Diagnosing malaria with machine learning and cell morphology

Diagnosis of malaria is undertaken with a range of tools, the gold standard being microscopy of stained blood samples. A range of alternatives exist with each their drawbacks and advantages, including Polymerase Chain Reaction (PCR) and antibody based tests such as self-contained Rapid Diagnostic Tests. Increasingly sophisticated and sensitive methods of detection are being developed in laboratories around the world, such as droplet microfluidics platforms and even mobile app-based technologies. In this article, we’ll discuss a new method reported by Han Sang Park and colleagues at Duke University.

The new approach, described as “machine learning algorithms with quantitative phase images”, is a combination of statistical analysis and a particular method of microscopy. Starting with the latter, Quantitate Phase Imaging (QPI) is a microscopy technique allowing the determination of structural and morphological characteristics of cellular samples. In this particular application, Park et al. use QPI to determine a range of morphological characteristics of normal and infected red blood cells (RBC’s), such as minor and major axis length, elongation and skewness. In total, 23 separate characteristics were examined, all of which were significantly different between infected and non-infected RBC’s, but none of which were able to discriminate between infected and non-infected populations with a high degree of accuracy. For this reason, the authors combined all 23 parameters in three machine learning approaches: Linear Discriminant Classification, Logistic Regression and k-nearest neighbour classification. These approaches differed slightly in their ability to distinguish infected from non-infected RBC’s across three life stages (early trophozoite, late trophozoite and schizont) of the parasite, but e.g. LCD exhibited sensitivity/specificity of 93.5%-99.4%/99.9%-100% across the three stages.

Although this method eliminates the need for human interpretation, a major drawback of traditional stained microscopy, it does require careful sample preparation to avoid contamination of the analytical sample with white blood cells, which the algorithm reportedly mistakes for infected RBCs. The authors also note that current overall computation time is not feasible for clinical use, and that separate training sets would be needed for patients with morphological pathologies of RBC’s.

The technology is an interesting example of new approaches to diagnosis. Most existing methods look for direct evidence of the pathogen, such as enzymatic activity, parasitic molecular structures detected with antibodies, or the presence of parasite DNA using either amplification or staining and visualisation with microscopy. In contrast the present approach utilises changes to the host environment, i.e. changes in the shape and form of RBC’s, to determine whether infection is present. On account of the machine learning approach, there is no human interpretation of sample readings, giving the potential for very high accuracy. On the other hand, current sample preparation using a centrifuge step and manual separation requires technical skill, which is time consuming and involves risk of contamination and biased readings.

For further information, see: Park et al, “Automated Detection of P. falciparum Using Machine Learning Algorithms with Quantitative Phase Images of Unstained Cells“, PLoS One (2016)

Be the first to comment

Leave a Reply

Your email address will not be published.