Increasing Sample Size Research Articles

An information theoretic limit to data amplification

Abstract In recent years generative artificial intelligence has been used to create data to support scientific analysis. For example, Generative Adversarial Networks (GANs) have been trained using Monte Carlo simulated input and then used to generate data for the same problem. This has the advantage that a GAN creates data in a significantly reduced computing time. N training events for a GAN can result in NG generated events with the gain factor G being greater than one. This appears to violate the principle that one cannot get information for free. This is not the only way to amplify data so this process will be referred to as data amplification which is studied using information theoretic concepts. It is shown that a gain greater than one is possible whilst keeping the information content of the data unchanged. This leads to a mathematical bound, 2 log(Generated Events) ≥ 3log(Training Events), which only depends on the number of generated and training events. This study determined the conditions for both the underlying and reconstructed probability distributions to ensure this bound. In particular, the resolution of variables in amplified data is not improved by the process but the increase in sample size can still improve statistical significance. The bound was confirmed using computer simulation and analysis of GAN generated data from the literature. 

The predictive power of profiling the DNA methylome in human health and disease.

Early and accurate diagnosis significantly improves the chances of disease survival. DNA methylation (5mC), the major DNA modification in the human genome, is now recognized as a biomarker of immense clinical potential. This is due to its ability to delineate precisely cell-type, quantitate both internal and external exposures, as well as tracking chronological and biological components of the aging process. Here, we survey the current state of DNA methylation as a biomarker and predictor of traits and disease. This includes Epigenome-wide association study (EWAS) findings that inform Methylation Risk Scores (MRS), EpiScore long-term estimators of plasma protein levels, and machine learning (ML) derived DNA methylation clocks. These all highlight the significant benefits of accessible peripheral blood DNA methylation as a surrogate measure. However, detailed DNA methylation biopsy analysis in real-time is also empowering pathological diagnosis. Furthermore, moving forward, in this multi-omic and biobank scale era, novel insights will be enabled by the amplified power of increasing sample sizes and data integration.

Overcoming methodological barriers in electronic nose clinical studies, a simulation data-based approach

Analysis of volatile organic compounds by electronic nose (e-nose) may address gaps in non-invasive screening for neoplasia. Machine learning impacts study design and sample size requirements, but guidance on clinical study design is limited. This study evaluates how neoplasia prevalence, augmented data, and the number of e-nose devices impact sample size requirements. Simulated e-nose breath test data were created using real-world study data. We examined the effect of varying neoplasia prevalence (50%-5%) and data augmentation on model performance, as well as the impact of using multiple devices. Prediction models were developed using single value decomposition and random forest, and convolutional neural networks. Model performance was displayed as area under the receiver operating characteristics curve and F1-score. Stable model performance was defined as the phase where additional data no longer increases model performance. We found that lower neoplasia prevalence significantly increased sample size requirements, with low-prevalence settings (5%) requiring up to five times more data than high-prevalence settings (50%) for stable model performance. Model performance varied between devices, and integrating data from multiple devices required larger sample sizes. Approximately 400 data points per device at 50% prevalence, and 2100 data points at 5% prevalence, were necessary to reach stable model performance. Concluding, sample size requirements for e-nose studies are heavily influenced by disease prevalence and the number of devices used. Limiting device variability and ensuring sufficient case and control samples per device are crucial for achieving reliable predictive performance. Specific requirements will vary based on sensor and disease characteristics.ClinicalTrials.gov Identifier:Clinicaltrials.gov Identifier NCT03346005 (model study) and NCT04357158 (validation study).

Comparative and evolutionary analysis of chloroplast genomes from five rare Styrax species

BackgroundStyrax, a vital raw material for shipbuilding, construction, perfumes, and drugs, represents the largest and most diverse genus in the Styracaceae. However, there is a relative scarcity of research on Styrax, particularly in evolution and genetics. Therefore, this study conducted comparative and evolutionary analyses of the chloroplast genomes of five rare Styrax species (S. argentifolius, S. buchananii, S. chrysocarpus, S. finlaysonianus, and S. rhytidocarpus).ResultsThe results indicated that, despite high levels of conservation in chloroplast genome structure among these species, specific mutation hotspot regions exist, particularly involving the expansion and contraction of the IR region. Additionally, evidence of positive selection was detected in eight genes (atpB, ccsA, ndhD, petA, rbcL, rpoC1, ycf1, and ycf2), which may be associated with adaptive evolution in response to environmental changes. Phylogenetic analysis revealed conflicts between trees constructed based on coding sequences and complete chloroplast genomes for several species, which were similar to previous phylogenetic studies.ConclusionThis study underscores the importance of increasing sample sizes to enhance the accuracy of phylogenetic analyses and provides a new perspective on understanding the adaptive evolution of Styrax species. These findings are not only important for the conservation and sustainable use of Styrax, but also provide valuable insights for research in plant evolution and ecology within the genus.

