SEARCH · PubMed Health

Results for “Machine Learning”

Explore indexed PubMed citations for clinical trials, systematic reviews and public health research. Read source abstracts and follow each citation to its original PubMed record.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 307 records · Page 17Linked to original sources

PubMed1993

Representation for discovery of protein motifs.

There are several dimensions and levels of complexity in which information on protein motifs may be available. For example, one-dimensional sequence motifs may be associated with secondary structure identifiers. Alternatively, three-dimensional information on polypeptide segments may be used to induce prototypical three-dimensional structure templates. This paper surveys various representations encountered in the protein motif discovery literature. Many of the representations are based on incompatible semantics, making difficult the comparison and combination of previous results. To make better use of machine learning techniques and to provide for an integrated knowledge representation framework, a general representation language--in which all types of motifs can be encoded and given a uniform semantics--is required. In this paper we propose such a model, called a spatial description logic, and present a machine learning approach based on the model.

Amino Acids↗

PubMed2001

Using Bayesian networks in the construction of a bi-level multi-classifier. A case study using intensive care unit patients data.

Combining the predictions of a set of classifiers has shown to be an effective way to create composite classifiers that are more accurate than any of the component classifiers. There are many methods for combining the predictions given by component classifiers. We introduce a new method that combine a number of component classifiers using a Bayesian network as a classifier system given the component classifiers predictions. Component classifiers are standard machine learning classification algorithms, and the Bayesian network structure is learned using a genetic algorithm that searches for the structure that maximises the classification accuracy given the predictions of the component classifiers. Experimental results have been obtained on a datafile of cases containing information about ICU patients at Canary Islands University Hospital. The accuracy obtained using the presented new approach statistically improve those obtained using standard machine learning methods.

Algorithms↗

PubMed2026

Proteomics-Driven Strategies for Proximity-Inducing Drug Discovery.

In recent years, proximity-inducing drugs have emerged as a novel therapeutic modality that induces or stabilizes protein-protein interactions, especially by recruiting effector proteins to specific target proteins, thereby achieving functions beyond traditional inhibitors. The potential of proximity-inducing drugs extends beyond targeted protein degradation (TPD), as studies have demonstrated their ability to regulate biological processes such as signal transduction, gene transcription, chromatin regulation, and protein trafficking by modulating protein interaction networks. Rational discovery of proximity-inducing drugs requires clarifying their effects on protein-protein interactions, determining drug selectivity, and developing suitable ligands for drug construction. Proteomics has become a central technology in drug discovery, enabling global identification of the direct drug targets and systematic characterization of proteome-wide downstream responses. This provides a more refined map of drug mechanisms. In parallel, advances in machine learning applied to proteomic data, together with the expansion of proteome-wide ligandability maps, are further accelerating the discovery and optimization of proximity-inducing drugs. This review summarizes recent advances of proximity-inducing drugs, with a particular emphasis on how proteomics facilitates target space expansion, drug efficacy optimization, and ligandability discovery, alongside the emerging contributions of machine learning. Collectively, these insights aim to support the rational development of next-generation proximity-inducing drugs.

Drug Discovery↗

PubMed2026

Dietary Polyphenol Acteoside-Related Molecular Signatures in Clear Cell Renal Cell Carcinoma: Multi-Omics Profiling and Functional Validation of IMPDH1.

Clear cell renal cell carcinoma (ccRCC) is characterized by substantial metabolic and molecular heterogeneity, but the disease-relevant programs associated with acteoside, a dietary polyphenol, remain poorly understood. We integrated predicted acteoside targets with bulk, single-cell, and spatial transcriptomic data from ccRCC and combined molecular subtyping with cross-cohort machine-learning analysis. Acteoside-related signatures were preferentially enriched in malignant compartments and increased with tumor grade and stage. Consensus clustering identified two molecular subtypes with distinct biological and clinical features. C1 was associated with immune activation, metabolic activity, and more favorable survival, whereas C2 showed greater genomic instability, reduced renal epithelial differentiation, and poorer outcomes. We further benchmarked multiple machine-learning strategies and established a 10-gene prognostic model that retained predictive performance across independent cohorts, with IMPDH1 emerging as the strongest risk-associated feature. Functional experiments confirmed the biological relevance of IMPDH1: its knockdown suppressed ccRCC cell proliferation, DNA synthesis, colony formation, and migration, whereas overexpression produced the opposite effects. Together, these findings indicate that acteoside-related molecular signatures capture clinically relevant heterogeneity in ccRCC and provide a framework for linking dietary-polyphenol-related molecular space with tumor biology. The identification and functional validation of IMPDH1 further highlight its potential importance in ccRCC progression.

