Risk Prediction Research Articles

From dosimetry to deep learning: personalized risk prediction models for radiation pneumonitis

Abstract Radiation pneumonitis (RP) is a common complication following radiotherapy for thoracic tumors, significantly impacting treatment efficacy and patient quality of life. Identifying and predicting risk factors for RP has become a key research focus. This study aims to summarize current knowledge by analyzing previously published studies and large clinical trials. A systematic literature search was conducted in Embase, PubMed, and Web of Science for publications up to November 1, 2024, using keywords such as “radiation pneumonitis”, “risk factors”, “machine learning”, etc. Inclusion criteria prioritized clinical relevance and methodological rigor. Identified RP-related factors include radiation dose parameters (e.g., V20, mean lung dose [MLD]), clinical characteristics (e.g., age, interstitial lung disease), inflammatory markers (e.g., IL-6, neutrophil-to-lymphocyte ratio [NLR]), and features from imaging and multi-omics analyses. In addition, traditional dosimetric indicators remain central, while recent advances integrate radiomics and artificial intelligence (AI)-driven models to improve predictive accuracy. Despite progress, challenges such as limited sample sizes, lack of standardization, and insufficient multi-center validation persist. Future efforts should prioritize data integration, model optimization, and clinical translation to better predict RP risk and guide individualized interventions.

Non-invasive urinary proteomic biomarkers for prognostic assessment in sepsis

Early identification of the death risk of sepsis may improve short-term prognosis. The objective of this study was to identify urinary proteomic biomarkers and create a model to predict short-term outcomes in sepsis patients. A total of 46 sepsis patients selected from the intensive care unit of a comprehensive tertiary hospital were enrolled in this study. We used data-independent acquisition (DIA) proteomics to detect proteins in the urinary of death patients (n = 14) and survivals (n = 32). KEGG and GO analyses were conducted to investigate the possible functions of these proteins. Feature variables were selected from the differentially expressed proteins using the Least Absolute Shrinkage and Selection Operator (LASSO) and the random forest algorithms and by determining whether their proteins had an area under the curve (AUC) greater than 0.8. Nomogram model and ROC curves were constructed to evaluate the predictive efficacy of these identified protein biomarkers. In total, 2570 proteins were identified in urine. Statistical analysis revealed that 255 proteins exhibited differential expression, with 146 being upregulated and 109 downregulated. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses highlighted the involvement of key genes in processes such as the negative regulation of hemostasis, organization of the cortical actin cytoskeleton, the Rap1 signaling pathway, and cytoskeletal dynamics in muscle cells. Utilizing LASSO regression, random forest analysis, and a receiver operating characteristic (ROC) curve with an area under the curve (AUC) greater than 0.8, we identified potential protein biomarkers for predicting sepsis prognosis. Additionally, a nomogram incorporating biomarkers Solute Carrier Family 25 Member 24 (SLC25A24), Ubiquilin-1 (UBQLN1), and Cyclic AMP-responsive element-binding protein 3-like protein 3 (CREB3L3) demonstrated superior predictive accuracy for assessing the risk of sepsis-related mortality. This study has identified several novel proteomic biomarkers and has developed a practical prediction nomogram utilizing SLC25A24, UBQLN1, and CREB3L3 for the individualized prediction of sepsis mortality risk. This nomogram serves as a valuable tool in facilitating personalized treatment strategies.

Incidental Finding of Coronary and Non-Coronary Artery Calcium: What Do Clinicians Need To Know?

