Search Results (37)

Prediction of the time to failure of pipeline materials supports inspection scheduling within risk-based inspection (RBI) frameworks. We compare three survival models, namely the Cox proportional hazards model (Cox PH), the random survival forest (RSF), and XGBoost with the survival:cox objective, on 823 corrosion incidents from the PHMSA database (1986–2025) retained after an engineering-validity screen. The comparison turns on a point that is easy to miss at the level of implementation. For the survival:cox objective, the predict() method of XGBoost returns the hazard ratio exp($\hat{\mathsf{\eta }}$), whereas the Breslow estimator of the baseline cumulative hazard expects the linear predictor $\hat{\mathsf{\eta }}$ and applies the exponential internally. Chaining the two evaluates exp(exp($\hat{\mathsf{\eta }}$)). Both maps are strictly increasing, so the concordance index is identical under either scale, to the last representable digit in our computation, and the defect does not appear in discrimination-based validation, while the estimated survival function is destroyed. A simulation with a known survival function isolates the magnitude: substituting the hazard ratio for the linear predictor raises the mean absolute deviation of Ŝ(t | x) from 0.073 to 0.398 and the integrated Brier score (IBS) from 0.154 to 0.395, at an unchanged concordance index of 0.723. On the pipeline corpus, the erroneous scale inflates the IBS to 0.356 ± 0.043 over 20 random splits. On the correct scale, XGBoost-Cox attains C = 0.871 ± 0.017 and IBS = 0.0614 ± 0.0056 and exceeds both reference models on both axes simultaneously (Cox PH 0.853 ± 0.018 and 0.0717 ± 0.0067; RSF 0.807 ± 0.020 and 0.0751 ± 0.0030; two-sided Wilcoxon p = 1.9 × 10⁻⁶ throughout). It also attains the lowest expected calibration error at every horizon from 10 to 40 years. No post hoc calibration is required, and adding one degrades the result. Out-of-time validation preserves the discrimination, whereas the absolute probabilities require periodic re-estimation of the baseline hazard. The feature-importance hierarchies agree with the ISO 8044 classification of corrosion factors. Full article

Background: Patients with stage I rectal cancer generally have favorable outcomes after curative surgery, but prognosis is not completely homogeneous. This study explored the prognostic association of preoperative inflammation–nutrition indices with overall survival and developed an interpretable internally validated machine learning survival model. Methods: We retrospectively included 475 patients with primary pathological stage I rectal adenocarcinoma who underwent curative-intent radical surgery at Sichuan University West China Hospital between 2018 and 2021. Patients downstaged to ypStage I after neoadjuvant therapy or treated by local transanal excision without lymph node dissection were excluded. Candidate predictors included age, sex, carcinoembryonic antigen, and routinely available preoperative inflammation–nutrition indices. LASSO-Cox regression was used for feature selection. Six survival models were developed and evaluated using 1000 bootstrap resamples with out-of-bag internal validation. Model performance was assessed at 60 months using time-dependent AUC, C-index, Brier score, calibration, and decision curve analysis. SHAP analysis was used for model interpretation. Results: During a median follow-up of 68 months, 30 deaths occurred. LASSO-Cox regression identified five predictors: age, lymphocyte-to-white blood cell ratio, fibrinogen-to-lymphocyte ratio, albumin-to-alkaline phosphatase ratio, and neutrophil-to-HDL cholesterol ratio. In bootstrap out-of-bag internal validation, XGBoost-Cox achieved the highest, although only marginally higher, discriminative performance among the evaluated models, with a 60-month time-dependent AUC of 0.783, a C-index of 0.774, and a Brier score of 0.0507. The calibration intercept and slope of XGBoost-Cox were 0.519 and 1.070, respectively. SHAP analysis identified age as the most influential predictor, followed by the selected inflammation–nutrition indices. Decision curve analysis suggested potential clinical utility within threshold probabilities from 1% to 20%, although this finding remains exploratory. Conclusions: Preoperative inflammation–nutrition indices may contribute to overall survival prognostic stratification in primary pathological stage I rectal cancer. External validation in larger multicenter cohorts is required before clinical application. Full article

