Screening of malaria infections using AI-powered infrared spectroscopy

Mshani, Issa H. (2025) Screening of malaria infections using AI-powered infrared spectroscopy. PhD thesis, University of Glasgow.

Abstract

Over the past two decades, malaria control efforts have averted 2.1 billion cases and saved 11.7 million lives globally, yet the disease still claims over 600,000 lives annually, mostly in sub-Saharan Africa. Key interventions like insecticide-treated nets, indoor spraying, and antimalarial drugs have driven success, but major challenges persist. Accurate, timely detection of malaria parasites and scalable population screening remain difficult, especially in low transmission areas. Although WHO promotes surveillance as a core pillar of elimination, resource constraints in low-income countries hinder the expansion of effective, affordable surveillance systems. Current malaria screening tools, such as rapid diagnostic tests (RDTs) and microscopy, are essential for detecting parasites but have limitations, particularly in low transmission settings and at low parasite densities, which hampers elimination efforts. While more sensitive methods like polymerase chain reaction (PCR) are available, they are costly and impractical for widespread use in resource-limited areas. As a result, there is an urgent need for sensitive, cost-effective, and scalable tools capable of detecting low-density infections, especially in low transmission contexts.

Recent studies have shown the potential of using artificial intelligence (AI) powered infrared spectroscopy to detect malaria parasites in human blood. This approach is reagent-free, robust, user-friendly, quick and potentially cost-effective. However, to address the gaps in current methods for malaria screening and diagnosis, it was important to also assess factors such as lowest detectable parasite density and performance in low transmission settings before adoption by malaria control programs.

The primary aim of my PhD research was to improve malaria surveillance by exploring the application of infrared spectroscopy and machine learning (IR-ML) for malaria screening in population surveys. To achieve this, I pursued five complementary objectives: (1) reviewing the potential applications of IR-ML for malaria surveillance, developing a target product profile, and identifying key considerations and research gaps for integrating IR-ML into control efforts; (2) demonstrating the performance of mid-infrared spectroscopy and machine learning (MIRs-ML) across varying parasite densities and anaemic conditions; (3) conducting cross-sectional surveys to map malaria burden in an endemic setting, assessing the performance of existing methods (RDTs, microscopy, and qPCR) for risk stratification; (4) evaluating MIRs-ML performance in areas with differing prevalence rates; and (5) developing a web-based platform to deliver MIRs-ML results to end users. The ultimate goal was to advance the development of MIRs-ML as a scalable malaria screening tool, adaptable to both high (prevalence rate >30%) and low transmission (prevalence rate <5%) settings, with the potential to transform malaria detection and monitoring.

To achieve the first objective, I reviewed the current state of infrared spectroscopy and machine learning (IR-ML) for malaria surveillance, comparing its advantages and limitations to existing tools like PCR, RDTs, and microscopy. This review identified research gaps and developed a target product profile (TPP) for integrating infrared technology into routine surveillance. For the second objective, I conducted lab experiments using blood from 70 malaria-free volunteers in Tanzania, diluted with cultured Plasmodium falciparum to create different parasitemia and anemia levels. These samples were used to create dry blood spots, which were then scanned using ATR-FTIR spectroscopy. Using supervised machine learning classifiers trained on a subset of the samples, we achieved over 90% accuracy in detecting malaria, even at low parasite densities, and across different anemia conditions. Field applications of these models demonstrated over 80% accuracy in predicting natural infections.

The third and fourth objectives involved cross-sectional surveys in 93 sub villages in southeastern Tanzania, screening 7,628 individuals using RDTs and microscopy, with two-thirds analyzed by qPCR. qPCR consistently detected higher transmission rates, revealing that RDTs and microscopy underestimate malaria prevalence, particularly in fine-scale mapping. I then used the survey data to evaluate MIRs-ML performance in areas with varying malaria prevalence rates from low to high. Again, the ML classifiers achieved over 90% accuracy and sensitivity in both high and low transmission settings. We also observed that performance was slightly lower in low transmission areas when trained exclusively on high transmission data, compared to when the models were trained with data from across all settings.

Finally, to make these models readily available to users in future, we developed a web-based platform that allows scientists and national programs to access pretrained ML models for instant malaria infection predictions. This platform is currently powered by models trained on over 5,000 human blood samples and 40,000 mosquitoes, and will continue to expand with data from Tanzania, Burkina Faso, and the UK. The ultimate goal is to democratize the applications of these models across different user groups in different countries.

In conclusion, through population surveys, I demonstrated the limitations of RDTs and microscopy for mapping malaria risk. I developed MIRs-ML in the lab to overcome these challenges and tested it in the field, showing its promise. This research has significantly advanced our understanding of the potential of MIRs-ML for malaria screening. It has demonstrated that the approach has high sensitivity and is capable of detecting parasite levels as low as one parasite/μl of blood, making it particularly suitable for large-scale population surveys and enhancing risk stratification efforts. The study also highlighted the limitations of current screening tools, such as RDTs and microscopy, which perform poorly in low transmission settings compared to the more sensitive PCR. This underscores the urgent need for new, more sensitive approaches for precise stratification. This PhD research shows that MIRs-ML could meet these needs, making it a valuable complement to existing surveillance methods and a promising tool for malaria screening, even in low transmission areas.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Additional Information: Supported by funding from Swiss TPH, the Royal Society, and the Bill and Melinda Gates Foundation.
• Subjects: Q Science > QA Mathematics > QA76 Computer software; Q Science > QR Microbiology
• Colleges/Schools: College of Medical Veterinary and Life Sciences > School of Biodiversity, One Health & Veterinary Medicine
• Funder's Name: Swiss TPH, The Royal Society (ROYSOC), The Bill and Melinda Gates Foundation
• Supervisor's Name: Babayan, Dr. Simon, Baldini, Dr. Francesco and Okumu, Professor Fredros
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-84885
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 06 Feb 2025 16:51
• Last Modified: 11 Feb 2025 09:17
• Thesis DOI: 10.5525/gla.thesis.84885
• URI: https://theses.gla.ac.uk/id/eprint/84885

Related URLs

• Article DOI
• Article DOI
• Article DOI