This review summarizes the role of incidentally and non-incidentally discovered coronary artery calcification (CAC) and the evolving role of non-coronary artery calcification in atherosclerotic cardiovascular disease (ASCVD) risk assessment. Additionally, this review explores the emerging use of artificial intelligence (AI), machine learning (ML), radiomics, and natural language processing (NLP) for automated detection, quantification, and communication of these incidentally discovered findings. This review summarizes recent findings in the space, including the development of various AI/ML-based approaches for automated calcification quantification and detection. Recent work leverages the use of incidentally discovered CAC and non-coronary calcification (e.g. aortic valve, aortic arch, carotid artery, breast arterial calcification) and their influence on clinical decision-making and prescribing practices. CAC and various forms of non-coronary artery calcifications are increasingly recognized as powerful and additive predictors of ASCVD risk. Advances in AI, ML, and radiomics enable scalable, automated measurement of both incidental and non-incidental CAC and non-coronary calcifications, which will facilitate more precise, personalized ASCVD risk stratification.

Understanding Prediabetes and Diabetes Among Vietnamese Americans: Exploring Biological, Psychological, and Social Factors.

Vietnamese Americans, a growing population in the United States, face unique challenges in managing diabetes due to cultural, social, and psychological factors. This study examined potential predictive for diabetes risk in Vietnamese Americans. A cross-sectional study was conducted with 304 Vietnamese American adults using validated PhenX-selected surveys and snowball sampling. Older age (adjusted odds ratios [AOR] = 1.18), mental health concerns (AOR = 4.50), higher BMI (AOR = 1.61), family history of diabetes (AOR = 16.11), and hypertension (AOR = 18.65) were significant independent predictors of diabetes or high diabetes risk (p ≤ .05). Gender, health numeracy, disability, and various social factors were initially significant but became non-significant after adjustment, suggesting confounding effects. Findings highlight the need for culturally tailored care for Vietnamese Americans with diabetes or high diabetes risk, focusing on body mass index as a modifiable predictor and other biological and health-related factors for focused targeting and disease management.

Circadian rhythms of gut microbiota and plaque vulnerability: mechanisms and chrono-microbiota modulation interventions.

The stability of atherosclerotic plaques constitutes the fundamental pathological basis for acute cardiovascular events, and their circadian rhythm characteristics highlight the essential role of dynamic interactions between the host and microorganisms. This review systematically elucidates the multifaceted mechanisms by which disruptions in the circadian rhythm of the gut microbiota contribute to plaque destabilization. Specifically, the microbiota modulates endothelial function, immune homeostasis, and vascular inflammation via rhythmic variations in metabolites. Perturbations in this rhythm compromise the structural integrity of plaques through a synergistic "metabolic-immune-vascular" network. Furthermore, the review unveils the bidirectional regulation between the host's circadian clock and the microbiota's rhythm. Innovatively, we propose "Chronotherapy-based Microbiome Modulation (CMM)," a strategy that reestablishes synchrony between the host and microbiota rhythms through time-restricted feeding, time-specific probiotics, and drugs targeting the circadian clock, thereby, it is possible to improve plaque stability by regulating the host's gut microbiota. The clinical translation of these findings requires overcoming technical challenges, such as personalized time window prediction and microbiota ecological risk assessment, and integrating multi-omics dynamic monitoring with AI modeling and optimization strategies. This review presents a novel perspective on the regulation of plaque stability.

Impact of soil particle size on lead distribution, geochemical speciation, and bioaccessibility in lead paint-contaminated residential soils.

