Beyond the microscope: Unveiling bacterial vaginosis with AI‐powered multiomics data analysis

Recent advancements in production and access to large datasets, faster computing and cheaper storage have encouraged the development of machine learning (ML) algorithms with human-like artificial intelligence (AI) capabilities. Such algorithms are emerging as transformative tools in dermatology, enhancing diagnostic accuracy and patient care. Challa et al.1 use these tools in an attempt to validate AI's capacity to diagnose bacterial vaginosis (BV) adopting a training dataset that combines the vaginal microbiome and metabolome of patients and controls. Bacterial vaginosis is one of the most prevalent vaginal disorders among women of childbearing age globally.2 However, its pathogenesis remains elusive, the diagnosis time-consuming and experience-dependent and its treatment highly unsatisfactory, with, 50% rate of relapse within 6–12 months post-antibiotic therapy, currently the only evidence-based treatment available.2 The diagnosis of BV is particularly challenging because of BV's intricate polymicrobial features and a wide range of clinical presentations. Current diagnostic modalities for BV are the Amsel's criteria or the Nugent's score. The Amsel's criteria are strongly physician-dependent and have low sensitivity. The Nugent score is based on the evaluation of the gram stain of the vaginal smear and is superior in sensitivity, but it is time-consuming and easily influenced by the examiner's skills.2 Deep learning models based on the Nugent's score have been developed to predict BV efficiently and accurately.3 A simplified version of the Nugent's scores, the Hay/Ison criteria, are easier and faster to use.4 However, all score-based methods, even if AI-powered, do not allow for the identification of the overgrowing bacterial species to tailor treatment options for individual patients. Laboratory-based molecular tests, which identify DNA from a subset of BV-associated species, have recently been FDA approved.5 However, the characterization of the vaginal flora with DNA-based methods could still deliver false positive results, as pathogenic anaerobic bacteria can occur without causing disease. Moreover, they are expensive and not usable as point-of-care (POC) tests. Current strategies to develop POC diagnostic tools for BV include the qualitative detection of high levels of bacteria-produced metabolites in vaginal fluid samples through enzymatic assays. This could help developing cheaper diagnostic tools and give information on the nature of the pathogenic flora. Additionally, the study of bacteria-produced metabolite detection could allow for personalized interventions, such as the use of specific biofilm disrupters, probiotics and pH modulators, which could be added to the patient specific antibiotic treatment to reduce the rate of recurrence4. However, clinical studies, still have to prove any benefits of combined personalized approaches in BV.4 As a proof of concept, Challa et al.1 successfully combined the input data of 16S rRNA amplicon sequencing with metabolomics detected by liquid chromatography-mass spectrometry (LC–MS) to train a ML tool for the diagnosis of BV. Interestingly, the metabolome-based diagnostic model alone, mostly based on 14 identified key metabolites that correlate with the pathogenic flora, could predict BV presence quite fairly, with only marginal gain if combined with the 16S rRNA amplicon sequencing. The study has limitations, particularly the small number of study participants (40 per group) lacking population characteristics and risk factors evaluations. In fact, prevalence rates of BV vary significantly worldwide, influenced by socioeconomic status, age, hormonal fluctuations, sexual practices, diet, comorbidities and other unknown factors.2 An ideal metabolome-based AI algorithm to predict BV should be created with a very large training data set, representative of the global community, combining information on patient and population characteristics, vaginal microbiome sequencing and antibiotic resistance to help predict individualized risk of recurrence. Such an ideal approach could lead to the identification of new key metabolites and the development of cheaper enzyme-essay-based POC diagnostics and allow for patient-tailored therapeutics. The authors declare no conflicts of interest. Data sharing is not applicable to this article as no new data were created or analyzed in this study.