Validation of HORIBA Medical Pentra XL/XLR and Microsemi CRP Malaria flag performance derived from algorithmic data-mining techniques

► Click here to download the original poster (PDF)

Parag Dharap¹, Sébastien Raimbault², Sylvie Arnavielhe³, Gérard Dray⁴, Stefan Janaqi⁴, Michel Plantié⁴, Pierre Jean⁴, Vincent Derozier⁴, Shubham Rastogi⁵
¹Dr Dharap’s Diagnostic Centre, Mumbai, India, ²Dept. of Innovation, HORIBA Medical, Montpellier, France, ³Kyomed, Montpellier, France, ⁴LGI2P/EMA, Nimes, France, ⁵HORIBA Medical, New Delhi, India

1. Abstract

Introduction:

Blood cell counter screening for the presence of malaria has been investigated for >20 years. Efficiency reported has varied by methodology and study design. Several manufacturers have introduced malaria flagging on high range instruments with reports claiming sensitivity and specificity >80%. However the economic reality is malaria endemic areas typically utilize low and medium range instruments. We applied contemporary computer machine-learning techniques to develop flagging algorithms for detection of malaria using two low to middle range HORIBA Medical blood counters. The validation included Dengue fever patients, which have similar clinical symptomatology as malaria.

Method:

290 blood specimens were serially analyzed the following instruments and testing modes: Microsemi CRP (CBC 3 part diff + CRP modes) and Pentra XLR (5 part diff or fluorescent reticulocyte analysis modes). Specimen selection used three diagnostic categories confirmed by antigen testing: healthy, malaria positive, dengue fever positive. Malaria confirmed cases were further speciated into P. vivax (N=103) and P. falciparum (N=29) and microscopically scored the predominate development stage (schizoints, ring or amoeboid). Normal (N=70) and dengue cases (N=87) were considered to be malaria negative samples for determination of sensitivity, specificity, negative and positive predictive values, correct classification rate and positive and negative likelihood ratios. Data-mining techniques were used to identify those instrument parameters that showed significant differences between malaria from normal in a training set of 550 cases (232 malaria positive) and the flagging algorithms developed included weighting of parameters showing higher discriminating power. The number of parameters selected from those tested for the malaria suspect flag varied by instrument (Microsemi CRP, 59 of 886; PXLR DIFF, 412 of 25,994; PXLR Retic, 402 of 26,395) with the datamining and machine-learning techniques. Samples were analyzed with instrument software integrating the malaria suspect flag with results compared to the confirmatory testing.

Results:

2. Background

Malaria is one of the most common communicable diseases in the world, affecting the populations in all tropical regions. Malaria is a life threatening disease caused by parasites that are transmitted to people through the bites of infected female Anopheles mosquitoes. About 3.2 billion people – almost half of the world’s population are at risk of malaria. Malaria is the most common cause of morbidity in Africa. Malaria is a major public health problem with WHO estimates of 207 million cases of malaria occurred globally in 2012 and 6,270,000 deaths.

Several hematology blood analyzers have reported the ability to detect the presence of malaria infection using various parameters with varying efficiency. Abbott analyzers utilized light depolarization and nucleic acid fluorescent dyes (1-3), Beckman Coulter instruments utilized off-line algorithms developed from cell positional or size parameters (4-6) and Sysmex likewise has utilized cell size parameters in the white blood cell matrix and changes in the fluorescence of red cell or reticulocyte matrix for their malaria flag (7-12, copied as well by Mindray). In common to all these methods of malaria detection, either performed as an automated flag or by user interpretation of scattergram patterns, is that less than four parameters are utilized in the detection of blood changes associated with malaria infection.

HORIBA Medical chose to develop an automated malaria suspect flag using datamining techniques to examine all parameters generated by the Microsemi CRP, Pentra XL and Pentra XLR instruments to determine those most useful in distinguishing malaria from both healthy and non-malarial infections. Using these contemporary machine-learning techniques applied to over 500 patient specimens including not only malaria positive and healthy cases, but also a group of patients with dengue fever, which has a clinical presentation similar to malaria (same signs and symptoms). From these data files 412 of 25,994 parameters for the Pentra XL and Pentra XLR in CBC+Diff mode, 402 of 26,395 for the Pentra XLR in modified Retic mode, and 59 of 886 variables for the Microsemi CRP were identified as having discriminatory power for malaria. From these useful parameters a flagging algorithm was developed and utilized in this validation study.