Deteriorating lead (Pb)-based paint is a major source of Pb contamination in urban areas. Lead contamination in homes poses significant health risks, especially to children. Children ingest Pb-contaminated soil through their hand-to-mouth activity. The Bioaccessibility Research Group of Europe (BARGE) recommends using the < 250μm soil fraction for oral bioaccessibility assessment, while the USEPA suggests the < 150μm fraction for Human Health Risk Assessment (HHRA). However, these practices may underestimate risk, as smaller soil particles are more likely to adhere to children's hands and be ingested; moreover, real-world exposure often involves a mix of particle sizes. Research on Pb speciation and bioaccessibility across soil sizes in residential soils is crucial for accurate risk assessment. This study examined the total Pb, geochemical fractionation, and bioaccessibility of Pb in different size fractions in Pb paint-contaminated residential soils in Detroit, Michigan, to identify the optimal particle size for health risk assessments. Lead-contaminated soil was collected from 10 homes known to have Pb-based paint contamination and separated into 3 size fractions, i.e., < 250μm, < 150μm, and < 63μm. Total Pb concentrations in the soils ranged from approximately 36-650mg/kg across the size fractions, with the highest concentrations observed in the < 63μm fraction. Each fraction was analyzed for total and bioaccessible Pb concentrations, and geochemical speciation of Pb was performed. Results showed that overall, the highest Pb fraction was organic matter-bound (47.5%). However, Pb was mainly in the exchangeable form in the < 63μm fraction (33.1%) and contained higher total and bioaccessible Pb (22%) compared to the < 250μm (13%) and < 150μm (17%) size fractions. The study suggests that including < 63μm fractions in risk assessment may improve health risk prediction; however, comprehensive assessments should consider contributions from multiple particle size fractions to better reflect real-world exposure.

Machine learning-based risk prediction model for central nervous system involvement in diffuse large B-cell lymphoma.

Accurate prediction of CNS relapse in DLBCL remains challenging despite existing models like IPI and CNS-IPI. This study aimed to develop a machine learning (ML)-based prognostic model. A retrospective cohort of 664 R-CHOP-treated DLBCL patients was analyzed; 44 (6.6%) experienced CNS relapse at a median of 9.3 months. ML models, including Random Survival Forests (RSF) and Gradient Boosting Machines (GBM), were developed and validated using the entire cohort (n = 664), irrespective of CNS relapse. RSF demonstrated high discriminative ability (C-index: 0.91) and low prediction error (Integrated Brier Score [IBS]: 0.057), while GBM yielded comparable performance (C-index: 0.88, IBS: 0.042), both outperforming traditional scores such as IPI and CNS-IPI. Key predictors included extranodal site number, high-risk organ involvement, and ECOG performance status, although ECOG lost significance in Fine and Gray competing risks analysis, likely due to early mortality. ML-based models offer enhanced predictive accuracy and support personalized CNS risk assessment in DLBCL.

Construction and verification of a risk prediction model for patients with kinesophobia after cerebral hemorrhage surgery

ObjectiveTo establish a risk prediction model of kinesophobia in patients after cerebral hemorrhage surgery and verify its effect.MethodsA total of 218 patients after cerebral hemorrhage surgery were selected, and the differences in clinical data between kinesophobia patients and non-kinesophobia patients were analyzed. Using 20 indexes as independent variables, the characteristic variables were screened by LASSO regression, and then multivariate Logistic regression analysis was carried out. Based on the results, the nomogram prediction model was constructed, and the model was verified from the aspects of clinical applicability, discrimination, and calibration.ResultsSignificant differences were found in age, electronic health literacy score, depression score, NIHSS score, VAS pain score, intraoperative blood loss, and anxiety score between patients with phobia and non-phobia (P < 0.05). 12 characteristic variables were selected by LASSO regression. Multivariate Logistic regression analysis showed that age, NIHSS score, VAS pain score and depression score were independent risk factors for the occurrence of kinesophobia after cerebral hemorrhage surgery (OR > 1 and P < 0.05), and electronic health literacy score was an independent protective factor (OR < 1 and P < 0.05). Based on age, NIHSS score, VAS pain score, e-health literacy score, and depression score, a nomogram prediction model was constructed. The DCA curve shows that the model has the highest clinical net benefit when the threshold probability is between 0.14 and 0.99, indicating good clinical applicability. The area under the ROC curve (AUC) is 0.836(95% CI: 0.782–0.890), which indicates good discrimination. Spiegelhalter’s z test and the calibration curve show that the calibration degree is good, and the C statistic after Bootstrap self-sampling internal verification is 0.820 (95% CI: 0.772–0.877), indicating that the prediction is robust.ConclusionThe nomogram prediction model of the risk of kinesophobia after cerebral hemorrhage based on multivariate regression analysis has a good prediction effect, which can provide reference for the clinical prevention of kinesophobia after cerebral hemorrhage.

