Explainable AI models sharpen disease prediction across four medical data studies

Artificial intelligence is moving from retrospective curiosity to clinical decision support, but a new editorial argues that accuracy alone will not get it there. Writing in Frontiers in Molecular Biosciences, Yulin Zhang of Shandong University of Science and Technology reviews a Research Topic of four studies that apply machine learning to medical big data spanning metabolomics, transcriptomics, population health surveys and electronic health records. The editorial’s central claim is pointed: conventional statistics and so-called black-box models often impede biological interpretability and real-world clinical adoption, so explainable artificial intelligence, or XAI, must sit at the core of any pipeline that hopes to reach a physician’s desk.

The first study tackles Down syndrome, the genetic condition caused by trisomy 21. Colak and colleagues searched for pathway-specific metabolic biomarkers that distinguish individuals with Down syndrome from healthy controls using plasma liquid chromatography-mass spectrometry metabolomics data from 316 participants with T21 and 103 controls. Rather than betting on a single algorithm, the team compared five ensemble classifiers, AdaBoost, LightGBM, Random Forest, KTBoost and XGBoost, and KTBoost delivered the best classification performance. A SHAP-based explainability analysis then ranked L-Citrulline, Kynurenin, Prostaglandin A2/B2/J2, Urate and Pantothenate as the top biomarker candidates, a molecular signature the authors link to oxidative stress, mitochondrial dysfunction and inflammatory dysregulation in T21.

The second contribution is less a model than a blueprint for proving one. Chen and Bo designed a five-phase, multi-center, double-blinded clinical trial to validate deep learning-based early detection of severe influenza using electronic health record data from 87 hospitals in Jiangsu, China. The trial hypothesizes that AI-assisted diagnosis will improve the area under the curve by more than 0.15 and cut the misdiagnosis rate by at least 30 percent compared with physician-only assessment, while maintaining consistent performance across subgroups defined by age, comorbidity and hospital resource level. In a field crowded with retrospective models that quietly fail on external data, a prospective, blinded, multi-hospital design is exactly the kind of rigor the editorial says the discipline needs.

Cardiovascular risk prediction gets the interpretable-ensemble treatment in the third study. Chen and coauthors built a two-level stacking-ensemble classifier that integrates Naive Bayes, Decision Tree, CatBoost and Gradient Boosting Tree to predict heart disease status from large-scale population health survey data. Stacking, which trains base learners and then feeds their outputs to a meta-learner, outperformed every single base model, reaching 86.69 percent accuracy and an AUC-ROC of 0.97. Global and local SHAP interpretation identified sleep duration, self-reported general health status, age category and BMI as the dominant cardiovascular risk factors, and characterized interaction patterns such as the joint effect of sleep and health status, giving clinicians person-level explanations rather than opaque probability scores.

The fourth study zooms in on the molecular boundary between benign and dangerous. He and colleagues combined bioinformatics and machine learning to screen for gene markers associated with the malignant transformation of colorectal polyps into colorectal cancer. By integrating bulk RNA-seq, single-cell RNA-seq and three feature selection algorithms, Boruta, LASSO regression and XGBoost, the team narrowed the field to six core genes, EIF2S3, GTF3A, HMGA1, HSP90AB1, PABPC1 and S100A11, and used them to build a ridge regression diagnostic model. The model achieved strong performance in both internal and external validation cohorts, qRT-PCR confirmed upregulated expression of these genes in colorectal cancer cell lines, and survival analysis tied high expression of EIF2S3 and S100A11 to poorer patient prognosis.

What makes the collection notable is not any single model but the shared architecture of the workflows. Metabolomics-oriented XAI prioritizes plasma biomarker candidates for a genetic disease; stacking ensembles combined with SHAP provide person-level risk explanation for population-derived epidemiological data; multi-algorithm transcriptomic filtering narrows gene signatures for pre-malignant lesions; and a multi-phase prospective trial framework translates a retrospective AI model toward clinical practice. Each study pairs custom-tailored datasets with well-designed analytical pipelines and, in most cases, explainability tooling to address unmet clinical needs for genetic syndromes, cardiovascular disease, colorectal pre-malignant lesions and severe influenza.

The editorial is equally candid about the limits. Common shortcomings across the four studies include a lack of large-scale prospective external validation, confounding clinical variables, and the persistent challenge of distinguishing correlation from causation. A model that achieves 0.97 AUC on survey data, for instance, still says nothing about whether changing sleep duration changes heart disease risk. The authors call for future research to prioritize multi-center prospective cohorts, harmonized multi-omics integration and intervention-oriented studies that can move biomarker candidates from discovery toward clinical application.

The takeaway for translational medicine is that neither high-accuracy machine learning nor omics profiling alone is sufficient. Interpretability and rigorous clinical validation are, in the editorial’s framing, the core pillars of translating AI-augmented precision medicine into practice. As XAI methods like SHAP become standard companions to ensemble classifiers, and as trial frameworks like the Jiangsu influenza study push models into blinded, multi-hospital testing, the editorial argues that the discipline now has a template for building diagnostic tools that clinicians can actually trust and verify.

Subject of Research: Explainable artificial intelligence for disease prediction and diagnosis using medical big data

Article Title: Editorial: Exploring AI’s role in disease prediction and diagnosis through medical big data

Article References: Zhang, Y. (2026). Editorial: Exploring AI’s role in disease prediction and diagnosis through medical big data. Frontiers in Molecular Biosciences, 13, Article 2005303. https://doi.org/10.3389/fmolb.2026.2005303

Image Credits: AI Generated

DOI: 10.3389/fmolb.2026.2005303

Keywords: artificial intelligence, explainable AI, machine learning, biomarkers, metabolomics, transcriptomics, Down syndrome, heart disease, colorectal cancer, influenza, predictive modeling, clinical validation