Background/Objectives: Contemporary endometrial cancer pathology integrates histological and molecular features for risk classification. Endometrioid endometrial carcinoma (EEC) with a squamous cell carcinoma (SCC) component was historically termed adenosquamous carcinoma but later incorporated into the squamous-differentiation subtype of EEC. This reclassification left an evidence gap, and its prognostic significance remains insufficiently defined. We compared clinicopathological aggressiveness and evaluated associations of an SCC component with overall survival (OS) and lymph node (LN) involvement. Methods: We retrospectively analyzed 7590 surgically staged patients from 24 institutions (2000–2019): 7495 pure EEC and 95 EEC with SCC components. Confounding was addressed using overlap weighting (OW), inverse probability of treatment weighting targeting the average treatment effect on the treated (IPTW-ATT), and 1:4 propensity score matching (PSM), each with doubly robust Cox regression. Survival machine-learning models with SHapley Additive exPlanations (SHAP) assessed histological subtype contribution. LN involvement was assessed by OW/IPTW-ATT-weighted logistic regression. Results: EEC with SCC component had higher grade, deeper myometrial invasion, more advanced stage, and worse unadjusted OS (hazard ratio [HR] 7.04; 95% confidence interval [CI] 3.50–14.16; p < 0.001). After adjustment, the OS disadvantage persisted: OW-adjusted HR 3.05 (95% CI, 1.36–6.85; p = 0.007), IPTW-ATT-adjusted HR 3.26 (95% CI, 1.51–7.01; p = 0.003), and PSM-adjusted HR 3.30 (95% CI, 1.24–8.77; p = 0.017). XGBoost-Survival showed stable discrimination (test C-index 0.797 ± 0.057; 5-year AUC 0.787 ± 0.065), and SHAP supported histological subtype as a prominent risk-associated contributor after class-imbalance correction. In contrast, adjusted LN involvement was not increased (OW odds ratio [OR] 1.000, 95% CI 0.505–1.979; p = 1.000; IPTW-ATT OR 1.035, 95% CI 0.527–2.035; p = 0.920). Conclusions: EEC with SCC component was associated with worse OS after adjustment, without an adjusted increase in LN involvement. These findings support explicit morphologic reporting and future validation incorporating systematically recorded molecular testing when available. Full article

Introduction: Head and neck cancers are frequently diagnosed at advanced stages, resulting in poor survival despite therapeutic advances, highlighting the need for earlier identification within routine clinical care. Routinely collected electronic health record data provide a potential source of longitudinal clinical signals, although the clinical applicability of machine learning models remains limited by concerns regarding interpretability and real-world integration. Methods: This retrospective cohort study included 157,031 patients treated at the University of Debrecen Clinical Centre between 2007 and 2022, with head and neck cancer defined using ICD-10 codes C01 to C14. A total of 1397 variables, including demographic characteristics, ICD based comorbidities, and laboratory parameters, were reduced to 91 features using variance filtering and elastic net penalized Cox regression. Three survival modelling approaches were developed and compared, including CoxNet, Random Survival Forest, and XGBoost with a Cox objective. Results: The XGBoost model demonstrated the highest predictive performance with a mean concordance index of 0.916, followed by Random Survival Forest at 0.892 and CoxNet at 0.886, with acceptable calibration across models. Risk stratification showed clear separation between low, medium, and high-risk groups. Model interpretability using SHapley Additive exPlanations indicated that predictions were driven by a combination of demographic factors, laboratory markers, and clinically relevant diagnosis codes, reflecting both distal risk gradients and proximal clinical signals. Conclusions: These findings suggest that explainable machine learning applied to routine clinical data can support accurate and clinically interpretable risk stratification, with potential utility for opportunistic early identification of high-risk patients within existing healthcare pathways. Clinical Relevance: Explainable EHR-based survival models may support opportunistic identification of patients at increased head and neck cancer risk within routine clinical workflows, potentially improving triage and referral. Full article