Radiomics-based tumor heterogeneity augments clinicopathological models for predicting recurrence in high-risk clear cell renal cell carcinoma after nephrectomy.

To investigate the association between CT radiomics-based tumor heterogeneity and recurrence-free survival (RFS) in high-risk clear cell renal cell carcinoma (ccRCC) after nephrectomy, and to determine whether integrating CT radiomics with clinicopathological model enhances recurrence risk prediction for adjuvant treatment decisions. This retrospective study included 194 patients with high-risk ccRCC undergoing nephrectomy. A radiomics model based on random survival forest was developed in the training set, using radiomics features extracted from pre-operative corticomedullary phase images. The performance of radiomics, Leibovich score, and the combined model were evaluated using Kaplan-Meier survival analysis, time-dependent receiver operating characteristic curves (time-AUC), time-dependent Brier scores, and decision curve analysis in external test set. During follow-up, 62 patients experienced recurrence. The radiomics model demonstrated superior predictive performance compared to the Leibovich score, with higher time-dependent AUCs (1-year: 0.882 vs. 0.781; 2-year: 0.865 vs. 0.762; 3-year: 0.793 vs. 0.797; all p < 0.05) and better calibration (lower Brier scores) in the test set. Decision curve analysis demonstrated that the combined model provided the highest net benefit, particularly for 2- to 3-year recurrence risk predictions. For high-risk ccRCC, CT radiomics provides incremental prognostic value beyond conventional clinicopathological models, enabling more precise recurrence risk stratification. This approach bridges imaging and precision oncology, with potential to optimize surveillance protocols and adjuvant therapy trial design.

A comprehensive survey on machine learning for workplace injury analysis: risk prediction, return to work strategies, and demographic insights

This survey paper explores the application of machine learning (ML) techniques in the domain of workplace injuries, focusing on three key areas: risk prediction, return to work (RTW) strategies, and demographic analysis. Through an extensive review of literature from January 2015 to July 2024, this paper examines the latest advancements in ML-driven approaches to workplace safety and identifies important research gaps. This paper highlights how classical ML techniques, such as ensemble models and decision trees, have become essential tools for identifying workplace injury risks, enabling more accurate interventions. It emphasizes the importance of leveraging ML in personalized RTW programs, which use data-driven insights to improve recovery outcomes and reduce economic demands. In the context of demographic analysis, this paper explores how ML algorithms can uncover disparities in injury rates across various age groups, industries, and occupations, underscoring the need for targeted safety measures. Moreover, research gaps are identified, particularly regarding the emerging potential of advanced ML techniques, such as deep learning and large language models (LLMs), for analyzing structured and unstructured safety data, methods that have not yet been widely applied in workplace injury research. As such, future research should apply recent advances in ML, integrating these approaches with comprehensive and accessible datasets to enhance the prediction and prevention of workplace injuries, provide more detailed analytics and insights, and improve safety protocols across all industries. This comprehensive survey is an invaluable resource for researchers and practitioners leveraging ML to address complex challenges in workplace safety.

Cross-population GWAS and proteomics improve risk prediction and reveal mechanisms in atrial fibrillation

Atrial fibrillation (AF) is a common cardiac arrhythmia with strong genetic components, yet its underlying molecular mechanisms and potential therapeutic targets remain incompletely understood. We conducted a cross-population genome-wide meta-analysis of 168,007 AF cases and identified 525 loci that met genome-wide significance. Two loci of PITX2 and ZFHX3 genes were identified as shared across populations of different ancestries. Comprehensive gene prioritization approaches reinforced the role of muscle development and heart contraction while also uncovering additional pathways, including cellular response to transforming growth factor-beta. Population-specific genetic correlations uncovered common and unique circulatory comorbidities between Europeans and Africans. Mendelian randomization identified modifiable risk factors and circulating proteins, informing disease prevention and drug development. Integrating genomic data from this cross-population genome-wide meta-analysis with proteomic profiling significantly enhanced AF risk prediction. This study advances our understanding of the genetic etiology of AF while also enhancing risk prediction, prevention strategies, and therapeutic development.