IMPDH1↗

PubMed2000

Visual management of large scale data mining projects.

This paper describes a unified framework for visualizing the preparations for, and results of, hundreds of machine learning experiments. These experiments were designed to improve the accuracy of enzyme functional predictions from sequence, and in many cases were successful. Our system provides graphical user interfaces for defining and exploring training datasets and various representational alternatives, for inspecting the hypotheses induced by various types of learning algorithms, for visualizing the global results, and for inspecting in detail results for specific training sets (functions) and examples (proteins). The visualization tools serve as a navigational aid through a large amount of sequence data and induced knowledge. They provided significant help in understanding both the significance and the underlying biological explanations of our successes and failures. Using these visualizations it was possible to efficiently identify weaknesses of the modular sequence representations and induction algorithms which suggest better learning strategies. The context in which our data mining visualization toolkit was developed was the problem of accurately predicting enzyme function from protein sequence data. Previous work demonstrated that approximately 6% of enzyme protein sequences are likely to be assigned incorrect functions on the basis of sequence similarity alone. In order to test the hypothesis that more detailed sequence analysis using machine learning techniques and modular domain representations could address many of these failures, we designed a series of more than 250 experiments using information-theoretic decision tree induction and naive Bayesian learning on local sequence domain representations of problematic enzyme function classes. In more than half of these cases, our methods were able to perfectly discriminate among various possible functions of similar sequences. We developed and tested our visualization techniques on this application.

Alcohol Dehydrogenase↗

PubMed2026

Spectral-Proteomic Integration Analysis (SPIA) Deciphers Molecular Trajectories of Breast Cancer and Enables Multitarget Therapeutic Assessment.

Raman spectroscopy and mass spectrometry-based proteomics offer deeply complementary yet largely disconnected views of cancer biology: the former provides a label-free, real-time biochemical phenotype, while the latter delivers a quantitative inventory of specific protein effectors. Bridging this gap remains a fundamental challenge in analytical biomedicine. Here, we introduce Spectral-Proteomic Integration Analysis (SPIA)─a novel, data-driven integrative framework that systematically links Raman spectroscopic phenotypes with quantitative proteomic profiles through machine learning and statistical correlation. Using a DMBA-induced rat breast cancer model with and without Toremifene (TOR) intervention, SPIA dynamically maps tumor microenvironment remodeling, capturing progressive collagen deposition and lipid metabolic reprogramming. An SVM classifier trained on Raman spectra achieves exceptional diagnostic accuracy (AUC ≥ 99.0%) and successfully predicts TOR therapeutic response. Proteomic analysis identifies 1,350 differentially expressed proteins, with convergent machine learning feature selection (LASSO, Random Forest, XGBoost) pinpointing core regulators including Luc7l2, Nucb1, Cbx3, and Csnk2a1. Crucially, Spearman correlation analysis between key Raman bands and core DEPs reveals strong, statistically robust associations (median ρ ∼ 0.75 in the 1533-1669 cm-1 region), empirically validating SPIA's core integrative logic. Leveraging this multimodal map, we elucidate a multitarget mechanism for TOR involving concurrent suppression of collagen deposition and correction of aberrant lipid metabolism. SPIA establishes a powerful, generalizable paradigm for integrating phenotypic and molecular data, with broad implications for biomarker discovery, drug mechanism elucidation, and precision oncology.

Animals↗

PubMed2026

Integrative TWAS and multi-omics analyses prioritize HSPE1 as a candidate risk gene for bipolar disorder with immune cell-specific regulatory evidence.

