in an CB cohort.MethodsICD-9/10 codes, CPT codes or PAH medications were used to screen an electronic medical record (EMR) for possible PAH. A subset (Development Cohort) was manually reviewed and adjudicated as PAH or "not PAH" and used to train and test MLAs. A second subset (Refinement Cohort) was manually reviewed and combined with the Development Cohort to make The Final Cohort, again divided into training and testing sets, with MLA characteristics defined on test set. The MLA was validated using an independent EMR cohort.Results194 PAH and 786 "not PAH" in the Development Cohort trained and tested the initial MLA. In the Final Cohort test set, the final MLA sensitivity was 0.88, specificity was 0.93, positive predictive value was 0.89, and negative predictive value was 0.92. Persistence and strength of PAH medication use and CPT code for right heart catheterization were principal MLA features. Applying the MLA to the EMR cohort using a split cohort internal validation approach, we found 265 additional non-confirmed cases of suspected PAH that exhibited typical PAH demographics, comorbidities, hemodynamics.ConclusionsWe developed and validated a MLA using only CB features that identified PAH in the EMR with strong test characteristics. When deployed across an entire EMR, the MLA identified cases with known features of PAH.
| S-EPMC9145474 | biostudies-literature
Project description:BackgroundImmunization reminders in electronic health records (EHR) provide clinical decision support (CDS) that can reduce missed immunization opportunities. Little is known about using CDS rules from a regional immunization information system (IIS) to power local EHR immunization reminders.ObjectiveThis study aimed to assess the impact of EHR reminders using regional IIS CDS-provided rules on receipt of immunizations in a low-income, urban population for both routine immunizations and those recommended for patients with chronic medical conditions (CMCs).MethodsWe built an EHR-based immunization reminder using the open-source resource used by the New York City IIS in which we overlaid logic regarding immunizations needed for CMCs. Using a randomized cluster-cross-over pragmatic clinical trial in four academic-affiliated clinics, we compared captured immunization opportunities during patient visits when the reminder was "on" versus "off" for the primary immunization series, school-age boosters, and adolescents. We also assessed coverage of CMC-specific immunizations. Up-to-date immunization was measured by end of quarter. Rates were compared using chi square tests.ResultsOverall, 15,343 unique patients were seen for 26,647 visits. The alert significantly impacted captured opportunities to complete the primary series in both well-child and acute care visits (57.6% on vs. 54.3% off, p = 0.001, and 15.3% on vs. 10.1% off, p = 0.02, respectively), among most age groups, and several immunization types. Captured opportunities for CMC-specific immunizations remained low regardless of alert status. The alert did not have an effect on up-to-date immunization overall (89.1 vs. 88.3%).ConclusionCDS in this population improved captured immunization opportunities. Baseline high rates may have blunted an up-to-date population effect. Converting Centers for Disease Control and Prevention (CDC) rules to generate sufficiently sensitive and specific alerts for CMC-specific immunizations proved challenging, and the alert did not have an impact on CMC-specific immunizations, potentially highlighting need for more work in this area.
| S-EPMC8674096 | biostudies-literature
Project description:IntroductionFrailty is a complex condition that affects many aspects of patients' wellbeing and health outcomes.ObjectivesWe used available Electronic Medical Record (EMR) and administrative data to determine definitions of frailty. We also examined whether there were differences in demographics or health conditions among those identified as frail in either the EMR or administrative data.MethodsEMR and administrative data were linked in British Columbia (BC) and Manitoba (MB) to identify those aged 65 years and older who were frail. The EMR data were obtained from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN) and the administrative data (e.g. billing, hospitalizations) was obtained from Population Data BC and the Manitoba Population Research Data Repository. Sociodemographic characteristics, risk factors, prescribed medications, use and costs of healthcare are described for those identified as frail.ResultsSociodemographic and utilization differences were found among those identified as frail from the EMR compared to those in the administrative data. Among those who were >65 years, who had a record in both EMR and administrative data, 5%-8% (n=191 of 3,553, BC; n=2,396 of 29,382, MB) were identified as frail. There was a higher likelihood of being frail with increasing age and being a woman. In BC and MB, those identified as frail in both data sources have approximately twice the number of contacts with primary care (n=20 vs. n=10) and more days in hospital (n=7.2 vs. n=1.9 in BC; n=9.8 vs. n=2.8 in MB) compared to those who are not frail; 27% (BC) and 14% (MB) of those identified as frail in 2014 died in 2015.ConclusionsIdentifying frailty using EMR data is particularly challenging because many functional deficits are not routinely recorded in structured data fields. Our results suggest frailty can be captured along a continuum using both EMR and administrative data.
| S-EPMC7893852 | biostudies-literature