Background
Accurate diagnosis of malaria infection in both humans and mosquitoes is essential for understanding transmission patterns, estimating epidemiological burden and informing appropriate management of cases. WHO’s Global Technical Strategy for Malaria 2016–2030 (GTS) recommends that surveillance strategies should be integrated as core interventions to provide better estimates of disease burden, improve resource allocation and accelerate progress towards elimination [
1]. The malERA Refresh Consultative Panel on Tools for Malaria Elimination reiterated in 2017 that new diagnostic tools are as important as new drugs, vaccines and vector control tools [
2].
Given health system challenges facing many low-income, malaria-endemic countries, there is particular interest in non-immunological point-of-care (POC) techniques that could be readily scaled up with minimum effort [
2]. Equally vital is the need for better quantification and identification of asymptomatic pathogen carriers in situations of low transmission and sub-microscopic parasitaemia [
2,
3]. Low-cost mass screening tools are also vital for large cross-sectional investigations such as national malaria indicator surveys, and for monitoring progress of interventions such as insecticide-treated bed nets [
4‐
6], indoor residual spraying [
7], larviciding [
8], and anti-malarial medicines [
9].
Malaria diagnosis from human samples currently relies on light microscopy with thin and thick blood smears [
10], polymerase chain reaction (PCR) assays [
11‐
13] and a variety of antigen-detecting rapid diagnostic tests [
14]. As countries approach elimination, there is greater need for accurate detection of malaria parasites in both symptomatic and asymptomatic individuals [
15]. Microscopy is still commonly used in diagnosis to support case management [
10]. With highly trained and experienced microscopists, it can offer reliable results and enable identification and quantification of sexual and asexual stages of infections by different malaria parasite species (
Plasmodium falciparum,
Plasmodium vivax,
Plasmodium malariae, and
Plasmodium ovale, being the most prevalent). However, microscopy-based diagnosis is labour-intensive and often requires more than one highly trained person for confirmation of conflicting results [
10]. Additionally, this technique has low sensitivity in many laboratories and misses most low-level parasitaemia cases [
16].
PCR, on the other hand, is highly specific and sensitive for malaria parasite detection, but is not widely used for primary diagnosis in most malaria-endemic places because it is expensive, requires highly skilled personnel, and is therefore impractical to implement in rural and remote facilities [
17]. Furthermore, until the impact of treating sub-microscopic parasitaemia on transmission is fully understood, WHO has recommended that PCR should not be part of routine malaria control or elimination programmes [
18]. It is however capable of identifying low-level parasitaemia that are otherwise undetectable by other methods, and distinguishing between individual parasite species. Recent developments in PCR applications have signaled the potential of non-invasive malaria diagnostic options, such as those relying on DNA detection in saliva, urine, sweat, and even faecal excreta [
19‐
21]. Besides, real time PCR assays enable quantitative assessments and comparison of infection loads [
15], but are expensive for most laboratories. A related technology is the loop-mediated isothermal amplification (LAMP), which is also increasingly used for diagnosis of multiple malaria species, and can be conducted at room temperatures without PCR instruments [
22].
At present, malaria rapid diagnostic tests (RDTs) are the best option for addressing the technical limitations of both light microscopy and DNA-based diagnosis. These tests target persistent specific antigens from malaria parasites [
14,
23,
24] and are quick and deployable at large scale. They also do not require highly skilled labour, electricity or sophisticated storage needs. Unfortunately, RDTs can be unreliable in low transmission settings, resulting in significant false negative and false positive results [
25]. Both microscopy and RDTs are recommended only when the number of malaria parasites exceed 100/µL, and are therefore not applicable for measuring low-level parasitaemia or identifying asymptomatic cases [
16]. In one example, mass RDT screening followed by treatment did not result in reduction of malaria incidence in the pre-elimination settings of Zanzibar, most likely due to poor reliability of the RDTs [
26]. However, a new rapid lateral flow technique that detects specific proteins of the infectious
Plasmodium stages, i.e., gametocytes, in saliva is showing encouraging results in detecting sub-microscopic parasites in children and adults [
27].
Recent studies have shown that non-molecular techniques such as near-infrared spectroscopy (12,500 cm
−1 to 400 cm
−1 frequencies) and mid-infrared spectroscopy (4000 cm
−1 to 400 cm
−1 frequencies), combined with advanced data analysis, could provide cheaper, quicker, reagent-free, and potentially simpler options for surveys of mosquitoes and mosquito-borne infections. Examples include detection of endosymbionts such as
Wolbachia bacteria, and pathogens such as
Plasmodium and Zika virus in mosquitoes [
28‐
31]. Such approaches have also been used for estimating ages of disease-transmitting mosquitoes [
32‐
38], distinguishing between vector species [
32,
38,
39] and assessing their blood-feeding histories [
40], all of which directly influence malaria transmission.