BACKGROUND: Bipolar disorder (BD) is a severe psychiatric disorder associated with substantial disability. Although genome-wide association studies have identified multiple BD-associated loci, the underlying genes and mechanisms remain incompletely understood. METHODS: We integrated a European-ancestry BD genome-wide association dataset with cross-tissue and tissue-specific transcriptome-wide association studies (TWAS) and complementary gene-based analysis. Candidate genes were further evaluated using differential expression analysis, consensus clustering, immune infiltration analysis, machine learning, summary-data-based Mendelian randomization, Mendelian randomization using single-cell expression quantitative trait locus data, single-nucleus transcriptomics, phenome-wide association analysis, and virtual screening. RESULTS: The integrative analyses prioritized 37 candidate genes. Peripheral-blood differential-expression analysis identified 14 genes that remained significant after FDR correction, and their expression profiles separated BD samples into two expression-defined clusters. Machine-learning analysis selected UNC50, LMAN2L, LYG2, HSPE1, and KANSL3 for an exploratory classification nomogram. SMR associated genetically predicted higher HSPE1 expression with increased BD risk in two blood eQTL datasets. Cell-type-specific analyses indicated HSPE1-related associations in T-cell and natural killer cell subsets, while single-nucleus analysis descriptively showed higher HSPE1 expression in medial thalamic T cells from BD samples. PheWAS identified no genome-wide significant associations for HSPE1, whereas virtual screening identified candidate compounds with favorable predicted docking scores against the HSPE1 structure. CONCLUSION: This integrative multi-omics study identified HSPE1 as a candidate BD risk gene with immune-cell-related regulatory evidence, providing insight into BD pathogenesis and supporting functional validation.

Humans↗

PubMed1997

Modeling obesity using abductive networks.

This paper investigates the use of abductive-network machine learning for modeling and predicting outcome parameters in terms of input parameters in medical survey data. Here we consider modeling obesity as represented by the waist-to-hip ratio (WHR) risk factor to investigate the influence of various parameters. The same approach would be useful in predicting values of clinical parameters that are difficult or expensive to measure from others that are more readily available. The AIM abductive network machine learning tool was used to model the WHR from 13 other health parameters. Survey data were collected for a randomly selected sample of 1100 persons aged 20 yr and over attending nine primary health care centers at Al-Khobar, Saudi Arabia. Models were synthesized by training on a randomly selected set of 800 cases, using both continuous and categorical representations of the parameters, and evaluated by predicting the WHR value for the remaining 300 cases. Models for WHR as a continuous variable predict the actual values within an error of 7.5% at the 90% confidence limits. Categorical models predict the correct logical value of WHR with an error in only 2 of the 300 evaluation cases. Analytical relationships derived from simple categorical models explain global observations on the total survey population to an accuracy as high as 99%. Simple continuous models represented as analytical functions highlight global relationships and trends. Results confirm the strong correlation between WHR and diastolic blood pressure, cholesterol level, and family history of obesity. Compared to other statistical and neural network approaches, AIM abductive networks provide faster and more automated model synthesis. A review is given of other areas where the proposed modeling approach can be useful in clinical practice.

Adult↗

PubMed1988

Physicochemical characteristics of non-electrolytes and their uptake by Brugia pahangi and Dipetalonema viteae.

The uptake of a diverse set of 14C-labelled non-electrolytes by Brugia pahangi and Dipetalonema viteae was measured relative to the free diffusion of tritiated water. Inulin was used as a non-absorbable surface marker to account for non-electrolyte adherent to the surface of the parasite which had not crossed the cuticle. B. pahangi and D. viteae took up the non-electrolytes to a similar degree; a comparison of tissue uptake indices gave a correlation coefficient of 0.99. Worm uptake could not be described by non-electrolyte octanol/aqueous partition coefficients alone. However, greater success was achieved using further descriptors and pattern recognition techniques for data analysis. The whole molecule descriptors log P, molar refraction, melting point, dipole moment and CNDO total energy were obtained from computer chemistry and the literature. Using a linear learning machine to relate uptake to these 5 physicochemical descriptors it was possible to successfully classify non-electrolytes as high or low uptake. Multivariate regression analysis of uptake versus these 5 parameters gave a correlation coefficient of 0.77. However, this was not statistically significant and therefore could not be used for quantitative predictions of substance uptake by worms. This illustrates the value of 'pattern recognition' techniques such as the linear learning machine. Using such 'pattern recognition' methods on a chemically related set of compounds it is anticipated that predictions of uptake can be achieved and improved upon. Such predictions could then be used in drug design.

Animals↗

PubMed2026

Firemaster 550 differentially alters gene expression underlying synaptic function in amygdala of prairie voles after gestational or lactational exposure.

