Screening of malaria infections in human blood samples with varying parasite densities and anaemic conditions using AI-Powered mid-infrared spectroscopy

Abstract

Background: Effective testing for malaria, including the detection of infections at very low densities, is vital for the successful elimination of the disease. Unfortunately, existing methods are either inexpensive but poorly sensitive or sensitive but costly. Recent studies have shown that mid-infrared spectroscopy coupled with machine learning (MIRs-ML) has potential for rapidly detecting malaria infections but requires further evaluation on diverse samples representative of natural infections in endemic areas. The aim of this study was, therefore, to demonstrate a simple AI-powered, reagent-free, and user-friendly approach that uses mid-infrared spectra from dried blood spots to accurately detect malaria infections across varying parasite densities and anaemic conditions.

Methods: Plasmodium falciparum strains NF54 and FCR3 were cultured and mixed with blood from 70 malaria-free individuals to create various malaria parasitaemia and anaemic conditions. Blood dilutions produced three haematocrit ratios (50%, 25%, 12.5%) and five parasitaemia levels (6%, 0.1%, 0.002%, 0.00003%, 0%). Dried blood spots were prepared on Whatman^™ filter papers and scanned using attenuated total reflection-Fourier Transform Infrared (ATR-FTIR) for machine-learning analysis. Three classifiers were trained on an 80%/20% split of 4655 spectra: (I) high contrast (6% parasitaemia vs. negative), (II) low contrast (0.00003% vs. negative) and (III) all concentrations (all positive levels vs. negative). The classifiers were validated with unseen datasets to detect malaria at various parasitaemia levels and anaemic conditions. Additionally, these classifiers were tested on samples from a population survey in malaria-endemic villages of southeastern Tanzania.

Results: The AI classifiers attained over 90% accuracy in detecting malaria infections as low as one parasite per microlitre of blood, a sensitivity unattainable by conventional RDTs and microscopy. These laboratory-developed classifiers seamlessly transitioned to field applicability, achieving over 80% accuracy in predicting natural P. falciparum infections in blood samples collected during the field survey. Crucially, the performance remained unaffected by various levels of anaemia, a common complication in malaria patients.

Conclusion: These findings suggest that the AI-driven mid-infrared spectroscopy approach holds promise as a simplified, sensitive and cost-effective method for malaria screening, consistently performing well despite variations in parasite densities and anaemic conditions. The technique simply involves scanning dried blood spots with a desktop mid-infrared scanner and analysing the spectra using pre-trained AI classifiers, making it readily adaptable to field conditions in low-resource settings. In this study, the approach was successfully adapted to field use, effectively predicting natural malaria infections in blood samples from a population-level survey in Tanzania. With additional field trials and validation, this technique could significantly enhance malaria surveillance and contribute to accelerating malaria elimination efforts.

Fig. 1

Schematic flow of experimental setup used to create DBS of different parasitaemia under non-anaemic (50%), moderate (25%), and severe anaemia (12.5%) and then scanned with MIR Spectrometer for spectra acquisitions and analysis using AI approaches

Fig. 2

The performance of the seven ML classifiers assessed through cross-validation on non-anaemic samples, following three approaches. a Accuracy scores of classifiers in high contrast against the negative class. b Evaluation with all concentrations, combining all parasite densities as positive against none. c Assessment in low contrast against the negative. Confusion matrices for trained and fine-tuned LR models on the 20% test set, with parasitaemia class similar to the training set, are displayed in panels d, e, and f. Receiver Operating Characteristics (ROC) and Area Under the Curve (AUC) of three tuned LR models are presented for high contrast (g), all concentrations (h), and low contrast (i)

Fig. 3

Spectral features with the greatest influence on the performance of the three models, a High contrast, b All concentrations combined, and c Low contrast training sets for the prediction of positive class (Red circle) and negative class (Blue circle). The size of the circle represents coefficient scores

Fig. 4

Visual comparison of mid-infrared spectra from blank filter paper with those containing varying levels of malaria parasitaemia in malaria-positive blood (6%, 0.1%, 0.002%, and 0.00003%) and malaria-negative blood (0%). The figure represents averaged spectra from 10 filter papers, each scanned 32 times, i.e. 320 spectral scans

Fig. 5

Performance of three LR models on a completely unseen dataset held out prior to training, for non-anaemic (a), moderate (b), and severe anaemia (c). d–f represent the three-dimensional representation of the LR performance on the validation set for high contrast, all concentrations, and low contrast models, respectively

Fig. 6

Evaluation of three trained logistic regression (LR) models for classifying malaria infections in laboratory and patient samples. a Mean accuracy for each anaemia level across various parasitaemia levels using laboratory and field-collected DBS. b Confusion matrix of the high-contrast model predicting laboratory-combined parasitaemia against negative in non-anaemic conditions, simulating realistic field collections. c Confusion matrix indicated the high-contrast model’s performance in detecting malaria infections in realistic field-collected DBS. d False positive and false negative predictions by the high-contrast model averaged by anaemic conditions using both laboratory and field samples