On the validity of fMRI mega-analyses using data processed with different pipelines

Abstract In neuroimaging and functional magnetic resonance imaging (fMRI), many derived data are made openly available in public databases. These can be re-used to increase sample sizes in studies and thus, improve robustness. In fMRI studies, raw data are first preprocessed using a given analysis pipeline to obtain subject-level contrast maps, which are then combined into a group analysis. Typically, the subject-level analysis pipeline is identical for all participants. However, derived data shared on public databases often come from different workflows, which can lead to different results. Here, we investigate how this analytical variability, if not accounted for, can induce false positive detections in mega-analyses combining subject-level contrast maps processed with different pipelines. We use the Human Connectome Project (HCP) multi-pipeline dataset, containing contrast maps for N = 1,080 participants of the HCP Young-Adult dataset, whose raw data were processed and analyzed with 24 different pipelines. We performed between-groups analyses with contrast maps from different pipelines in each group and estimated the rates of pipeline-induced detections. We show that, if not accounted for, analytical variability can lead to inflated false positive rates in studies combining data from different pipelines.

Interventions to improve mood and/or social interaction in residents of long-term care facilities with dementia: A systematic review.

Interventions to improve mood and/or social interaction in residents of long-term care facilities with dementia: A systematic review.

The application of fNIRS in studies on occupational workload: a systematic review.

Occupational workload can contribute to significant health problems such as chronic stress, fatigue and burnout. To investigate the underlying mechanisms, it is necessary to monitor brain activity in real work environments. Functional near-infrared spectroscopy (fNIRS) is a portable, non-invasive neuroimaging method that captures neural correlates of occupational workload under natural conditions. However, despite its increasing application, a comprehensive overview of fNIRS-based research in this field is lacking. Therefore, this systematic review examines how fNIRS can be utilized to investigate occupational workload. Following PRISMA 2020 guidelines, we conducted our systematic review by searching Web of Science, PubMed, and Scopus between November 15, 2023 and March 20, 2025. We included all studies published in English or German at any date, as long as they examined healthy adult professionals performing occupational tasks with functional near-infrared spectroscopy (fNIRS). Extracted data included study characteristics, workload details, signal processing methods, main fNIRS findings, and study quality, assessed using the JBI Critical Appraisal Tool. We included 41 studies. Of these, 23 reported a significant increase in oxygenated hemoglobin (HbO) concentration and functional connectivity in the prefrontal cortex (PFC) under higher occupational workload conditions. Only five studies examined typical office tasks. Nine studies analyzed differences in cortical activation between experts and novices, with experts showing increased HbO concentration in the PFC than novices. Regarding methodology, 26 studies used standardized optode placements, while only 17 applied systemic and extracerebral artifact correction. Small sample sizes and the absence of randomized controlled trials limited the reliability and reproducibility of the findings. Functional near-infrared spectroscopy effectively detects neural correlates of occupational workload and provides objective insights into cognitive demands in real-world work settings. Standardizing optode placement, harmonizing signal-processing methods, and increasing sample sizes would enhance the validity and comparability of future research. Expanding investigations to typical office environments is also crucial for understanding daily workload and for developing interventions that promote employee well-being and productivity. Overall, fNIRS represents a promising tool for establishing evidence-based workplace health promotion strategies across diverse occupational settings.