Radiomics has the potential to improve risk stratification in head and neck squamous cell carcinoma (HNSCC), but clinical adoption is limited by inconsistent performance across institutions. A key source of variability is how radiomic features are generated, preprocessed, and selected prior to model development. This multicenter study evaluated how radiomics parameterization and feature selection strategies affect external model performance, feature stability, and time-to-event risk stratification. We studied pre-treatment CT scans from 752 patients with primary HNSCC from three hospitals. For each scan, 1648 radiomic features were computed using 20 different preparation methods that varied in scaling, outlier removal, and gray-level bin width. We compared five feature selection methods: Graph-FS with connected components, Boruta, Lasso, RFE-RF, and mRMR. The classification models used were Random Forest, XGBoost, CatBoost, and Logistic Regression. We measured performance using external ROC-AUC, bootstrap confidence intervals, Brier score, and RobustScore. Stability of feature selection was assessed using the Kuncheva and Jaccard indices. Cox proportional hazards models confirmed time-to-event results, and consensus SHAP analysis helped explain the models. Radiomics parameterization influenced model performance, and no single configuration was optimal across all analyses. Radiomics-only models outperformed clinical-only models, while clinical–radiomics models achieved the highest overall performance. mRMR and Lasso produced the highest average external AUCs, while Graph-FS showed the greatest stability. The best classification model achieved an external AUC of 0.817. In Cox validation, the best clinical–radiomics configuration achieved an external C-index of 0.662 and separated high- and low-risk patients in the external cohort. Full article

Biochemical recurrence (BCR) after radical prostatectomy (RP) remains a major clinical challenge. Although metabolic reprogramming drives prostate cancer (PCa) progression, its predictive value for BCR and its interplay with the tumor immune microenvironment (TIME) remain incompletely understood. By integrating weighted gene co-expression network analysis (WGCNA) with machine learning, we identified four metabolic-related hub genes (GDPD1, PLA2G7, PTGDS, and SRD5A2) and developed an XGBoost-Cox model that accurately stratified BCR risk (training 5-year AUC: 0.858; validation 5-year AUC: 0.745). SHAP analysis enhanced the model’s interpretability, while immunohistochemistry (IHC) validated differential protein expression of these targets across 32 clinical specimens. Furthermore, immune profiling demonstrated that these genes are closely linked to M2 macrophage-mediated immunosuppression and altered T-cell infiltration. To translate these biomarkers into therapeutic targets, we employed in silico screening, molecular docking, and molecular dynamics simulations, identifying (-)-epigallocatechin gallate (EGCG) as a promising multi-target candidate. Subsequent in vitro assays confirmed that EGCG binds stably to GDPD1, PTGDS, and SRD5A2, effectively suppressing malignant PCa phenotypes and prostate-specific antigen (PSA) secretion. In summary, we established a robust and interpretable model for predicting BCR after RP, and our in vitro validation suggests that EGCG holds promise as a therapeutic agent to delay PCa progression. Full article

Background/Objectives: Metabolic syndrome (MetS) affects more than 1.5 billion adults worldwide and is present in 37–70% of STEMI patients. Its ten-year prognostic value after primary PCI—particularly for heart failure, which is rarely examined as a primary endpoint—remains incompletely characterized. Methods: In total, 506 STEMI patients treated with primary PCI (December 2009–June 2010) were followed for ten years. MetS was defined at admission using AHA/NHLBI criteria. Co-primary endpoints were all-cause mortality, MACE, and hospitalization for heart failure. Multivariable Cox regression was adjusted for sex, age, LVEF, previous MI, Killip class, and multivessel disease. Four ML models were evaluated by 10-fold stratified cross-validation with SHAP-based feature, with a Fine–Gray subdistribution-hazard sensitivity analysis for heart failure. Feature attribution used TreeSHAP on XGBoost and permutation importance on a Random Survival Forest. Results: MetS(+) patients were older, more frequently female, and had higher SYNTAX scores (all p < 0.05). MetS was present in 216 patients (42.7%). It did not independently predict mortality (HR 1.09, p = 0.66) but did predict MACE (HR 1.47, p = 0.028) and heart failure hospitalization (cause-specific HR 2.86, 95% CI 1.57–5.22; Fine–Gray HR 2.61, 95% CI 1.44–4.75; both p ≤ 0.002). The null mortality finding coincided with differential statin discontinuation and a selective obesity paradox: in non-obese patients, MetS doubled mortality (42.9% vs. 21.1%, p = 0.008), while in obese patients, the effect disappeared (26.5% vs. 23.2%, p = 0.529). Two independent ML frameworks ranked the cumulative number of MetS criteria—rather than the binary diagnosis—among the leading individual-level features for heart failure prediction (Random Survival Forest c-index 0.843). Conclusions: In primary PCI-treated STEMI survivors, MetS independently predicts ten-year MACE and heart failure but not mortality. The number of MetS criteria at baseline, rather than the binary classification, was more strongly associated with heart failure risk; whether prospective modification of individual components reduces this risk requires dedicated interventional studies. The lean MetS-positive phenotype may represent a candidate subgroup warranting further investigation. Full article

