Antifungal agents comprise four principal chemical categories: azoles, echinocandins, polyenes, and flucytosine. Azole antifungals inhibit cytochrome P-450-mediated ergosterol biosynthesis, resulting in fungal cell membrane disruption. Echinocandins inhibit cell wall biosynthesis by blocking (1→3)-β-D-glucan synthase (GS). Amphotericin B, primarily administered as a liposomal formulation (L-AmB), represents the only systemically applicable polyene, exerting fungicidal effects through interaction with ergosterol, membrane extracellular aggregation, and/or pore formation (e37). Flucytosine undergoes deamination in fungal cells and inhibits fungal nucleic acid biosynthesis.

Recent years have witnessed alterations in non-albicans Candida susceptibility to antifungal agents. Global antifungal surveillance analysis indicates that between 2016 and 2017, over 99% of C. albicans isolates remained susceptible to fluconazole and voriconazole. C. glabrata demonstrated overall fluconazole resistance of 6.5%, highest in the United States (13.0%), while voriconazole resistance increased to nearly 50% (Castanheira et al., 2020). C. tropicalis resistance to fluconazole and voriconazole has continuously increased from below 10% in 2009 to exceeding 30% (Xiao et al., 2018; Fan et al., 2019; Castanheira et al., 2020; Borjian Boroujeni et al., 2021).

C. auris exhibits extensive multidrug resistance, demonstrating resistance to nearly all first-line antifungal agents. In 2019, the CDC classified C. auris among five pathogens posing urgent threats to public health and as the first fungal pathogen in the urgent antimicrobial resistance threat category (CDC, 2022). C. auris demonstrates fluconazole resistance exceeding 90%, amphotericin B resistance of approximately 30%, and relatively lower echinocandin resistance at approximately 7% (Lockhart et al., 2017). Research on C. auris resistance mechanisms reveals an outer cell wall layer with high mannopolysaccharide density and low structural complexity, contributing to reduced innate immune recognition. Additionally, C. auris exhibits weaker activation of MAPK signalling pathways controlling pro-inflammatory cytokine and chemokine expression in macrophages, facilitating evasion of host immune responses (Wang et al., 2022). Azole-resistant C. auris strains demonstrate ERG11(K143R) and CDR1(V704L) gene mutations. Echinocandin-resistant strains initially exhibit FKS1(S639Y) mutations, followed by distinctive FKS1(F635C) mutations. Flucytosine-resistant strains manifest FCY1, FUR1, and ADE17 mutations. Pan-resistant C. auris isolates demonstrate uridine phosphoribosyltransferase deletion (FUR1[1Δ33]) and FUR1 expression elimination (Lockhart et al., 2017; Jacobs et al., 2022). Amphotericin B resistance mechanisms require further investigation. For C. auris infections, echinocandins may represent the optimal initial therapeutic approach. The multidrug resistance profile and high mortality associated with C. auris necessitate heightened vigilance, with accurate species identification being crucial for appropriate treatment and prevention.

Molecular detection technologies encompass traditional polymerase chain reaction (PCR), real-time fluorescent quantitative PCR (qPCR), loop-mediated isothermal amplification (LAMP), and high-throughput sequencing techniques (Pham et al., 2024). These methodologies, predicated on the detection of fungal DNA or RNA in clinical specimens, are critical for selecting appropriate antifungal therapy.

Fungal-specific antigen detection represents one of the commonly employed technologies for clinical auxiliary diagnosis of IFI, encompassing cell wall polysaccharide component determination, metabolite determination, and biomarker detection methodology platforms. Traditional detection methods include assays for 1,3-β-D-glucan (BDG) and galactomannan (GM). BDG represents a polysaccharide component of most fungal cell walls, including Candida species, Pneumocystis jirovecii, Aspergillus species, and Fusarium species. Currently, an FDA-approved detection method exists—the Fungitell assay (Associates of Cape Code, Inc., East Falmouth, MA, USA). BDG testing has been extensively employed for IC diagnosis and therapeutic guidance. In IC diagnosis, sensitivity ranges between 70% and 90%, but specificity demonstrates poorer performance at 50% to 70% (Haydour et al., 2019). Notably, Mucorales cell walls lack glucans, while Cryptococcus cell walls, although containing glucans, primarily comprise α-glucans and β-1,6-glucans. Consequently, BDG detection is unsuitable for Cryptococcus and Mucorales detection.