Abstract 637: Single-cell analysis of systemic immune response to sequential radiation and immune checkpoint blockade in metastatic colorectal cancers

Abstract Immune checkpoint blockade (ICB) has shown significant clinical efficacy in mismatch repair deficient colorectal cancers (CRCs). However, an effective means to use these therapies for mismatch repair proficient (pMMR) CRC has yet to be developed. To enhance the efficacy of ICB for patients with oligometastatic pMMR CRC, we performed a clinical trial evaluating the sequential use of stereotactic body radiotherapy (SBRT), pembrolizumab, surgical resection, and then adjuvant pembrolizumab (NCT02837263). The primary endpoint of 1-year recurrence-free survival (1 yr RFS) was met in 60% of trial participants compared to a historical control of 50%. From this trial, 13 subjects could have blood samples collected for peripheral blood mononuclear cell isolation pretreatment and after operative management. We hypothesize that the systemic immune response changes upon combined treatment and is distinct between responders and non-responders. We performed scRNA-Seq on these samples to compare the circulating immune cell populations with clinical outcomes. Despite a small sample size, integrated bioinformatics analysis identified (i) pre-existing CD8+ T cells were observed with treatment response (p = 0.031) (ii) combined RT and ICB induces changes in myeloid cell phenotypes, including reduced pro-inflammatory monocytes (p = 0.038) and conventional cDC2 (p = 0.030) in subjects without recurrence and increasing pDCs (p = 0.024) in most patients. CD8+ T cells from those recurrence-free subjects showed terminally differentiated effector phenotypes, expressing CX3CR1, GZMB, and KLRG1 before treatment. cDC2s expressed interferon activity markers, including STAT1, IFI44, and IRF7, that presented a more activated state. CD8+ T cell abundance remained stable, with no significant changes before and after treatment. In contrast, myeloid cells showed greater variability, reflecting unique polarization states through differences in cytokine (CCL3, CXCL8), interferon (ISG15, IFI44), and complement (C1QA, C1QB) associated genes. Future work will increase sample size in independent patient cohorts to elucidate the extent to which circulating phenotypes are also identified in tumors. Overall, this study demonstrates pre-existing T cell abundance and changes in myeloid cell phenotypes may be associated with clinical outcomes from combined immunotherapy, which have the potential to serve as predictive biomarkers for therapeutic response. Citation Format: Joshua Brand, Sean G. Kraus, Michael Bassetti, Dustin A. Deming, Huy Q. Dinh. Single-cell analysis of systemic immune response to sequential radiation and immune checkpoint blockade in metastatic colorectal cancers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 637.

Deep learning-based automatic segmentation of cerebral infarcts on diffusion MRI

We explored effects of (1) training with various sample sizes of multi-site vs. single-site training data, (2) cross-site domain adaptation, and (3) data sources and features on the performance of algorithms segmenting cerebral infarcts on Magnetic Resonance Imaging (MRI). We used 10,820 annotated diffusion-weighted images (DWIs) from 10 university hospitals. Algorithms based on 3D U-net were trained using progressively larger subsamples (ranging from 217 to 8661), while internal testing employed a distinct set of 2159 DWIs. External validation was conducted using three unrelated datasets (n = 2777, 50, and 250). For domain adaptation, we utilized 50 to 1000 subsamples from the 2777-image external target dataset. As the size of the multi-site training data increased from 217 to 1732, the Dice similarity coefficient (DSC) and average Hausdorff distance (AHD) improved from 0.58 to 0.65 and from 16.1 to 3.75 mm, respectively. Further increases in sample size to 4330 and 8661 led to marginal gains in DSC (to 0.68 and 0.70, respectively) and in AHD (to 2.92 and 1.73). Similar outcomes were observed in external testing. Notably, performance was relatively poor for segmenting brainstem or hyperacute (< 3 h) infarcts. Domain adaptation, even with a small subsample (n = 50) of external data, conditioned the algorithm trained with 217 images to perform comparably to an algorithm trained with 8661 images. In conclusion, the use of multi-site data (approximately 2000 DWIs) and domain adaptation significantly enhances the performance and generalizability of deep learning algorithms for infarct segmentation.

Data-Driven Assessment of Lung Cancer Patients Using Performance Status and Wearable Device Metrics.

