Host-based infectious disease diagnostics, when integrated with machine learning, represent a significant technological effort to improve the speed and accuracy of identifying infection states. This approach shifts focus from pathogen-specific detection to analyzing the patient’s systemic biological response, which may include changes in cytokine profiles, transcriptional signatures, metabolites, and physiological signals. By interpreting these host responses, such systems could theoretically support clinicians in distinguishing between bacterial and viral causes of infection or identifying syndrome-based presentations like sepsis in environments where traditional laboratory testing is unavailable, delayed, or inconclusive. Recent investigations have sought to synthesize the current landscape of these diagnostic models, examining how researchers pair data modalities with specific machine-learning architectures and how performance is validated across diverse clinical cohorts.
A review of 144 potential records published between January 2020 and August 2025 identified 21 studies that met strict criteria for host-based machine-learning diagnostics. These included research on bacterial-versus-viral classification, pathogen detection in mixed febrile cohorts, and syndrome-specific diagnosis. The diverse input data utilized in these models were categorized into structured clinical variables, gene or molecular-expression measurements, medical imaging, and physiological signals. The analysis examined 54 specific evaluations of four focal machine-learning model classes: random forests, neural networks, support vector machines, and logistic regression. While other methods like gradient boosting and k-nearest neighbors were also reported, they were less frequently evaluated. The heterogeneity in patient populations, disease targets, and experimental designs prevented a direct meta-analysis of diagnostic accuracy, highlighting significant limitations in the current evidence base.
Random forests were the most frequently reported model family in the included studies, accounting for 23 of the 54 focal-class evaluations. This approach was particularly common in transcriptomic and molecular-expression studies, where researchers leveraged random forests for tasks such as tuberculosis classification and the identification of host-response signatures for viral infection. The popularity of random forest models in these contexts stems from their ability to handle high-dimensional, nonlinear data with relatively limited distributional assumptions. However, feature importance rankings generated by these models can be influenced by correlated predictors and variations in feature scales, necessitating careful interpretation. Despite their broad application, the study found no definitive evidence that random forests consistently outperformed other methods across different clinical scenarios.
Neural networks were the second most frequent class, appearing in 17 of the 54 quantitative evaluations. These models were predominantly applied to imaging and physiological-signal data, such as computed tomography scans, electrocardiographic traces, and microscopy images. Because neural networks can learn hierarchical spatial or temporal representations directly from high-dimensional raw inputs, they are well-suited for imaging-based diagnostic tasks. However, this flexibility comes at the cost of high data requirements and potential vulnerability to overfitting, particularly in studies lacking substantial training sets or external validation. The lack of interpretability in many neural network implementations remains a concern for clinical deployment, as these models often operate as black boxes without extensive post hoc analysis such as SHAP or gradient-weighted class activation mapping.
Support vector machines and logistic regression models were utilized in 9 and 5 evaluations, respectively. Support vector machines were frequently favored for their performance in high-dimensional, small-sample settings, common in proteomics and certain transcriptomic datasets. By constructing decision boundaries that maximize class separation, support vector machines can remain effective when the number of features exceeds the number of observations, provided that hyperparameter tuning and feature scaling are managed appropriately. Logistic regression, while simpler than the other three classes, offered significant advantages in interpretability. Its use was largely confined to structured clinical variables and compact biomarker panels, where researchers aimed to translate coefficients directly into clinically actionable scores or probability estimates.
Reporting practices across the 21 included studies revealed substantial variability that complicates reproducibility and clinical assessment. While accuracy and area under the receiver operating characteristic curve (AUC) were widely reported, these metrics frequently served as the sole indicators of performance. Specifically, 14 evaluations reported accuracy alone, while 20 relied on AUC alone. Relying on accuracy in imbalanced cohorts, where the majority class may dominate, can be particularly misleading, potentially masking poor performance in detecting the clinically critical minority class. The reliance on these metrics without supplemental information like confidence intervals or confusion matrices leaves a critical gap in understanding how these tools might behave in real-world clinical settings.
Patient-level covariates, including age, sex, comorbidities, and immune status, were rarely incorporated as predictive features in the diagnostic models. The analysis found that nearly half of the study-domain entries for these demographic and clinical variables were not reported, and when they were, they were often treated descriptively rather than as active model inputs. This lack of integration suggests that existing models may be vulnerable to confounding variables that could shift performance when applied to different patient populations. Without rigorous assessment of fairness or interaction effects, it remains unclear how these algorithms would perform across demographically and clinically distinct settings, raising concerns about the generalizability of reported diagnostic results.
Standardization in the field remains an urgent priority for future diagnostic research. The evidence synthesis observed that feature selection, preprocessing, and model validation were often insufficiently described, with several studies failing to include independent external validation cohorts. Without prespecified evaluation plans and common data dictionaries, comparing models across different modalities or disease targets becomes problematic. Future multimodal diagnostic studies should define how discordant predictions are converted into a final clinical result.