Background: Peripheral artery disease (PAD) impacts more than 200 million individuals globally. Despite its prevalence, management remains suboptimal, partly due to the lack of reliable blood-based biomarkers. The ankle–brachial index (ABI), the current gold-standard test for PAD, is limited by inter-operator variability, misinterpretation, and reduced accuracy in patients with diabetes. Fatty acid binding protein 3 (FABP3) has emerged as a potential biomarker for PAD; however, its prognostic performance relative to ABI remains unclear. This study compared FABP3 and ABI for predicting PAD outcomes using statistical and machine learning approaches. Methods: A total of 1001 participants were prospectively recruited, including 644 patients with PAD and 357 without PAD. The primary outcome was 2-year major adverse limb event (MALE), defined as a composite of vascular intervention, major amputation, or acute limb ischemia. At enrollment, plasma FABP3 was quantified using a validated multiplex immunoassay. Kaplan–Meier analysis of MALE-free survival was performed across pre-specified FABP3 tertiles (high [>3.55 ng/mL], moderate [1.55–3.55 ng/mL], and low [<1.55 ng/mL]) and ABI tertiles (severe [<0.40], moderate [0.40–<0.70], and mild [0.70–0.90]), with curve separation assessed using log-rank tests. Multivariable Cox proportional hazards modelling was used to evaluate the independent relationships of FABP3 and ABI with 2-year MALE after adjustment for baseline demographic and clinical covariates. To assess predictive performance for 2-year MALE, an extreme gradient boosting (XGBoost) classification model incorporating 10-fold cross-validation was trained using a combination of clinical covariates, plasma FABP3 levels, and ABI. Discriminatory performance was assessed using the area under the receiver operating characteristic curve (AUC). Results: The average participant age was 68 years (SD 12), and 34% (n = 340) were women. Mean ABI was 0.75 ± 0.25 and mean FABP3 concentration was 2.97 ± 2.06 ng/mL. Among the 644 participants with PAD, 558 (86.6%) had complete time-to-event data for MALE status, FABP3, and ABI. Over the median follow-up period of 2 years, 140 (25.1%) participants with PAD experienced MALE. Kaplan–Meier analyses demonstrated significant separation in MALE-free survival across FABP3 tertiles (log-rank p < 0.001). At 24 months, MALE-free survival was 100.0% in the FABP3 < 1.55 group, compared with 71.1% in the FABP3 1.55–3.55 group and 67.7% in the FABP3 > 3.55 group. In contrast, ABI severity groups showed less pronounced separation, with 24-month MALE-free survival rates of 80.3% for mild ABI, 73.2% for moderate ABI, and 71.3% for severe ABI, without a statistically significant overall difference (p = 0.170). In adjusted Cox proportional hazards models, FABP3 demonstrated strong prognostic performance for 2-year MALE. A 1 SD increase in log-transformed FABP3 was independently associated with a higher risk of 2-year MALE (HR 1.90, 95% CI 1.60–2.25; p < 0.001), with minimal change after additional adjustment for ABI (HR 1.90, 95% CI 1.60–2.24; p < 0.001). Machine learning analyses similarly favored FABP3 over ABI, with the FABP3-based model achieving an AUC of 0.773 compared to 0.686 for the ABI-based model. Adding ABI to the FABP3 model did not improve discrimination. Conclusions: Circulating plasma levels of FABP3 are strongly associated with PAD outcomes. Specifically, FABP3 demonstrated a stronger and more robust association with 2-year MALE compared to ABI. This study validates the prognostic value of FABP3 for PAD outcomes in comparison to ABI. Full article

