Suchergebnisse

IntroductionThe South African HIV Cancer Match (SAM) study is a probabilistic record linkage study involving creation of an HIV cohort from laboratory records from the National Health Laboratory Service (NHLS). This cohort was linked to the pathology based South African National Cancer Registry to establish cancer incidences among HIV positive population in South Africa. As the number of HIV records increases, there is need for more efficient ways of de-duplicating this big-data. In this work, we used clustering to perform big-data deduplication.
Objectives and ApproachOur objective was to use DBSCAN as clustering algorithm together with bi-gram word analyser to perform big-data deduplication in resource-limited settings. We used HIV related laboratory records from entire South Africa collated in the NHLS Corporate Data Warehouse for period 2004-2014. This involved data pre-processing, deterministic deduplication, ngrams generation, features generation using Term Frequency Inverse Document Frequency vectorizer, clustering using DBSCAN and assigning cluster labels for records that potentially belonged to the same person. We used records with national identification numbers to assess quality of deduplication by calculating precision, recall and f-measure.
ResultsWe had 51,563,127 HIV related laboratory records. Deterministic deduplication resulted in 20,387,819 patient record deduplicates. With DBSCAN clustering we further reduced this to 14,849,524 patient record clusters. In this final dataset, 3,355,544 (22.60%) patients had negative HIV test, 11,316,937 (76.21%) had evidence for HIV infection, and for 177,043 (1.19%) the HIV status could not be determined. The precision, recall and f-measure based on 1,865,445 records with national identification numbers were 0.96, 0.94 and 0.95, respectively.
Conclusion / ImplicationsOur study demonstrated that DBSCAN clustering is an effective way of deduplicating big datasets in resource-limited settings. This enabled refining of an HIV observational database by accurately linking test records that potentially belonged to the same person. The methodology creates opportunities for easy data profiling to inform public health decision making.

ABSTRACT
BackgroundSub-Saharan Africa is the region most heavily affected by the HIV/AIDS epidemic. HIV increases the risk of developing cancer but the ascertainment of cancers in patients attending antiretroviral therapy (ART) treatment programs might be incomplete. To estimate the under-ascertainment of cancer we compared incidence rates of AIDS-defining cancers in South African HIV cohorts with and without cancer case ascertainment through record linkage with the National Cancer Registry.
MethodsWe used the data of adult (≥16 years) HIV-positive persons receiving care between 2004 and 2011 at one of four ART programs in South Africa. These programs collaborate with the International Epidemiologic Databases to Evaluate AIDS Southern Africa (www.iedea-sa.org) and collected data for AIDS-defining cancers but not for other cancers. To improve cancer ascertainment we probabilistically linked patient records (using first name, surname, age, and gender) from two HIV cohorts with the cancer records of the South African National Cancer Registry. We calculated incidence rates per 100,000 person-years after starting ART for the AIDS-defining cancers, i.e. Kaposi sarcoma (KS), invasive cervical cancer (ICC) and non-Hodgkin lymphoma (NHL). We compared incidence rates before and after inclusion of record linkage identified cancer cases using the attributable fraction of cancers identified with 95% confidence intervals (CI).
ResultsA total of 49,207 adults starting ART in South Africa were included. 65% of patients were female, median age at starting ART was 35 years (interquartile range 30-41 years). We identified a total of 471 incident cancer cases. With record linkage the incidence increased from 81 to 292 for KS, from 1 to 119 for NHL and 12 to 497 for ICC per 100,000 person-years. The attributable fraction of cancers identified was 72% (95% CI 63-79%) for KS, 98% (95% CI 94-99%) for NHL and 98% (95% CI 95-99%) for ICC.
ConclusionAscertainment of cancer in HIV program data in African settings is incomplete. This case study has shown that probabilistic record linkage to cancer registries is both feasible and essential for cancer ascertainment in HIV cohorts in South Africa.

BackgroundAccording to the Global Strategy to Accelerate the Elimination of Cervical Cancer launched by the World Health Organizations (WHO), 70% of all women should undergo cervical cancer (CC) screening by 2030. While integrated HIV care and CC screening services exist, lack of unified data and patient management systems hinder reaching the set WHO goals. We used record linkage methods to assess the number of women receiving HIV care and visual inspection with acetic acid (VIA) based CC screening at Kanyama First Level Hospital (FLH) in Lusaka, Zambia.
MethodsWe retrieved HIV patient records from SmartCare, Zambia's National Electronic Health Records system, and the CC screening database for the years 2004 to 2018, and 2010 to 2018, respectively. We restricted this pilot study to women attending Kanyama FLH for HIV services, and its surrounding clinics for CC screening. We used the following linkage variables: first name, last name, facility, and year of birth. We used a machine learning based approach with Random Forest classifiers to identify records from the HIV and CC screening services belonging to the same person. We created synthetic patient record datasets to train and test the linkages. After the linkage, we anonymized the data and retrieved screening status from the CC screening database. We assessed the percentage of women in HIV care undergoing CC screening and the respective screening results. We calculated the time to CC screening with reference to the start of the HIV Antiretroviral Therapy treatment (ART).
ResultsWe retrieved a total of 32,970 records from SmartCare and 176,736 records from CC screening databases. With the linkage we determined that of all women that ever accessed HIV treatment services at Kanyama FLH between 2004 and 2018, 2,999 (9%) received CC screening. Of those that were screened, 30% received the CC service at Kanyama FLH, while the rest from 13 surrounding facilities. 61% of women accessed CC services before starting ART. Of the women screened for CC, 371 had a positive VIA result (12%), 2,395 (80%) a negative and 233 (8%) had a missing or indeterminate result. When we restricted the analysis to women, who were eligible for screening (aged 25 to 59 years), the screening coverage was 9.6%.
ConclusionOur study demonstrates the significance of record linkage efforts in determining CC screening coverage among women living with HIV. According to these findings, CC screening coverage among patients receiving ART services between 2004 and 2018 was at 9%, showing the depth of effort required to meet WHO targets at the time. The linkage output was primarily influenced by the scarcity of available linkage variables and data missingness, resulting in potentially overestimating or underestimating the CC coverage. This is typical of data systems in low- and middle-income countries; however, this study demonstrates that record linkage is useful in bridging these widespread data gaps.

AbstractINTRODUCTIONTo eliminate cervical cancer (CC), access to and quality of prevention and care services must be monitored, particularly for women living with HIV (WLHIV). We assessed implementation practices in HIV clinics across sub‐Saharan Africa (SSA) to identify gaps in the care cascade and used aggregated patient data to populate cascades for WLHIV attending HIV clinics.METHODSOur facility‐based survey was administered between November 2020 and July 2021 in 30 HIV clinics across SSA that participate in the International epidemiology Databases to Evaluate AIDS (IeDEA) consortium. We performed a qualitative site‐level assessment of CC prevention and care services and analysed data from routine care of WLHIV in SSA.RESULTSHuman papillomavirus (HPV) vaccination was offered in 33% of sites. Referral for CC diagnosis (42%) and treatment (70%) was common, but not free at about 50% of sites. Most sites had electronic health information systems (90%), but data to inform indicators to monitor global targets for CC elimination in WLHIV were not routinely collected in these sites. Data were collected routinely in only 36% of sites that offered HPV vaccination, 33% of sites that offered cervical screening and 20% of sites that offered pre‐cancer and CC treatment.CONCLUSIONSThough CC prevention and care services have long been available in some HIV clinics across SSA, patient and programme monitoring need to be improved. Countries should consider leveraging their existing health information systems and use monitoring tools provided by the World Health Organization to improve CC prevention programmes and access, and to track their progress towards the goal of eliminating CC.