Machine Learning Classification of False-Positive Human Immunodeficiency Virus Screening Results

Area of Science:

• Clinical diagnostics
• Machine learning applications
• Public health

Background:

• Fourth- and fifth-generation HIV (HIV5G) screening tests, incorporating p24 antigen detection, have reduced the diagnostic window.
• However, HIV5G tests in low prevalence populations generate numerous false-positive results, raising questions about orthogonal testing's impact on outcomes.
• The clinical utility of current HIV screening protocols requires optimization.

Purpose of the Study:

• To develop and validate a machine learning classifier for predicting true and false positivity in HIV5G screening tests.
• To assess the potential of this classifier to improve diagnostic workflow and public health outcomes.

Main Methods:

• A cohort of 60,587 HIV5G screening tests with molecular and clinical data from 2016-2018 was analyzed.
• Machine learning, specifically support vector machines and principal component analysis, was employed to build a predictive classifier.

Main Results:

• The machine learning classifier achieved 94% accuracy in identifying false-positive screens.
• It demonstrated 92% accuracy in classifying true-positive screens.
• The model effectively distinguished between true and false positives.

Conclusions:

• Implementing the machine learning classifier can optimize HIV5G screening workflows by enabling immediate reporting of likely true positives.
• This facilitates prompt initiation of follow-up testing, treatment, and reduces spread of HIV infection.
• Likely false positives can be efficiently managed with orthogonal testing, minimizing patient distress and healthcare visits, thereby enhancing patient care and public health.