Neurodevelopmental disorders often share similar behavioral diagnostic criteria including socioemotional and cognitive deficits. The prairie vole is a uniquely suitable model to study these deficits because they demonstrate strong social affiliation, bi-parental care, and partner attachment. Previously, we have shown that developmental exposure to the flame-retardant mixture Firemaster 550 (FM 550) impairs socioemotional behavior in the prairie vole and alters underlying neuroanatomy and function. However, the mechanisms for impaired pair bonding in males and increased anxiety in females remain unknown, along with the specific critical window(s) of vulnerability. Herein, we exposed prairie vole dams to FM 550 during gestation or lactation, and performed bulk RNA-seq on the amygdala, a hub of socioemotional processing, in their adult offspring. Two mathematically orthogonal methods were utilized for analysis, a linear statistical method and an ensemble machine learning method, incorporating sex as a biological variable. Gene ontology (GO) pathway analysis was performed following both and results compared to identify potential mechanisms of toxicity. GO results indicated consistent expression changes in the Synapse cellular component in all conditions, and implicated glutamatergic signaling specifically. Additionally, gestational exposure (GE) altered genes underlying modulation of synaptic transmission and neural development, while lactational exposure (LE) impacted genes underlying synaptic plasticity, axon guidance, and mitophagy. Machine learning identified disruption of endocrine system development, regulation of biosynthetic processes in GE animals, and suppression of various neuroinflammatory genes across multiple groups. Finally, we performed RNA expression analysis using Nanostring and demonstrated stronger correlation with the differentially expressed genes (DEG) of interest in females than males. Overall, this study demonstrates both the intersecting and distinct impacts of FM 550 exposure on amygdalar gene expression depending on sex and timing of exposure.

Animals↗

PubMed1999

Comparison of genetic algorithms and other classification methods in the diagnosis of female urinary incontinence.

Galactica, a newly developed machine-learning system that utilizes a genetic algorithm for learning, was compared with discriminant analysis, logistic regression, k-means cluster analysis, a C4.5 decision-tree generator and a random bit climber hill-climbing algorithm. The methods were evaluated in the diagnosis of female urinary incontinence in terms of prediction accuracy of classifiers, on the basis of patient data. The best methods were discriminant analysis, logistic regression, C4.5 and Galactica. Practically no statistically significant differences existed between the prediction accuracy of these classification methods. We consider that machine-learning systems C4.5 and Galactica are preferable for automatic construction of medical decision aids, because they can cope with missing data values directly and can present a classifier in a comprehensible form. Galactica performed nearly as well as C4.5. The results are in agreement with the results of earlier research, indicating that genetic algorithms are a competitive method for constructing classifiers from medical data.

Algorithms↗

PubMed2026

Integrated multi-omics analyses identify an RAS-SLC11A2-associated molecular framework linking iron metabolism with PCOS-related cardiometabolic risk.

INTRODUCTION: PCOS is a common endocrine disorder with elevated cardiometabolic risk, yet the role of the renin-angiotensin system (RAS)-iron metabolism axis in this comorbidity remains unclear. We explored its underlying mechanisms and evaluated the therapeutic potential of gentiopicroside. METHODS: Integrated multi-omics analyses combining transcriptomics, single-cell RNA sequencing, Mendelian randomization, machine learning, molecular docking, and in vitro functional assays were performed to identify shared molecular pathways and therapeutic targets across PCOS, hypertension, NAFLD, and T2DM. RESULTS: SLC11A2 was consistently dysregulated in PCOS transcriptomic datasets, and associated with iron metabolism, inflammatory response and oxidative stress pathways. Genetic analyses validated RAS-related regulation in hypertension susceptibility and revealed shared genetic architecture between PCOS and cardiometabolic traits. Network and single-cell analyses characterized SLC11A2-associated molecular patterns in disease-relevant cell types; machine learning identified disease-classifying molecular signatures. Gentiopicroside alleviated inflammatory and oxidative stress phenotypes, including reduced IL-6 expression and reactive oxygen species accumulation. CONCLUSION: This study defines an RAS-SLC11A2 molecular framework linking iron metabolism dysregulation to PCOS-related cardiometabolic risk, elucidating the mechanisms connecting ovarian dysfunction, inflammation, oxidative stress and hypertension, and supports gentiopicroside as a promising therapeutic candidate.

Humans↗

PubMed1996

Evaluation of automatic knowledge acquisition techniques in the diagnosis of acute abdominal pain. Acute Abdominal Pain Study Group.