Despite the challenges, some studies demonstrated promising translational pathways, such as the use of targeted assays to reduce the feature space identified by high-dimensional broad discovery platforms. For instance, research combining viral-responsive genes with markers of bacterial infection suggests that multidimensional panels might offer better diagnostic resolution than individual gene signatures. Nevertheless, the transition from successful retrospective modeling to a functional clinical decision-support system requires more than just high AUC values. Prospective clinical trials are essential to determine whether these models can genuinely improve decision-making and patient outcomes compared to current standard-of-care diagnostics, which rely on conventional laboratory methods and symptomatic evaluation.
In conclusion, while machine learning for host-based diagnostics has the clear potential to enhance the classification of infectious states, the available evidence is too methodologically heterogeneous to identify a universally optimal model class. Model selection must be strictly guided by the specific nature of the input data and the clinical constraints of the diagnostic environment. The immediate path forward involves improving the reproducibility of studies through the consistent reporting of preprocessing, feature-selection, and validation methods. By shifting the focus toward generalizable, disease-agnostic outputs and validating models against external, independent datasets, the field can move toward developing practical decision-support tools that supplement traditional clinical expertise rather than merely serving as isolated demonstrations.
The literature search across Nature, PLOS, PubMed, and Elsevier initially retrieved 144 records published between January 2020 and August 2025. This intake yielded 21 records from Nature, 43 from PLOS, 27 from PubMed, and 53 from Elsevier. Screening occurred in two distinct phases: an initial host-diagnostics filter retained 51 candidates while excluding 93 off-topic or non-host studies, and a secondary filter eliminated 30 publications lacking an eligible machine-learning diagnostic model, leaving the final cohort of 21 investigations. Notably, the reviewers acknowledged that the search strategy was developed without a prospectively registered protocol, and they did not employ formal risk-of-bias assessment tools such as PROBAST-AI or QUADAS-2, conducting instead a descriptive appraisal of extraction fields across the heterogeneous corpus.
Individual studies illustrated distinct architectural strategies for integrating multiple diagnostic inputs. In an evaluation of COVID-19 severity, Elgedawy and colleagues merged 17 routine blood measurements with targeted metabolomic data into a unified feature matrix; this feature-level fusion increased the random forest area under the curve from 0.88 to 0.97 while maintaining an accuracy of 0.92. At the decision level, Ramachandran and colleagues paired cerebrospinal fluid host gene expression with metagenomic next-generation sequencing for tuberculous meningitis. In that workflow, metagenomic detection overruled the host classifier when non-tuberculous organisms were identified; concordance was 50 percent, increasing to 61 percent after excluding probable cases in which metagenomics detected a non-tuberculous pathogen, while the combined assay improved the blinded test area under the curve from 0.74 to 0.87. In contrast, Malic and colleagues translated a six-gene sepsis signature identified via RNA sequencing onto reverse transcription quantitative polymerase chain reaction and microfluidic platforms using gradient boosting, demonstrating assay adaptation rather than multi-assay decision fusion.
Methodological vulnerabilities were also evident in data preprocessing and validation rigor across the surveyed literature. Across the 21 studies, formal batch-effect assessments were absent or unreported in 13 publications, and six investigations performed feature selection on entire datasets outside their validation loops, creating potential risks of optimistic performance estimates. Furthermore, the peril of relying solely on headline accuracy was demonstrated by Vu and colleagues in a Kenyan cohort of over 6,000 cases. Although six candidate models achieved dengue classification accuracies between 0.91 and 0.92, five models recorded zero percent sensitivity and boosted logistic regression attained just 5.5 percent sensitivity, underscoring how summary accuracy can obscure an almost complete failure to detect active cases.
Subject of Research: Machine learning in host-based infectious disease diagnostic performance and standardization.
Article Title: Machine learning implementation for host-based infectious disease diagnostics
Article References: Harrington, S., Nace, S., Martinez, K. M., & Kubicek-Sutherland, J. Z. (2026). Machine learning implementation for host-based infectious disease diagnostics. Frontiers in Digital Health, 8, Article 1986989. https://doi.org/10.3389/fdgth.2026.1986989
Image Credits: AI Generated
DOI: 10.3389/fdgth.2026.1986989
Keywords: Machine Learning, Infectious Disease, Host-Based Diagnostics, Clinical Diagnostics, Transcriptomics, Random Forest, Neural Networks, Support Vector Machines, Logistic Regression, Diagnostic Accuracy, Evidence Synthesis, Clinical Data, Biomarkers, Model Generalization, Validation Standards, Sepsis Classification, Bacterial vs Viral Discrimination
News Source: Ophelia Keating. (November 17, 2026). Machine Learning Diagnostic Tools for Infections Lack Universal Performance Standards. Scienmag.