Association of Right Ventricular Dysfunction with Blood Volume Expansion to Outcomes in Chronic Left-Sided Heart Failure.

Background: The clinical significance of right ventricular dysfunction (RVD) and concomitant blood volume expansion (BVE) in chronic left-sided heart failure (HF) is not well characterized. Accordingly, we assessed the relationship of RVD severity to quantitative measures of BV on clinical outcomes. Methods: BV was measured in 150 clinically euvolemic patients at hospital discharge using nuclear medicine indicator-dilution methodology. Patients were stratified by severity of RVD and BVE ≥+25% of normal volume. Data were analyzed for association of RVD and BVE to composite outcome (HF-related mortality/1st re-hospitalization) at one-year post index hospitalization. Results: Absolute BV was larger in patients with ≥ moderate RVD (N=80) compared with normal RV function (N=36) (6.8±1.6 vs. 6.1±1.5 liters, p=0.029), Distribution of BV profiles within the two subgroups, however, varied significantly. BVE was demonstrated in 59% of patients with RVD [compared to 33% with normal RV (p=0.010)] and normal BV in 25% with RVD {compared to 44% with normal RV (0.042)]. Kaplan-Meir and Cox regression analyses support BVE with erythrocythemia (relative to normal) as independent predictors of better composite outcomes (p=0.037) while RVD was not associated with an additional increase in risk (p=0.137). Conclusions: This prospective analysis demonstrated a high prevalence of RVD (76% of cohort) and BVE (51%) at hospital discharge. Importantly, BVE with erythrocythemia was associated with better composite outcomes despite a high prevalence of RVD, while RVD alone was not a predictor of further risk. This suggests that BVE may serve as an adaptive response in part compensating RVD in chronic left-sided HF.

Liver Retransplantation: Identifying Factors and Developing a Risk Prediction Model to Predict Futility in the Modern Era.

Liver retransplantation (rLT) results have traditionally been inferior compared with those of primary liver transplantation. Understanding the risks and anticipated outcomes is essential for patient counseling and obtaining informed consent. Using the Scientific Registry of Transplant Recipients database, we analyzed a large cohort of adult rLT cases in the contemporary era to identify variables associated with posttransplant outcomes (with a focus on 1- and 5-y survival). Model predictions were made with random survival forests, a machine learning approach integrated into survival analysis. The difference in the out-of-bag C-index between a model with and without the variable was used to define variable importance. A prospective holdout cohort was used to validate the model predictions. Of the 3774 patients studied, the overall adjusted 1- and 5-y patient survival rates increased from 76.3% and 63.0%, respectively, for those transplanted in 2010, to 81.1% and 67.4% in 2019, then decreased to 78.0% and 63.9%, respectively, in 2022. The most important predictors of posttransplant mortality include recipient characteristics (being on life support before transplant, number of previous liver transplants, age, body mass index, and Karnofsky score) and donor organ characteristics (cold ischemia time and donor age). In a prospective validation cohort stratified into risk tertiles, the high-risk group had significantly lower 1-y survival (63.7%) compared with medium-risk (83.2%) and low-risk (88.7%, P < 0.001) groups. We developed a user-friendly online application using recipient and donor characteristics to predict 1- and 5-y survival. The study model could be used as an additional tool to predict 1- and 5-y patient survival to help counsel prospective rLT candidates and guide donor selection in this technically challenging recipient group.

Development and validation of personalized risk prediction models for patients with IgA nephropathy: anationwide multicenter cohort study.