Clinical diagnosis in acute abdominal pain is still a major problem. Computer-aided diagnosis offers some help; however, existing systems still produce high error rates. We therefore tested machine learning techniques in order to improve standard statistical systems. The investigation was based on a prospective clinical database with 1254 cases, 46 diagnostic parameters and 15 diagnoses. Independence Bayes and the automatic rule induction techniques ID3, NewId, PRISM, CN2, C4.5 and ITRULE were trained with 839 cases and separately tested on 415 cases. No major differences in overall accuracy were observed (43-48%), except for NewId, which was below the average. Between the different techniques some similarities were found, but also considerable differences with respect to specific diagnoses. Machine learning techniques did not improve the results of the standard model Independence Bayes. Problem dimensionality, sample size and model complexity are major factors influencing diagnostic accuracy in computer-aided diagnosis of acute abdominal pain.

Abdominal Pain↗

PubMed2001

Comparing syntactic complexity in medical and non-medical corpora.

With the growing use of Natural Language Processing (NLP) techniques as solutions in Medical Informatics, the need to quickly and efficiently create the knowledge structures used by these systems has grown concurrently. Automatic discovery of a lexicon for use by an NLP system through machine learning will require information about the syntax of medical language. Understanding the syntactic differences between medical and non-medical corpora may allow more efficient acquisition of a lexicon. Three experiments designed to quantify the syntactic differences in medical and non-medical corpora were conducted. The results show that the syntax of medical language shows less variation than non-medical language and is likely simpler. The differences were great enough to question the applicability of general language tools on medical language. These differences may reduce the difficulty of some free text machine learning problems by capitalizing on the simpler nature of narrative medical syntax.

Artificial Intelligence↗

PubMed2026

A regulatory network underlying idiopathic pulmonary fibrosis.

BACKGROUND: Idiopathic pulmonary fibrosis (IPF) is a progressive interstitial lung disease in which genetic susceptibility interacts with epithelial, immune, and mesenchymal remodeling. Although the chromosome 11p15.5 locus contains established IPF susceptibility signals near MUC5B and TOLLIP, the broader regulatory architecture of this region remains incompletely resolved. METHODS: We integrated IPF genome-wide association study summary statistics with methylation, expression, and protein quantitative trait loci using summary-data-based Mendelian randomization (SMR). SMR-prioritized candidates were evaluated in independent transcriptomic and methylation cohorts and further contextualized using microRNA, transcription-factor, protein-interaction, machine-learning, single-cell, and spatial transcriptomic analyses. Fibrosis-associated expression patterns were assessed in a bleomycin-induced pulmonary fibrosis rat model. RESULTS: The analyses recovered the established MUC5B and TOLLIP signals and prioritized BRSK2 as a comparatively underexplored candidate supported by eQTL-based SMR and independent molecular evidence. The BRSK2 pQTL association did not pass the HEIDI test and was therefore not interpreted as convergent protein-level genetic evidence. Network analyses linked BRSK2 to cell-cycle, metabolic-stress, and senescence-related programs, while cross-cohort machine learning prioritized FOXA2, CDC25B, and NFE2 as informative network features. Single-cell and spatial analyses localized BRSK2 preferentially to fibroblast and myofibroblast compartments and to regions with greater histological fibrosis severity. In fibrotic rat lungs, BRSK2 expression increased, whereas FOXA2 and CDC25B decreased at the transcript and protein levels. CONCLUSIONS: These findings refine the molecular landscape of the chromosome 11p15.5 IPF susceptibility locus and prioritize BRSK2 as a candidate component of an IPF-associated profibrotic fibroblast state. Its causal contribution, direct regulatory relationships, and therapeutic tractability require targeted mechanistic validation.

Idiopathic Pulmonary Fibrosis↗

PubMed1999

Molecular hashkeys: a novel method for molecular characterization and its application for predicting important pharmaceutical properties of molecules.

We define a novel numerical molecular representation, called the molecular hashkey, that captures sufficient information about a molecule to predict pharmaceutically interesting properties directly from three-dimensional molecular structure. The molecular hashkey represents molecular surface properties as a linear array of pairwise surface-based comparisons of the target molecule against a common 'basis-set' of molecules. Hashkey-measured molecular similarity correlates well with direct methods of measuring molecular surface similarity. Using a simple machine-learning technique with the molecular hashkeys, we show that it is possible to accurately predict the octanol-water partition coefficient, log P. Using more sophisticated learning techniques, we show that an accurate model of intestinal absorption for a set of drugs can be constructed using the same hashkeys used in the aforementioned experiments. Once a set of molecular hashkeys is calculated, its use in the training and testing of property-based models is very fast. Further, the required amount of data for model construction is very small. Neural network-based hashkey models trained on data sets as small as 30 molecules yield statistically significant prediction of molecular properties. The lack of a requirement for large data sets lends itself well to the prediction of pharmaceutically relevant molecular parameters for which data generation is expensive and slow. Molecular hashkeys coupled with machine-learning techniques can yield models that predict key pharmacological aspects of biologically important molecules and should therefore be important in the design of effective therapeutics.

