Background: Cervical cancer remains a major global health challenge, causing over 300,000 deaths each year and disproportionately affecting women in low? and middle?income countries (LMICs). Despite the long?term promise of HPV vaccination, there is an urgent and immediate need for rapid, affordable, and scalable screening technologies that are practical in real?world, resource?constrained settings. Current diagnostic methods are costly, infrastructure?dependent, and often inaccessible to women in these regions, highlighting the necessity for alternative approaches that are both clinically effective and feasible at the point of care.
Objectives: This project aims to address this gap through a two?fold research strategy. First, it investigates the use of infrared (IR) spectroscopy to analyse urine samples as a non?invasive method for detecting high?risk HPV. Second, it applies advanced chemometric techniques—including principal component analysis, cluster analysis, and linear discriminant analysis—to develop a decision?support system capable of early identification of HPV infection. By integrating these methods with machine learning (ML) and artificial intelligence (AI), the goal is to establish a robust, rapid, and low?cost screening tool suitable for LMIC contexts.
Infrared spectroscopy, a form of vibrational spectroscopy, generates a detailed molecular fingerprint of a sample by measuring the vibrational frequencies of chemical bonds. These signals reflect the biochemical composition of the sample at a genomic, proteomic, and metabolomic level. IR spectroscopy has previously demonstrated high diagnostic accuracy across conditions such as dementia, brain cancer, and endometrial cancer, and early research suggests similar potential for gynaecological cancers using urine. Our team has also used IR signatures to discriminate between different microbial species, underscoring its power for biological classification.
Methods: In Phase I of this study, 1,098 urine samples were analysed: 794 tested negative for high?risk HPV and 304 were positive. Among the positive cohort, 42 samples contained HPV16 and 48 contained HPV18. As with many medical datasets, class imbalance presents a challenge; algorithms trained on disproportionately distributed classes may produce misleading accuracy, a phenomenon known as the accuracy paradox. Despite this, the current AI model demonstrates performance that meets analytical requirements. The next phase will focus on improving robustness, particularly through Random Forest models, which inherently resist overfitting. Future development will incorporate deep?learning architectures to better manage unbalanced datasets, reduce noise (including hydration?related spectral drift), and model outlier behaviour more effectively.
Results: To support real?world deployment, a spectral?quality assessment tool has been developed to provide immediate user feedback, reducing reliance on highly trained operators. A prototype hosted on the Streamlit platform allows browser?based access and can operate offline once loaded. These quality?control features will become integral to a future standalone application and can be embedded directly into an FTIR?based spectrometer optimised for HPV-screening.
Key Findings: 
?1- A standardised spectroscopy protocol was developed in a high-resource setting (UK) and successfully transferred to an LMIC, where local technicians were trained to perform spectral data acquisition, demonstrating cross-site feasibility.
2- The model for detecting high-risk Human Papillomavirus (hr-HPV) infection showed promising discriminatory power, with a sensitivity of 82% and a specificity of 75%.
3- workflow was completed in under 10 minutes.

 

 

Detection for HPV