Effective prediction of immunoglobulin A nephropathy (IgAN) progression is crucial for early intervention and management. We aimed to develop and validate distinct IgAN prediction models for clinical and researchapplications. We analyzed data from the Japanese Nationwide Retrospective Cohort Study in IgAN (n = 1174) gatheredover 10years. The models were developed and tested using data from general physicians in primary care, specialists in tertiary care hospitals, and researchers at academic research institutes. Three tailored prediction models (Primary Care, Tertiary Care, and Research Institute Models) were created to address the unique needs of different clinical environments. The primary outcome was a composite renal event defined as a 1.5-fold increase in serum creatinine level or progression to kidney failure. The predictive performance was assessed using C-statistics. In the derivation cohort, the primary care model included predictors such as estimated glomerular filtration rate < 45mL/min/1.73 m2, proteinuria ≥ 0.5g/day, and non-use of corticosteroids, achieving a C-statistic of 0.796 (95% confidence interval [CI] 0.686-0.895). The tertiary care model showed a C-statistic of 0.807 (95% CI 0.713-0.886), using predictors such as glomerular number and histological severity. The research institute model, incorporating 38 variables, demonstrated a C-statistic of 0.802 (95% CI 0.686-0.906). The prediction models for primary and tertiary care settings provided effective tools for forecasting renal outcomes in IgAN patients and are competitive with more complex machine learning-based models used in research. These models can help guide clinical decisions in various healthcare settings.

How I Treat Pediatric Pulmonary Embolism.

How I Treat Pediatric Pulmonary Embolism.

Inflammatory burden index predicts long term mortality in a nationally representative population from NHANES

This study investigated the relationship between the Inflammatory Burden Index (IBI) and risks of all-cause and cancer-specific mortality, focusing on its potential to enhance risk stratification. The research included a cohort of 14,835 participants from the American National Health and Nutrition Examination Survey. IBI was calculated using the formula CRP × (neutrophil / lymphocyte). Cox regression analysis was applied to assess the associations. During 223,719.71 person-years of follow-up, 3483 deaths (23.48%) occurred, including 778 (5.24%) from cancer. Mortality rates were 15.57 (all causes) and 3.48 (cancer) per 1,000 person-years. Kaplan–Meier analysis showed the highest IBI quartile had the lowest survival rates for all-cause and cancer-related mortality (Log-Rank p < 0.001). Adjusted models revealed a 23.4% higher risk of all-cause mortality and a 14.1% higher cancer-specific mortality per standard deviation increase in IBI. Smooth curve fitting confirmed a proportional relationship between IBI and mortality risk. ROC curve and reclassification analyses supported IBI’s role in improving mortality risk prediction. The findings of this study indicate noteworthy associations between IBI and both all-cause and cancer-specific mortality. Moreover, the results highlight the potential of IBI in enhancing risk stratification for incident all-cause and cancer-specific mortality within the general population.

Visualization of Early and Late Molecular Events Associated with the Development of Type 1 Diabetes Using NMR Spectroscopy.

Type 1 diabetes mellitus (T1DM) is an autoimmune disorder caused by the loss of insulin-producing pancreatic β-cells. This study aims to explore the correlation between hyperglycemia and concurrent metabolic perturbations during T1DM development to help identify biomarkers that differentiate between the early and established stages. Streptozotocin (STZ), a glucosamine nitrosourea compound, induces T1DM. Dose- and time-dependent studies were conducted in 7-8-week-old male C57BL/6 mice, who were administered increasing numbers of STZ injections (N = 0-5) and were followed for 15 (early) and 60 (late) days. The development of hyperglycemia was confirmed by performing an oral glucose tolerance test and an insulin tolerance test. A total of 50 abundant aqueous serum metabolites were identified and quantified using 1H NMR spectroscopy. In addition to glucose, a well-established biomarker for T1DM, a panel of 5 significantly perturbed metabolites (namely, leucine, choline, lactate, lysine, and mannose), Diagnostic Molecular Fingerprint (DMF), was identified. Unlike glucose levels, the proposed DMF (in combination with glucose) could differentiate not only between early and established stages of T1DM but also between young and aged healthy controls. However, these results need validation in humanized animal models and well-characterized patient cohorts of different ethnicities. In conclusion, the results obtained have contributed toward increasing the understanding of the pathophysiology and mechanism of T1DM establishment and progression that would possibly aid in accurate diagnosis, prognosis, risk prediction, defining the distinct stages of T1DM, and help in enhancing patient outcomes in the future.