Drug Design↗

PubMed2026

Assessing Metal Ion Assignment Accuracy in Protein Data Bank Models via Elemental Spectroscopy.

Accurate representation of metal ions in macromolecular structures is critical for chemical interpretation, computational modeling, and machine-learning methods that rely on Protein Data Bank (PDB) entries. However, the elemental identity of metals modeled in crystallographic structures is often inferred indirectly and rarely validated experimentally. Here, we combine Particle Induced X-ray Emission (PIXE) and X-ray Fluorescence Spectroscopy (XRFS) to determine the elemental composition of protein samples used to generate 70 deposited metalloprotein crystal structures. By analyzing the original protein material employed for crystallization, but before the addition of crystallization buffer solutions, we assess whether the modeled metal ions in deposited structures are consistent with experimentally detectable elemental content. We find that in a majority of cases, the metals modeled in the corresponding PDB entries are inconsistent with the metals present in the protein samples before crystallization, or that additional metals are present but not represented in the structural models. Spectroscopic results were integrated with automated crystallographic validation metrics, including real-space Z-difference (RSZD) analysis and systematic rerefinement, to evaluate atomic-number mismatch at metal sites. PIXE and XRFS show strong agreement for dominant elemental signals and provide complementary, scalable approaches for identifying suspect metal assignments. This work does not address physiological or functional metalation but instead highlights a widespread data integrity issue in deposited macromolecular structures, PDB-wide. These results establish an experimentally corroborated link between elemental identity and crystallographic validation metrics, enabling the large-scale detection of chemically inconsistent annotations in structural databases used for computational modeling and machine learning.

Databases, Protein↗

PubMed2026

An Exosomal Signature for Preoperative Detection of Occult Liver Metastasis in Pancreatic Cancer.

IMPORTANCE: Early liver metastasis (early-LiM) after pancreatectomy represents an aggressive biological phenotype of pancreatic ductal adenocarcinoma (PDAC) and is associated with markedly poor survival. Reliable preoperative biomarkers to identify occult hepatic micrometastasis remain lacking. OBJECTIVE: To develop and externally validate a circulating exosomal microRNA (exo-miRNA)-based machine learning model for preoperative detection of occult early-LiM in PDAC. DESIGN, SETTING, AND PARTICIPANTS: This multicenter retrospective case-control study included 3 phases: genome-wide discovery using exo-miRNA sequencing (discovery cohort), model development (training cohort), and independent external validation (2 validation cohorts). The study took place at 4 medical centers in China, Japan, and South Korea. A total of 372 patients were enrolled between 2011 and 2024. Data were analyzed from July 2024 to November 2025. EXPOSURES: Circulating plasma-derived exosomal miRNA expression profiles. MAIN OUTCOMES AND MEASURES: The primary outcome was early-LiM, defined as liver recurrence within 6 months after curative-intent resection. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and survival outcomes were assessed using Kaplan-Meier analysis. RESULTS: Among 372 patients with PDAC (median [IQR] age, 67 [59-73] years; 229 [61.6%] male and 143 [38.4%] female; median follow-up among survivors, 969 days),early-LiM was associated with significantly worse overall survival compared with other recurrence patterns (median OS, 9.1 months vs 26.6-31.8 months; log-rank P < .001). A 7-exo-miRNA extreme gradient boosting model demonstrated discrimination in the training cohort (AUC, 0.899; 95% CI, 0.822-0.976) and maintained performance in external testing cohorts (AUC, 0.876; 95% CI, 0.846-0.951 and AUC, 0.862; 95% CI, 0.744-0.981). The exo-miRNA panel score remained an independent identifier of early-LiM in multivariable analysis (odds ratio, 26.49; 95% CI, 18.45-55.28; P < .001) and stratified overall survival (log-rank P < .001). Decision curve analysis suggested improved net clinical benefit compared with conventional clinicopathologic variables. CONCLUSION AND RELEVANCE: In this multicenter study, a circulating exo-miRNA-based machine learning model enabled preoperative detection of occult early liver metastasis risk in PDAC. These findings support the potential of exosomal biomarkers to inform biology-guided treatment sequencing and warrant prospective validation.

Journal Article↗