Khoshmanesh et al. [
41] used mid-infrared (MIR) spectroscopy combined with partial least-squares (PLS) regression to detect early ring stages of laboratory-cultured
P. falciparum with detection limits less than 1 parasite/µL. To improve this approach, Roy et al. [
42] used
Plasmodium cultures to spike whole blood obtained from six uninfected volunteers, then aliquoted these mixtures multiple times to obtain 132 specimens containing different quantities of
P. falciparum parasites, glucose and urea. Based on PLS regressions analysis of MIR spectra from these aliquots, they correctly identified 98% of specimens with parasitaemia densities above 0.5% (~ 25,000 parasites/µL). Sensitivity was however only 70%, possibly because the model included only a small number of negative samples. Although limited to laboratory cultures and small number of samples with low genetic variability, these studies were the first to demonstrate direct potential of MIR for malaria parasite detection.
This current study has extended the approach used by Khoshmanesh et al. [
41] and Roy et al. [
42], to provide the first demonstration of MIR spectroscopy coupled with supervised machine learning (MIR-ML) to diagnose malaria in human dried blood spots (DBS) obtained from field surveys of naturally infected individuals in a malaria-endemic community in Tanzania.
Discussion
This study is an initial demonstration, using field-collected blood specimens, that MIR spectroscopy coupled with logistic regression analysis could potentially be harnessed for detection of infectious parasitic diseases, in this case malaria. Malaria cases were identified by both MIR spectroscopy and PCR, then the results were compared. It considered 296 individual samples, to demonstrate potential application of MIR spectroscopy for malaria detection. The main finding was that spectral signatures collected from human DBS can be relied upon to identify malaria-infected and non-infected specimens.
Validity of this approach was verified by PCR tests, and corroborates the earlier evidence by Khoshmanesh et al. [
41] and Roy et al. [
42], who first demonstrated applications of MIR spectroscopy to detect
Plasmodium. Other related approaches include the use of surface-enhanced Raman spectroscopy (SERS), requiring silver nanoparticles to be added to lysed blood, and magneto-optic scanning to detect haemozoin, a waste product of
Plasmodium infection of red cells pigment, haemoglobin [
57,
58]. Both of these approaches have been tested in vitro, although Newman et al. also evaluated the magneto-optic systems in a small pre-clinical trial with 13 participants [
59].
An important advancement in the work presented here is that it has demonstrated the first direct application of MIR spectroscopy on field-collected DBS specimens on filter papers. The technique therefore requires no additional reagents or pre-processing of samples. The samples analysed came from multiple age groups of both male and females in villages with varying malaria prevalence rates, thereby providing considerable variability between individual infections. When the final optimized LR model was applied to new blood samples that had not been exposed to the classifier, they correctly identified 92% of the malaria-free individuals and 93% of malaria-infected individuals compared to PCR.
This study adds to the growing evidence showing the potential role of infrared spectroscopy and chemometrics in surveillance of mosquito-borne diseases. In a recent study conducted using near-infrared rather than MIR spectra, it was possible to detect infectious
P. falciparum sporozoites in laboratory-reared and laboratory-infected
Anopheles mosquitoes with up to 90% accuracy [
31]. Earlier studies had also demonstrated detection of various pathogens such as
Wolbachia,
Plasmodium and Zika virus in different mosquito species [
28‐
31]. Many of these earlier approaches relied on spectroscopy at near-infrared (NIR) frequencies (12,500 cm
−1 to 400 cm
−1) where the absorption intensity is due to overtone and combination bands, which are relatively weak. Spectroscopy at MIR frequencies (4000 cm
−1 to 400 cm
−1) captures the fundamental vibrational modes of biological samples, which are stronger and more information rich [
38], thereby offering stronger signals than NIR, which reflects secondary modes. Advances in data analysis and machine learning techniques now make the large spectral datasets amenable to processing.
The distinctive features of the MIR spectra may be related to the biochemical changes in red blood cells following malaria infection [
41,
42]. These differences may also result from
Plasmodium-specific proteins present in infected human blood [
60], or simply due to pathological manifestations of malaria such as anaemia, iron deficiency or other inflammatory responses. Based on previous analyses of body fluids [
55,
56,
61,
62], and earlier work by Roy et al. [
42], Kozicki et al. [
56], Khoshmanesh et al. [
41], the spectral bands putatively responsible for the differences between
Plasmodium-infected and
Plasmodium-free specimens were examined (Table
1). Further analysis of the dominant spectral wavelengths showed that they were mostly in the fingerprint region associated with amino acids, carbohydrates, lipids, and proteins (Table
1 and Fig.
6). Nonetheless, more detailed analysis beyond this current work is still required to examine how variations in these specific bands influence the diagnostic capacity of this technology.