GM represents another cell wall biomarker for IFI, produced by various fungi including Aspergillus, Penicillium, Paecilomyces, Lichtheimia, and Histoplasma species. The Platelia Aspergillus enzyme immunoassay (Bio-Rad, Marnla cooche, France) represents an FDA-approved method for Aspergillus GM detection, demonstrating sensitivity and specificity of 78% and 93%, respectively, in bronchoalveolar lavage fluid (BALF) (de Heer et al., 2019). Traditional ELISA detection is limited by prolonged processing time and stringent detection environment requirements, consequently not practically implemented in most hospitals.

Lateral flow chromatography testing technology represents an emerging point-of-care detection method for invasive aspergillosis, with two currently developed assays: the AspLFD lateral flow device (LFD) by OLM Diagnostics (Newcastle upon Tyne, UK) and the sōna Aspergillus galactomannan lateral flow assay (LFA) by IMMY (Norman, OK, USA) (Mercier et al., 2020). These typically target Aspergillus cell wall GM components, enabling serum antigen detection within 15 minutes to 1 hour. They demonstrate excellent performance in both serum samples and BALF, with BALF sensitivity of 0.73-0.94 and specificity of 0.8-0.9 (Mercier et al., 2020; Fan et al., 2025).

Computed tomography (CT), owing to its high resolution and rapid imaging capabilities, has become an essential tool for evaluating pulmonary infections. The most frequently employed pulmonary imaging modality in clinical practice is high-resolution computed tomography (HRCT). Common manifestations of IPA on HRCT include the “halo sign”, “air crescent sign”, and “cavitation” (Barac et al., 2023). Early radiological manifestations of IPA typically include ground-glass opacities and small nodules, which may progress to larger nodules with the halo sign, consolidations, cavitations accompanied by the air-crescent sign, and ultimately residual fibrotic remodeling (Alexander et al., 2021). During the COVID-19 pandemic, the incidence of invasive fungal infections in critically ill patients increased significantly, particularly COVID-19-associated pulmonary aspergillosis (CAPA) and COVID-19-associated pulmonary mucormycosis (CAPM) (Koulenti et al., 2023). Pulmonary mucormycosis predominantly occurs in patients with diabetic ketoacidosis, solid-organ transplantation, or profound immunosuppression. CT typically reveals large areas of parenchymal consolidation (often >4 cm), cavitation, and the reverse halo sign (RHS)—a central ground-glass opacity encircled by a rim of consolidation—which is observed more frequently in pulmonary mucormycosis than in IPA (Jang et al., 2025). CT provides sensitive detection of pulmonary nodules, infiltrates, and other structural abnormalities, facilitating early recognition of radiological features suggestive of invasive fungal infection.

Magnetic resonance imaging (MRI), as another crucial imaging technology, offers unique advantages in detecting central nervous system, sinus, deep soft tissue, and other deep-seated fungal infections. MRI provides more detailed soft tissue imaging, facilitating identification of infectious foci and their impact on surrounding tissues, which is essential for comprehensive assessment of critically ill patients. Research indicates that MRI effectively enhances early diagnosis of acute invasive fungal rhinosinusitis (Voruz et al., 2025).

Positron emission tomography-computed tomography (PET-CT) combines the advantages of metabolic functional imaging (PET) with anatomical structural imaging (CT), capable of revealing metabolically active infectious or inflammatory lesions throughout the body by tracking radioactively labeled glucose analogs (such as ¹⁸F-FDG) (Kung et al., 2019). In invasive fungal infections, PET-CT holds significant value for detecting disseminated systemic infections (such as candidemia, disseminated invasive aspergillosis) or occult lesions. For example, in immunocompromised patients, PET-CT sensitively identifies deep tissue infections (such as pancreatic candidiasis, Aspergillus myocarditis, arthritis) that are difficult to detect with conventional imaging, while assessing infection extent and activity (Xue et al., 2024; Kim et al., 2025; Ravindra et al., 2025). Additionally, PET-CT can be integrated with microbiological or molecular detection results to optimise antifungal treatment duration and intensity. However, clinical application of PET-CT is limited by high costs, radiation exposure, and non-specific uptake (tumours or non-infectious inflammation may interfere with result interpretation).

Overall, innovative imaging technologies provide crucial support for early diagnosis in critically ill patients, assisting clinicians in making more accurate diagnoses at early stages, thereby facilitating timely implementation of effective therapeutic measures to reduce patient mortality and complication rates. With continued development and application of imaging technologies, additional innovative approaches are anticipated to be introduced into clinical practice, providing more comprehensive support for management of critically ill patients.

