Using multistage pooling mechanisms to optimise HIV viral load testing.

Date
2022-04
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH SUMMARY: HIV viral load (VL) monitoring is necessary to determine if antiretroviral (ART) treatment is successful. When the VL is virologically suppressed, patients are less likely to transmit HIV to others. The use of ART has resulted in a decreased risk of virologic failure (VF). Therefore, a lower prevalence of VF in patients receiving treatment and the high sensitivity (assays can quantify down to 20 cp/mL) of VL tests, ensure pooled testing for VF monitoring to be effective. However, such testing methods have never been implemented on a large scale, likely due to manual pooling of samples and testing each sample in a positive pool (referred to as pool deconvolution) being mundane and repetitive processes. The various steps in the pooling process make it more susceptible to human error. In this project, four models of four pooling methods were developed to reduce costs in testing. These methods are two multi-stage (Qualitative and Quantitative) and two matrix pooling (Matrix-or and Matrix-and) methods. The effects of VL dilution within a pool and error rates coupled to assays were included within testing. Data was simulated using gamma distributions fit to over 17 000 samples recorded by anonymized extraction of routine HIV-1 VL data from a South African public service diagnostic laboratory. These simulations provided a basis for optimisation by allowing an optimisation database to record the results (error rates as well as test reductions) of testing using each possible deconvolution for every prevalence of HIV-positive samples. The tests were conducted on more than 2 000 samples per deconvolution. A cost function was developed that calculates a single score based on false-negative and false-positive errors received during testing, running costs, and testing time. This score was used to find optimal testing procedures based on HIV prevalence in the data. The HIV prevalence within data is estimated using two types of estimation, with each containing three different methods, including machine learning prediction. Each optimised testing method and its cost function, which was weighted by evaluating each method using a training dataset, was then tested on unseen data. The results indicated that the Quantitative method was optimal for testing on the dataset, with prevalence larger than 24.3%. This method decreased costs by 18.0%, increased throughput by 17.7% and achieved a meaningful sensitivity of 96.0%. This method is effective for prevalence range up to 27.0%. It was observed that the Qualitative method achieved optimal results for higher prevalence than the other methods and when the prevalence is reduced the Matrix-and method would be optimum. A combined testing method was developed which increased efficiency to 19.0% and sensitivity increased to 98.5%. Each method was implemented on a Raspberry PI, to reduce human error, and is accessible using a graphical user interface (GUI). Guidelines for the adaptation of these methods to Covid-19 and other diseases were also developed.
AFRIKAANS SUMMARY: MIV virus lading (VL) moet gemonitor word om te bepaal of antiretrovirale (ART) behandeling suksesvol is. As die VL virologies onderdruk word, is dit minder waarskynlik dat pasiënte MIV aan ander sal oordra. Die gebruik van ART het tot ’n verminderde risiko vir virologiese mislukking, gelei. Die lae voorkoms van virologiese mislukking by pasiënte wat behandeling ontvang en die hoë sensitiwiteit (toetse kan tot 20 cp/mL kwantifiseer) van VL-toetse, verseker dat die toets van VL-poele (saamgevoegde bloedmonsters) vir VFmonitering effektief is. Hierdie toetsmetodes is egter nog nooit op groot skaal gëmplementeer nie, waarskynlik as gevolg van die alledaagse en herhalende proses van die samevoeging van die monsters met die hand en die verder toets van monsters (verwys na as dekonvolusie), as ’n poel positiewe getoets is. Die verskillende stappe in die poelproses maak dit meer vatbaar vir menslike foute. In hierdie projek word vier modelle van vier poel toetsmetodes opgestel en gebruik om die toetskostes te verlaag. Hierdie is twee meervoudige dekonvolusie (Kwalitatiewe en Kwantitatiewe) en twee matriks (Matriks-of en Matriks-en) toetsmetodes. Die invloed van VL-verdunning binne ’n poel en misklasifikasie, gekoppel aan toetse, is ingesluit in die modelle. Data word gesimuleer deur gamma-distribusies te gebruik wat gepas is op meer as 17 000 MIV-1 VL-monsters wat in ’n Suid-Afrikaanse staatsdiens diagnostiese laboratorium gemeet is. Hierdie simulasies bied ’n basis vir optimalisering deur ’n optimaliseringsdatabasis in staat te stel om die resultate (foutkoerse sowel as toetsverminderings) van toetsing aan te teken deur elke moontlike dekonvolusie vir elke voorkoms van MIV-positiewe monsters te gebruik. Die toetse is op meer as 2 000 monsters per dekonvolusie uitgevoer. ’n Kostefunksie is ontwerp wat ’n enkele telling gebaseer op vals-negatiewe en vals-positiewe misklassifikasie, bedryfskoste en tyd bereken wat gebruik word om ’n optimale dekonvolusie wat gebaseer is op die uitkoms van die data, te vind. Die voorkoms van MIV positiewe monsters in die data word beraam met behulp van twee tipes vergelykings, elkeen bestaande uit drie verskillende metodes, insluitend voorspelling deur masjienleer. Elke geoptimaliseerde toetsmetode en die kostefunksie daarvan, wat geweeg was deur die evaluering van elke metode met behulp van ’n opleidingsdatastel, was dan getoets op voorheen ongetoetse data. Die resultate het aangedui dat die Kwantitatiewe metode optimaal was vir toetsing op die datastel met voorkoms groter as 24.3%. Hierdie metode het die koste met 18.0% verlaag, die deurset van monsters verhoog met 17.7% en ’n betekenisvolle sensitiwiteit van 96.0% behaal. Hierdie metode is effektief vir ’n voorkoms van MIV van tot 27.0%. Daar is opgemerk dat die Kwalitatiewe metode optimale resultate behaal het vir ’n hoër voorkoms van MIV as die ander metodes, en as die MIV voorkoms verminder word, sou die Matriks-en metode optimaal wees. ’n Gekombineerde toetsmetode is ontwikkel wat die koste vermindering tot 19.0% verhoog het en die sensitiwiteit tot 98.5% verhoog het. Elke metode is op ’n Raspberry PI gëmplementeer om sodoende menslike foute te verminder en is toeganklik met ’n grafiese gebruikerskoppelvlak (GUI). Riglyne om hierdie metodes aan te pas by Covid-19 en ander siektes is ook ontwikkel.
Description
Thesis (MEng)--Stellenbosch University, 2022.
Keywords
HIV-testing, Antiretroviral agents, Pooling, Optimisation, Viral load -- Monitoring, UCTD
Citation