Bands in the C–H stretching regions are usually attributed to lipids synthesized during development of
P. falciparum and can change the make-up of infected erythrocytes and become detectable in the spectra. On the other hand, the region between 1250 and 800 cm
−1 is sensitive to nucleic acid vibrations associated with proliferation of
Plasmodium-specific ribosomes during parasite development and cell invasion [
42]. Several peaks were observed in this region characteristic of sugars and peptides, consistent with parasite DNA (Table
1 and Fig.
6). In previous assessments, it was also shown that reduced absorbance in the carbohydrate regions at 1144 cm
−1, 1101 cm
−1 and 1085 cm
−1 are likely associated with lower glucose content in infected red blood cells, since these parasites metabolize glucose faster than normal cells [
41,
42]. Perhaps the most obvious are the different haem vibration regions (Table
1), which show higher haemoglobin levels in non-infected than infected samples because
Plasmodium catabolizes the complex haemoglobin protein into constituent lipids and bilirubin. By tracking these key characteristics, it is possible to predict the infection status of the dried whole blood specimen.
This study should be considered only as an initial evaluation of the approach, and more studies are required before the technology is field-deployable or effective. Filter papers containing human DBS were scanned and the resulting spectra used to train algorithms to predict outcomes of nested PCR [
40]. Comparison was made with results of tests considering only
P. falciparum and also with results of tests considering any species of
Plasmodium. In both cases, the technique achieved high accuracy in distinguishing between infected and uninfected specimen. While this finding indicates that the technique could be applicable for diagnosis of different malaria parasites, it also highlights the need to improve the approach so as to distinguish between different parasite species.
The field survey from which the DBS specimens were collected also had RDT data. Relative to PCR as reference, the RDT results achieved slightly higher sensitivity (97.6 vs 92.8%) but lower specificity (84.4 vs 91.7%) than the MIR spectroscopy approach (Table
2). In previous studies, the same RDTs achieved sensitivity of 99.7% for
P. falciparum, and 95.5% for non
P. falciparum and a specificity of 99.5% for both [
46,
63].
Significant improvements and field validation studies are still necessary before this technique can be deployed for actual screening and diagnosis. One avenue of improvement is the availability of more field data to validate this approach as a complementary tool for screening malaria, and potentially other blood-borne infections. Also, this current study did not include any quantification of infection intensities, which will be necessary to examine validity in areas of different transmission intensities, as well as potential role of this technology in malaria elimination settings. Such future studies should also include: (a) greater analysis on biological basis of the observed signals and how this may be influenced by the natural history of malaria infections in humans, or its manifestations, such as anaemia; (b) determination of whether the method can distinguish between the different parasite stages, such as the asexual stages versus the sexual stages (i.e., gametocytes); (c) detailed examination and characterization of cases incorrectly identified by this new approach; and, (d) assessment of parasite detection thresholds, and factors that may influence it if MIR spectroscopy and machine learning are used.
It takes 30 s to scan a single blood spot under MIR spectrometer, and the approach is considerably lower cost than PCR platforms. The MIR equipment presently costs ~ US$29,000 as an initial outlay, but there are no extended costs for reagents except for occasional replacement of the desiccants. It takes less than a minute to clean the crystal and anvil, position DBS on the crystal and collect the MIR spectra, thus any experienced staff member can scan more than 250 specimens per day. Comparatively, average PCR systems cost between US$3000 and 10,000, and require reagents repeatedly. The cost of processing a DBS sample can be US$2–4 per unit and regularly takes up to 2 days to get back full results in batches typically not exceeding 100 specimens. Assuming a modest analysis of just 5000 samples per year, it would take ~ 2 years to recover full costs of switching from PCR to MIR-ML-based systems and break even. The onward costs of servicing are also low as the MIR equipment is robust. Compared to PCR, use of MIR-ML could therefore potentially be developed into a cost-effective, quick and scalable approach. Besides, the same spectra, once collected, can be analysed for multiple characteristics, potentially making this approach a one-stop system for assessing multiple disease indicators with different specimen types, such as DBS papers, blood slides, mosquitoes, and fluids. It is however not expected that the technology, in its current form, can replace current best practices for malaria surveillance or diagnostics, without further field validation.
One limitation of this current study is that the number of samples used was low, totalling only 296 (123 Plasmodium positive and 173 Plasmodium negative). It is expected that the quality of the predictions will improve as more data are available to train the models, especially if there is variation in localities from where the specimen originates, parasite densities, geographical locations, and demographics of infected individuals. A related limitation was that it was not possible to balance the Plasmodium-positive and Plasmodium-negative specimens by other variables such as anaemia, gender, age, period of storage, or parasitaemia prevalence in the different villages, all of which may influence malaria risk and the predictive values of test methods. Future studies should therefore test whether these factors can significantly influence outcomes of the prediction models.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.