In the clinical domain, prediction tool development has traditionally been based on background knowledge and statistical models, using parametric (such as logistic regression) or semi-parametric (such as Cox regression) approaches. Poissy et al. conducted a prospective multicentre matched case-control study incorporating 192 candidemia patients (Poissy et al., 2020). Using logistic regression and stepwise regression methods to screen for IC risk factors, the study identified total parenteral nutrition, acute kidney injury, cardiac disease, previous septic shock, and aminoglycoside exposure as independent risk factors for candidemia in ICU populations; they subsequently developed a candidemia prediction score for ICU patients with an AUC of 0.768. Playford et al. performed a multicentre prospective cohort study of 6,685 ICU patients in Australia, constructing an invasive candidiasis risk prediction model that stratified ICU patients into high-risk, moderate-risk, and low-risk groups, facilitating more targeted early antifungal therapy implementation (Playford et al., 2016). This model demonstrated an AUC of 0.82. Guillamet et al. utilised information from 2,597 patients to establish a clinical prediction rule for candidemia in patients with severe sepsis and septic shock based on logistic regression methodology. This prediction rule achieved an AUC of 0.798 (Guillamet et al., 2015). Stanzani et al. employed clinical information from 1,944 adult patients with haematological malignancies to construct a model using logistic regression methodology to predict patients’ risk of developing invasive mould disease (IMD) within 60 days after hospital admission for haematological malignancy treatment. This model demonstrated an AUC of 0.85 (Stanzani et al., 2019). Rauseo et al. developed a prediction model based on 539 candidemia patients using multivariate logistic regression and stepwise regression methodologies to identify patient populations with lower fluconazole resistance risk. This model demonstrated a c-statistic of 0.771 and sensitivity of 90.3% (Rauseo et al., 2022). Ostrosky-Zeichner et al. developed a risk assessment model predicting fluconazole treatment failure in candidemia patients (Ostrosky-Zeichner et al., 2017). This predictive model identified high-risk patients for fluconazole failure, demonstrating the potential value and feasibility of personalised candidemia treatment, with a model c-statistic of 0.65. Zhao et al. prospectively analysed 93 renal transplant patients receiving voriconazole treatment, employing logistic regression to identify predictors of adverse events while determining that CYP2C19 phenotype, platelet count, haemoglobin, and concomitant esomeprazole use were determinants of voriconazole trough concentrations (Zhao et al., 2021).

Collectively, these studies demonstrate that traditional statistical approaches—predominantly logistic regression–based models—have successfully yielded clinical prediction tools for invasive candidiasis and invasive mould infections across diverse high-risk populations. Such models consistently achieve moderate-to-good discriminative performance, underscoring their clinical utility in early risk stratification and timely initiation of antifungal therapy. Nonetheless, their dependence on a limited set of predefined clinical variables constrains their capacity to capture the biological complexity and heterogeneity inherent to invasive fungal infections, resulting in variable generalizability across healthcare settings and patient cohorts. Consequently, the integration of dynamic, high-dimensional clinical and laboratory parameters, and potentially multi-omics data, represents a highly promising avenue to advance individualized risk assessment.

Machine learning algorithms effectively model non-linear relationships between variables and outcomes, as well as non-linear relationships among predictive variables themselves. The establishment of disease early warning models using machine learning algorithms has been widely applied to early prediction of various diseases. Within the field of invasive fungal infections, Bhavani et al. employed logistic regression and gradient boosting machine (GBM) models to develop prediction algorithms for bacteremia and fungemia among hospitalized patients. Prior history of fungal disease and the interval between admission and the first positive blood culture emerged as the most influential predictors of fungemia. The GBM model demonstrated robust performance, achieving an AUC of 0.88 in internal validation, outperforming conventional clinical scoring systems such as the SIRS criteria, the Infection Probability Score, and the Modified Early Warning Score, and exhibiting superior discriminatory capacity (Bhavani et al., 2020). Meng et al. constructed a machine-learning–driven prediction model for candidemia in intensive care unit patients and conducted external validation involving 77 individuals. The random forest model yielded the strongest performance, maintaining an AUC of 0.87 in the external cohort, indicative of promising generalizability (Meng et al., 2024). Wang et al. developed an early diagnostic model for invasive pulmonary aspergillosis (IPA-NET) based on deep learning, utilising patient information including CT images. IPA-NET provides a non-invasive, objective, and reliable methodology for early IPA diagnosis, demonstrating accuracy of 89.7%, sensitivity of 0.88, specificity of 0.91, and AUC of 0.95 (Wang et al., 2023). Ripoli et al. utilised data from 295 patients to develop a random forest-based early prediction model for candidemia in medical wards, achieving an AUC of 0.847 (Ripoli et al., 2020).