Lung cancer remains a leading cause of cancer-related deaths globally, with non-small cell lung cancer (NSCLC) accounting for 85% of cases. Traditional methods for assessing the clinical status of cancer patients, such as Performance Status (PS), are subjective and may lack consistency across clinicians. Lung cancer remains a leading cause of cancer-related mortality worldwide. Monitoring the physical activity and PS of patients undergoing treatment is crucial for tailored therapeutic interventions. The LUPA study is a non-interventional, two-phase observational study aimed at assessing the usability of wearable devices and a mobile application for monitoring activity, sleep quality, and symptoms in lung cancer patients. A mixed-methods approach was used in Phase I to assess usability and data utility, while Phase II involved a one-group observational clinical study with 61 patients to explore correlations between clinician-reported PS and data collected through wearables. The results suggest moderate correlations between wearable data and ECOG-PS scores, but challenges remain in applying machine learning (ML) models to predict changes in patient condition. Future work should address model refinement, increased sample size, and the incorporation of additional features from wearable devices to enhance predictive accuracy.

Pharmacogenomics of chemotherapy induced peripheral neuropathy using an electronic health record-derived definition: a genome-wide association study.

Prior studies evaluating the genetic predisposition to chemotherapy induced peripheral neuropathy (CIPN) have been limited by small populations due to difficulty with real-world data extraction. This genome-wide association study (GWAS) evaluates the genetic differences between patients who developed CIPN against those unaffected, using an electronic health record (EHR) definition of CIPN. This study included all patients who received chemotherapy associated with CIPN and had germline genetic data within the biobank at the Colorado Center for Personalized Medicine. CIPN was defined by a new neuropathic pain medication or an ICD-diagnosis of neuropathy after specified chemotherapy initiation. GWAS were stratified by (1) total population, (2) platinum chemotherapy, (3) taxane chemotherapy, and (4) vinca alkaloid chemotherapy. Genes previously associated with CIPN were analyzed within each GWAS. Nine hundred fifteen patients received chemotherapy associated with CIPN, with 528 patients (57%) developing CIPN. Median age at chemotherapy initiation was 60.5years; female sex (n = 517, 56.5%) and White or Caucasian race (n = 822, 89.8%) were most common. Among single nucleotide polymorphisms (SNPs) that reached suggestive levels of genome-wide significance (p < 1 × 10-5), 60 SNPs occurred within 11 genes that may play a role in the development of or protection against CIPN, including RCOR1, CLDN14, TRIM5, and TMC2. No SNPs previously associated with CIPN achieved genome-wide significance in this population. This pharmacogenomic study suggests several genomic loci that may modulate the development of CIPN. This EHR-definition may allow for increased sample sizes and improved statistical power in future genetic studies of CIPN.

A Robust Framework for Probability Distribution Generation: Analyzing Structural Properties and Applications in Engineering and Medicine

This study introduces a novel trigonometric-based family of distributions for modeling continuous data through a newly proposed framework known as the ASP family, where ‘ASP’ represents the initials of the authors Aadil, Shamshad, and Parvaiz. A specific subclass of this family, termed the “ASP Rayleigh distribution” (ASPRD), is introduced that features two parameters. We conducted a comprehensive statistical analysis of the ASPRD, exploring its key properties and demonstrating its superior adaptability. The model parameters are estimated using four classical estimation methods: maximum likelihood estimation (MLE), least squares estimation (LSE), weighted least squares estimation (WLSE), and maximum product of spaces estimation (MPSE). Extensive simulation studies confirm these estimation techniques’ robustness, showing that biases, mean squared errors, and root mean squared errors consistently decrease as sample sizes increase. To further validate its applicability, we employ ASPRD on three real-world engineering datasets, showcasing its effectiveness in modeling complex data structures. This work not only strengthens the theoretical framework of probability distributions but also provides valuable tools for practical applications, paving the way for future advancements in statistical modeling.

Structural Development of Speech Networks in Young Children: A Cross-Sectional Study.

Abstract Understanding the structural development of the neural speech network in early childhood is critical for characterizing speech acquisition. This study investigated speech in the developing brain by scanning 94 children aged 4-7 years using diffusion-weighted imaging (DWI) MRI. To increase sample size and performance variability, children with attention-deficit hyperactivity disorder (ADHD) were included from a larger ongoing study (n = 47). Each child completed the Syllable Repetition Task (SRT), a validated measure of phoneme articulation. DWI data were modeled with restriction spectrum imaging (RSI) to examine restricted and hindered diffusion properties in both grey and white matter. Analyses included both whole brain and automated fiber quantification (AFQ) approaches to assess six fiber pathways critical for speech development. Whole brain analysis revealed associations between SRT performance and restricted diffusion in bilateral inferior frontal gyrus (pars opercularis), right pre-supplementary and supplementary motor areas, and bilateral cerebellar grey matter (p &amp;lt; .005). Age moderated associations in left pars opercularis and the frontal aslant tract (FAT); however, only cerebellar findings survived cluster correction. Further associations between SRT performance and restricted diffusion emerged in cortical association fiber pathways, particularly left FAT, and in cerebellar peduncles. AFQ analysis revealed differences between high and low performers, notably in left FAT, bilateral SLFIII, and cerebellar peduncles. These findings suggest individual differences in speech performance are reflected in grey and white matter structure and offer key insights into the developing neural networks that support speech in young children.