Abstract A054: Multimodal LLM-Driven Intervention for Precision Risk Prediction in Lung Cancer Surgery

Abstract Introduction: Lung cancer surgery, while potentially curative, carries a 30% risk of serious postoperative complications, such as pneumonia or respiratory failure, increasing morbidity, mortality, and healthcare costs. Traditional risk assessment tools, based on static scores or clinical judgment, lack precision and adaptability to identify high-risk patients or reassess surgical candidacy. These tools also fail to provide interpretable outputs aligned with surgical workflows, limiting their clinical utility. To address these gaps, we propose a novel multimodal deep learning framework integrating clinical variables, imaging-derived radiomic features, and large language model (LLM) insights to predict complications accurately. Our model generates editable, clinician-friendly risk summaries, enhancing transparency, trust, and personalized surgical decision-making to improve outcomes and optimize planning. Methods: We analyzed data from 3,440 lung cancer surgery patients, combining 17 preoperative clinical variables (e.g., age, smoking history, pulmonary function) with CT imaging from 3,205 cases. From these scans, 113 radiomic features (e.g., texture, shape) were extracted using PyRadiomics. Our model integrates three modules: (1) a clinical data encoder, (2) a radiomics module for imaging-based risk, and (3) an LLM-based interpreter (using Llama 3.3, DeepSeek R1-Distil, OpenBioLLM, Clinical Longformer) to generate surgeon-like risk narratives from unstructured data. These narratives, linked to a binary postoperative pulmonary complication outcome, allow real-time edits that update risk predictions, reflecting surgeon expertise. The model was trained with a hybrid loss balancing accuracy and usability and evaluated using AUC-ROC. Results: Benchmarking traditional machine learning models (e.g., logistic regression, random forests) on our dataset yielded AUC-ROC values ranging from 76.8-78.2%. Our multimodal framework achieved an AUC-ROC of 75.0%, comparable to baseline models, while also providing interactive interpretable risk summaries not possible in baseline models that enable surgeons to refine predictions based on clinical expertise. Human surgeon assessments had a True Positive Rate of 44.9% and a False Positive Rate of 20.0%, underscoring our model’s prediction precision. Editable risk summaries allow surgeons to adjust predictions in real-time, enhancing transparency and supporting personalized surgical decisions. Conclusions: Our multimodal framework transforms preoperative risk assessment in lung cancer surgery by integrating clinical data, radiomic features, and LLM-driven insights. It surpasses human judgment and competes with traditional risk tools in predictive precision while offering exceptional interpretability through editable, clinician-friendly outputs. By enabling real-time risk adjustments, the model supports dynamic surgical planning, reduces complications, and enhances patient-specific care in lung cancer, with the potential to improve resource allocation and support AI-augmented precision medicine in thoracic oncology. Citation Format: Shubham Pandey, Bhavin Jawade, Srirangaraj Setlur, Venugopal Govindaraju, Kenneth P. Seastedt. Multimodal LLM-Driven Intervention for Precision Risk Prediction in Lung Cancer Surgery [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A054.