Beyond early diagnosis, source control and early effective systemic antifungal therapy are crucial for successful management of IFD. Early appropriate antibiotic therapy effectively reduces patient mortality risk and economic burden while decreasing antimicrobial resistance development. Antifungal resistance represents an increasingly serious problem in fungal therapy. Some Candida species considered fully susceptible to fluconazole, such as C. albicans, C. tropicalis, and C. parapsilosis, may develop fluconazole resistance, while isolates of typically resistant non-albicans Candida species (such as C. glabrata) may maintain susceptibility.

Currently, clinical treatment decisions are primarily guided by species identification and antifungal susceptibility testing. However, microbiological culture cycles are relatively lengthy, unfavourable for early patient treatment. Utilising machine learning algorithms for patient risk discrimination, implementing prophylactic antifungal therapy for high-risk patients, and employing higher-level antifungal agents for potentially resistant cases may improve IFD patient survival rates. Yin et al. developed the Deep Learning-Quantitative Structure-Activity Relationship Empirical Screening (DL-QSARES) method, integrating deep learning and quantitative structure-activity relationship screening to design novel antifungal peptides de novo. This approach successfully identified 49 candidate antifungal peptides, whose anti-Candida activity was validated in both in vitro and in vivo experiments (Yin et al., 2025). Sharma et al. developed the Deep-AFPpred model using transfer learning and the 1DCNN-BiLSTM deep learning algorithm to identify antifungal peptides from protein sequences. Deep-AFPpred significantly outperformed other state-of-the-art AFP classifiers on both validation and test datasets, achieving accuracies of approximately 96% and 94% respectively (Sharma et al., 2022). Rozova et al. employed natural language processing techniques and various machine learning algorithms, integrating available electronic health record data including cytology and histopathology reports, to develop an automated IFI detection tool (Rozova et al., 2023).

Li et al. utilised data from 246 IC patients to construct prediction models using random forest, logistic regression, and support vector machine algorithms. The research identified primary death-associated risk factors including serum creatinine, serum albumin, CRP, PCT, total bilirubin levels, age, hospitalisation duration, ICU length of stay, leukocyte count, and neutrophil count (Li et al., 2022). The random forest-based prediction model demonstrated optimal performance with an AUC of 0.919. A multicentre retrospective cohort study investigated clinical characteristics and risk factors for adverse events in cirrhotic patients with concurrent candidemia and intra-abdominal Candida infections using logistic regression models and backward stepwise regression algorithms (Bassetti et al., 2017).

In our review, the vast majority of studies reported only internal validation, with very few conducting external validation in independent cohorts. Multiple factors likely contribute to this gap: diagnostic practices differ across regions and institutions, patient populations vary substantially, and the quality of data from disparate sources is often inconsistent. These objective limitations hinder the broader implementation of prediction models in clinical practice. Consequently, many investigations remain confined to the model development phase, and substantial progress is still required before these tools can be reliably deployed in real-world clinical settings. Additionally, many existing models in related fields use microbiological cultures or colonisation markers (e.g. airway Aspergillus isolation in ICU patients) as outcomes, which may reflect fungal colonisation rather than proven/probable IFI as per EORTC/MSGERC criteria. More precise differentiation between colonisation and invasive disease will likely require incorporating specific fungal biomarkers—such as serum or BAL galactomannan, β-D-glucan, and fungal PCR assays—into future algorithms, either as predictors or as part of composite outcome definitions.

In summary, research regarding invasive fungal infections remains in the exploratory phase, with relatively few studies employing machine learning for fungal infection prediction models, primarily utilising traditional algorithms. Additionally, current research in related fields may represent colonisation rather than invasive infection, necessitating further clarification of IFI definitions. Finally, retrospective studies may demonstrate significant discrepancies from clinical realities. Consequently, prospective research is required to further promote interdisciplinary application of artificial intelligence technologies in the IFI field, leveraging their advanced capabilities to drive intelligent development of clinical diagnosis and treatment for invasive fungal infections.