Virtual Escape Rooms in Anatomy Education: Case Studies from Two Institutions.

Virtual escape rooms (ERs) require learners to solve puzzles and answer riddles while trying to "escape" a digital room. While the educational merit of such gamified learning activities continues to be realized, guides on the development of ERs are lacking, as well as student perceptions on how, if, and where they should be integrated into medical curricula. Therefore, the aim of this study was to describe the experiences of building anatomy-themed virtual ERs of differing formats at two separate institutions: Queen's University Belfast (QUB) and Edward Via College of Osteopathic Medicine (VCOM), focusing on abdominal and upper limb anatomy respectively. Google Workspace applications served as the primary platform. 3D models were built using photogrammetry techniques or Virtual Human Dissector software (www.toltech.net) and integrated into the ER. Of 69 students and staff invited at QUB, 9 (13%) participated in the in-person virtual ER in teams of 2-3 (7 medical students, 2 anatomy instructors). Of 27 VCOM medical students invited, 8 (30%) agreed to participate and individually completed VCOM's virtual ER remotely. Anonymous surveys and a focus group revealed the ERs to be enjoyable and engaging and encouraged participants to think about material in a new way while helping them to identify knowledge gaps. Strengths and weaknesses of different designs (linear versus nonlinear), delivery methods (in-person versus remote) and grouping of participants (team-based versus individual) were realized and discussed, revealing opportunities for optimizing the experience. Future studies would benefit from increasing sample sizes to assess the learning gain of such activities.

Unraveling overoptimism and publication bias in ML-driven science.

Unraveling overoptimism and publication bias in ML-driven science.

From genes to phenotypes: A review of multilevel omics techniques in beef quality.

From genes to phenotypes: A review of multilevel omics techniques in beef quality.

Magnetization of Iron Meteorites up to the Meter in Size as Possible Analogs for Asteroid Psyche

AbstractMeteorite paleomagnetic studies indicate planetesimal generated magnetic fields, but spacecraft magnetic measurements have yet to identify asteroidal natural remanent magnetization (NRM). This apparent discrepancy is of particular interest in the context of the NASA Psyche mission, which will search for evidence of past magnetic activity of the metal‐rich asteroid (16) Psyche. Here, we aim to test whether the NRM of meteorites inevitably drops below detectable values as specimen size increases, which could explain why asteroidal NRMs could never be detected. We focus on iron meteorites as possible analogs to (16) Psyche's constituent material. To do so, we measure the remanent magnetic field and estimate the NRM of samples of four iron meteorites with volumes between mm3 and m3. We find that their estimated NRMs decrease with increasing sample size but appear to plateau. These data are compatible with the idea that the bulk NRM of increasingly large objects becomes dominated by the fraction of this NRM produced by assemblages of magnetic minerals sharing a common magnetization direction. Moreover, all m3‐sized meteorites carry NRMs that are two orders of magnitude above the detectability limit of the Psyche Magnetometer, three of which are possibly pre‐terrestrial. These data, acquired on some of the largest masses of iron meteorites available on Earth, support the range of plausible NRM values for km‐size regions of (16) Psyche, used to establish the spacecraft Magnetometer's performance requirements. Nevertheless, large‐scale events such as brecciation of the asteroid following magnetization acquisition could always lower the asteroid's NRM below the detectability limit.

Impact of pre-hospital delays on the prognosis of older patients with diabetic foot ulcers: a 10-year retrospective study.