3. Methods

290 patient samples, selected from diagnostic groups of normal, malaria positive or dengue fever positive, were included in the study with each sample run in duplicate. Rapid antigen diagnostic tests manufactured by SD – Bio Standard Diagnostics Pvt Ltd; Gurgaon, India were used to screen for Plasmodium falciparum (histidine rich protein II), Plasmodium vivax (Plasmodium lactate dehydrogenase), and dengue fever (NS1 Antigen &/ IgM Antibody).

All malaria cases were confirmed by microscopy and scored for the predominant life cycle forms (trophozoites, schizoints, gametocytes).

Blood samples were analyzed in duplicate on a HORIBA Medical Pentra XLR instrument (CBC 5 part differential counts and thiazole orange based reticulocyte analysis) using both the CBC DIFF mode and again with a modified (decreased) fluorescence gain Retic modes and on a Microsemi CRP instrument (CBC 3 part differential counts and whole blood C-reactive protein) in both CBC+Diff mode or CBC+Diff+CRP modes.

Results statistically analyzed for the efficiency, sensitivity, specificity, positive Youden J index, correctness of classification, positive predictive value and negative predictive value of the malaria suspect flag compared to the usual laboratory methods for malaria confirmation.

Figure 1: Summary of datamining process. All parameters measured and generated by the blood counter instrument, including counts, histogram information; and matrix plot positions are compared for each diagnostic category (normal; malaria positive and dengue fever). Differences between the diagnostic groups are normalized for both negative and positive differences. Variables with discriminating ability are identified and further weighted in the flagging algorithm developed to classify malaria as distinct from healthy individuals and dengue fever patients.

4. Results

Table 1: Performance of malaria suspect flag on Pentra XLR using CBC and DIFF mode. The number of true negative (TN), false negative (FN), true positive (TP) and false positive (FP) results are determined for flagging on duplicate analysis. Statistical analysis for negative predictive value (NPV), positive predictive value (PPV), positive likelihood ratio (PLR); negative likelihood ratio (NLR), Youden J index with correct classification rate with specificity and sensitivity are good for both A, malaria vs. all negative samples (normal and dengue fever) and B, malaria vs. normal samples.

Table 2: Performance of malaria suspect flag on Pentra XLR using CBC and DIFF mode. The statistical analysis for negative predictive value (NPV), positive predictive value (PPV), positive likelihood ratio (PLR), negative likelihood ratio (NLR), Youden J index; correct classification rate; specificity and sensitivity are equally good for both A, P. vivax vs. all negative samples (normal and dengue fever) and B, P with falciparum vs. normal samples.

Table 3: Performance of malaria suspect flag on Pentra XLR RETIC mode is equally good for both A, malaria vs. all negative samples (normal and dengue fever) and B, malaria vs. normal samples. However, the modified reticulocyte analysis mode did not improve performance over the CBC + DIFF mode.

Table 4: Performance of malaria suspect flag on Pentra XLR using CBC + DIFF mode (A) and Pentra XLR using RETIC mode (A) show equally good distinction between malaria vs. dengue fever samples.

Table 5: Performance of malaria suspect flag on Microsemi CRP using CBC + 3-part DIFF + CRP mode. The statistical performance are equally good for both A, malaria vs. all negative samples (normal and dengue fever) and B, malaria vs. normal samples.

Table 6: Performance summary of malaria suspect flag on Pentra XLR and Microsemi CRP for all modes tested and comparison among the various clinical groups (All = normal + dengue).

5. Conclusions

The malaria suspect flag developed with data-mining techniques on the Pentra XLR, both in CBC + Diff and RETIC modes, and Microsemi CRP in CBC + DIFF + CRP mode provide effective screening for malaria.
Reticulocyte analysis does not significantly improve the performance of malaria detection on Pentra XLR over routine CBC + Diff mode.
The Microsemi CRP malaria flag requires the addition of C-reactive protein measurement to achieve suitable performance of sensitivity and NPV.
The malaria flags also provide good distinction between malaria and dengue fever and equal performance for both P. vivax and P. falciparum.
Effective malaria suspect flags can be developed on low and middle range instruments and are more practical solutions for malaria endemic areas, which tend to be in economic developing countries.