Abstract A056: Integrative Machine Learning Approaches for Predicting Prostate Cancer Risk Using Multi-Omics Data

Abstract Introduction: Prostate cancer is one of the most common cancers affecting men globally. According to the American Cancer Society, it is estimated that in 2025, there will be approximately 313,780 new cases of prostate cancer and about 35,770 deaths from the disease in the United States. This study aims to improve prostate cancer risk predictions by integrating multi-omics data (mRNA, miRNA, and methylation) using advanced machine learning techniques. Methods: We analyzed multi-omics data from 493 patients in the Cancer Genome Atlas Prostate Adenocarcinoma (TCGA-PRAD) dataset. Patients were stratified into low (Gleason score &amp;lt;=7) and high-risk (Gleason score &amp;gt;=8) groups based on their Gleason scores. Data normalization was performed using z-scores, and missing values were imputed using the missForest method. Differential expression analysis (DEGs) was conducted for mRNA, miRNA, and methylation data. To enhance predictive accuracy, machine learning models, including Lasso, Random Forest, SVM, XGBoost, and Gradient Boosting, were applied to various data combinations, employing 5-fold cross-validation. Model performance was evaluated using ROC curves and AUC values generated by the 'pROC' package, with DeLong's test for AUC comparisons between models. Two-side P value &amp;lt; 0.05 were considered statistical significance. All the analyses were performed using R. Results: The analysis identified significant differential expression: 186 upregulated and 468 downregulated genes in mRNA; 21 upregulated in miRNA (downregulated data not specified); 651 upregulated and 955 downregulated methylation sites. The dataset was randomly divided into training and testing sets in a 6:4 ratio. Gradient Boosting model showing exceptional effectiveness, especially those integrating mRNA with methylation, and miRNA with methylation with 100% and 89% AUC in the training and testing sets. Further analysis identified 70 target mRNAs, which were used to explore potential biological pathways implicated in prostate cancer. Pathway analysis using Ingenuity Pathway Analysis (IPA) highlighted the Calcium signaling and ABRA signaling pathways as potentially crucial in miRNA-mRNA interactions, suggesting their significant roles in modulating prostate cancer risk. These pathways are known to be critical for various cellular processes that could influence cancer progression. Conclusions: The integration of multi-omics data via machine learning significantly improves the prediction of prostate cancer risk, highlighting the potential of such models in clinical applications. Pathways analysis may provide new targets for therapeutic intervention. Our findings need to be validated with larger, independent external cohorts. Acknowledgement: This research is supported by The Hawaii Advanced Training in Artificial Intelligence for Precision Nutrition Science Research (AIPrN) (T32DK137523) Citation Format: Zhanwei Wang, Lenora WM. Loo, Herbert Yu, Youping Deng. Integrative Machine Learning Approaches for Predicting Prostate Cancer Risk Using Multi-Omics Data [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A056.

Factors influencing recurrence of preeclampsia in pregnant women with a history of preeclampsia and the establishment of a predictive model.

To establish and verify the prediction model of recurrent preeclampsia (PE) in pregnant women with a history ofPE. Totally 130 pregnant women with a history of PE from Jan 2021 to Jan 2023 were selected retrospectively. The patients were randomly matched according to the proportion of 1:4 to establish a verification group (nasty 26) and a modeling group (nasty 104). The modeling patients were divided into two groups according to the occurrence of preeclampsia: recurrent group (nasty 50) and non-recurrent group (nasty 54). Multivariate logistic regression analysis of influencing factors was established. Calibration curve was performed to verify, decision curve analysis (DCA) was used to evaluate the clinical practicability of the prediction model, and ROC analysis was used to show the prediction value of themodel. Multivariate logistic regression analysis showed that there were significant differences in age, gestational age, gestational interval, systolic blood pressure and diastolic blood pressure of previous pregnancy. (p<0.05) According to the results of logistic regression analysis, a prediction model was constructed. Logit(P)=(0.910Age)+(0.987Age of onset of previous pregnancy)+(1.167Gestational interval)+(1.186Systolic blood pressure in previous pregnancy)+(0.970Diastolic blood pressure in previous pregnancy).The slope of the calibration curve was close to one in the training set and verification set. The results showed that the prediction of recurrent PE risk of pregnant women with history of eclampsia was consistent with the actual risk. ROC analysis showed that the area under curve was 0.991. The results of DCA analysis showed that the model had good clinical practicability. In this study, a prediction model is successfully established and verified according to the influencing factors.