This study aimed to evaluate the impact of pre-hospital delay on the prognosis of older patients with diabetic foot ulcers (DFUs). In this 10-year retrospective study, enrolled patients were divided into three groups based on the number of days before admission to hospital: <7 days (Group 1); 7-14 days (Group 2); and >14 days (Group 3). Electronic medical data were gathered, including: independent variables (such as age, sex, body mass index, education level and duration of diabetes); and dependent variables (such as number of surgeries, therapeutic regimen, Wagner grade, white blood cell counts, length of hospital stay (LoS), outcomes and treatment costs). A total of 288 patients were included in the study, 192 of whom were male, and the median age was 65 years. Of the studied participants, 27 arrived at the hospital within seven days, while 223 arrived after >14 days since the onset of their DFU. Significant differences were observed in the distribution of age (p=0.03) among the three groups, with the age of Group 1 lower than that of Group 3 (p=0.02). Significant differences in the distribution of number of surgeries (p=0.01), LoS (p=0.04), outcomes (p=0.04) and costs (p=0.03) were also observed among the three groups. Analyses showed that Group 1 LoS was lower than that of Group 2 (p=0.02); the number of surgeries in Group 1 was lower than that of Group 2 (p<0.01) and Group 3 (p<0.01); and Group 1 costs were lower than those of Group 2 (p=0.03) and Group 3 (p=0.03). A positive relationship was observed between pre-hospital delay and Wagner grade (0.122; p=0.04), and a positive relationship was observed between Wagner grade and LoS (0.181; p<0.01), outcomes (0.294; p<0.01), and costs (0.289; p<0.01). The findings of this study showed that longer pre-hospital delays adversely affected outcomes, such as extended hospital stays, increased numbers of surgeries and elevated hospitalisation expenses, in patients with DFUs. Age may be an underlying factor for this; further study with an increased sample size and comprehensive data is warranted.

A Latent Trait-based Measure as a Data Harmonization and Missing Data Solution Applied to the Environmental Influences on Child Health Outcomes Cohort.

Collaborative research consortia provide an efficient method to increase sample size, enabling evaluation of subgroup heterogeneity and rare outcomes. In addition to missing data challenges faced by all cohort studies like nonresponse and attrition, collaborative studies have missing data due to differences in study design and measurement of the contributing studies. We extend ROSETTA, a latent variable method that creates common measures across datasets collecting the same latent constructs with only partial overlap in measures, to define a common measure of socioeconomic status (SES) across cohorts with varying indicators in the Environmental influences on Child Health Outcomes Cohort, a consortium of pregnancy and pediatric cohorts. Starting with 52 indicators of prenatal SES from 39,372 participants across 53 cohorts, ROSETTA created three factors representing key domains of SES: income and education, insurance and poverty, and unemployment. At least one factor score was available for 34,528 participants and two factors were available for more participants than any single indicator. Factors fit the data well, had content validity, and were correlated with alternative measures of SES (for income and education factor, r = 0.40-0.89). Higher SES as measured by the factor scores was associated with lower odds of prenatal smoking: odds ratio income and education : 0.42 (95% confidence interval: 0.38, 0.45). Missing data were reduced compared with most methods, except for multiple imputation. ROSETTA aids in pooled analysis of individual participant data by creating measures on a common scale and maximizing data in the presence of missing and mismatched measures.

Gaussian Process Regression Model Based on Random Sampling and Secondary Encoding Techniques

Gaussian Process Regression (GPR) is a flexible non-parametric method that has been widely used in various prediction tasks due to its superior performance in fitting nonlinear functions. However, as the sample size increases, the computational complexity of GPR models grows exponentially, limiting their application to large-scale datasets. To address this issue, this paper proposes a GPR model based on the Stacking framework. The core innovation of the model consists of two parts: first, random sampling techniques are employed to extract multiple subsamples from the original dataset, and independent GPR models are trained for each subsample. Since the subsample sizes are relatively small, this strategy effectively reduces computational complexity and further improves efficiency through parallel processing of multiple models. Second, to overcome the performance variance among different submodels, a model fusion mechanism is adopted. The predictions from the individual submodels are treated as new features, and a secondary GPR model is trained as a combiner to optimize the aggregation of these predictions. This two-layer structural design not only significantly reduces the computational cost of GPR but also enhances the generalization capability of the predictive model through model fusion. Simulation experiments and real-world data analyses demonstrate that the proposed method exhibits a clear competitive advantage over traditional regression models.