Pathology informatics selected abstracts

Create PDF

Editors: Liron Pantanowitz, MD, director of anatomical pathology, Department of Pathology, University of Michigan, Ann Arbor, and David McClintock, MD, associate chief medical information officer for pathology, Department of Pathology, University of Michigan.

Applying machine learning to lab workflows to reduce false-positive HIV results

January 2022—With modern HIV testing incorporating p24 antigen detection, fourth- and fifth-generation HIV screening tests have halved the time between acquiring and detecting the infection. However, even with an enhanced detection window, these tests are not ideal because of complicated results reporting and a high number of false positives in low-prevalence populations. The latter can cause stress for patients who may have to undergo repeat testing. The authors hypothesized that machine-learning classification systems could be used with fifth-generation HIV screening tests (HIV5G) to more accurately discriminate between true-positive and false-positive results. This could reduce both turnaround times for true-positive cases and the number of false positives reported to local health agencies, in part by routing those samples for more appropriate reflex testing. The authors conducted a study in which they applied MATLAB R2019b to implement a machine-learning method designed to distinguish between HIV5G assay true positives and false positives. They used a cohort of 60,587 HIV5G screening tests, from 2016 to 2018, along with their molecular and clinical correlates to create the model. The authors found that using support vector machines for feature extraction and then transforming results with principle component analysis produced the best classifier for the HIV5G laboratory workflow. Application of this classifier resulted in 94 percent correct classification of false-positive screening tests and 92 percent correct classification of true-positive screens, with an overall improvement in assay screening from a 73.5 percent baseline without machine-learning assistance to 92.6 percent with machine-learning assistance. The authors concluded that using machine-learning–based methods to augment HIV5G screening may greatly improve laboratory and diagnostics workflows by permitting increased numbers of probable true positives to be reported immediately, thereby reducing the spread of infection and allowing for more efficient patient follow-up and treatment. The methods also may shunt likely false positives into separate repeat/confirmatory testing pathways that use the same specimen, thereby reducing patient stress.

Elkhadrawi M, Stevens BA, Wheeler BJ, et al. Machine learning classification of false-positive human immunodeficiency virus screening results. J Pathol Inform. 2021;12. www.jpathinformatics.org/text.asp?2021/12/1/46/330787

Correspondence: Dr. Bradley J. Wheeler at wheelerse3@upmc.edu

Challenges of using artificial intelligence in anatomic pathology

FDA clearance of the first artificial intelligence-based algorithm for clinical diagnostic use with prostate biopsies has created optimism among members of the pathology community regarding the use of AI in anatomic pathology. However, its use in the latter capacity is unlikely to happen any time soon as several considerable challenges must first be overcome in the areas of AI development, deployment, and regulation. Challenges related to developing useful AI-based algorithms include solving unmet needs without disrupting clinical workflow, achieving suitable data curation and annotation, training algorithms without introducing bias, and acquiring expensive computer hardware. Deploying an AI-based system in a clinical pathology work environment also poses challenges. Pathology laboratories initially will have to transition to fully digital imaging platforms, revamp legacy information technology infrastructures, and modify workflows. Appropriate reimbursement or cost offsets, or both, are also necessary to support a sensible business use case for adopting AI. The success of AI in routine clinical practice too will require that pathologists willingly participate in the use and oversight of the technology. Regulations being developed will promote safe and effective use.