Background: Lung cancer remains one of the leading causes of cancer-related mortality worldwide. However, prognostic models developed within a specific population may not be accurate when applied to another population due to differences in demographics and clinical practices. In the present study, we investigated the cross-national applicability of machine learning (ML)-based survival prediction models trained on population data from the United States and validated on an independent Chinese clinical cohort. Methods: Cox proportional hazards, Random Survival Forest (RSF), and XGBoost-Cox models were developed and externally validated. Model discrimination was evaluated using the concordance index (C-index) and time-dependent AUC at 1, 3, and 5 years, along with calibration and decision curve analysis. Hyperparameter tuning was performed using cross-validation to reduce overfitting and improve model generalizability. Results: Three survival prediction models were developed using the U.S. SEER database (n = 13,260) and externally validated in an independent Chinese cohort (n = 505). Baseline characteristics differed between the cohorts, with the Chinese cohort being younger and having a higher proportion of stage IA disease. Despite these differences, all models demonstrated acceptable discrimination. The RSF model was the most stable across cohorts and time horizons, with a C-index of 0.740 (95% CI: 0.735–0.746) in SEER and 0.782 (95% CI: 0.720–0.844) in the Chinese cohort. RSF showed good calibration at 1 and 3 years but slightly overestimated 5-year mortality risk in the Chinese cohort. Conclusions: Machine learning-based survival prediction models, such as the Random Survival Forest model, are promising and robust tools for predicting cross-population survival in early-stage non-small cell lung cancer (NSCLC). However, differences in patient characteristics and treatment patterns may influence long-term model performance. These findings highlight the potential of flexible machine learning models in oncology and the essential role of rigorous external validation. Full article

Backgrounds/Objectives: Non-small cell lung cancer (NSCLC) exhibits substantial prognostic heterogeneity that is not fully captured by conventional anatomical staging, highlighting the need for individualized risk assessment. Radiomics enables non-invasive characterization of tumor phenotype, yet high dimensionality and inter-feature correlations often limit model stability and interpretability. Methods: To address these challenges, we developed a multimodal late-fusion framework integrating radiomic, clinical, and demographic information to predict patient-specific absolute risk in the Lung1 cohort (N = 398). Radomic features (N = 107) were extracted from primary tumor volumes and refined using a Group Lasso–penalized Cox model, preserving biological coherence and producing a parsimonious imaging signature. This signature was combined with clinical and demographic variables using five different late-fusion strategies: weighted averaging, Cox regression, logistic stacking, Random Survival Forests (RSF), and XGBoost. Model performance was evaluated using 5-fold cross-validation based on discrimination, calibration, and risk stratification metrics. Results: Using 5-fold cross validation, the radiomics-only model outperformed conventional clinical staging in patients’ risk prediction (C-index 0.5717 vs. 0.5350) and accuracy, demonstrating the prognostic value of imaging biomarkers. All fusion strategies improved risk prediction performance, with the Cox fusion model slightly better than other fusion methods with C-index of 0.58, time-dependent AUC of 0.60, and the distinct risk stratification with log-rank χ² of 22.85. Conclusions: These findings suggest that multimodal late fusion may provide robust and interpretable risk estimates with potential clinical relevance, supporting personalized risk prediction for informed decision-making in NSCLC. Full article

Background: Owing to significant clinical heterogeneity, the achievement of accurate survival forecasting for individuals with colorectal cancer and peritoneal metastasis continues to be a complex undertaking. We aimed to transcend traditional prognostic limitations by evaluating machine learning boosting models against standard regression-based methods in terms of estimating overall survival (OS). Methods: Utilizing a multi-institutional registry of 150 patients diagnosed with synchronous peritoneal metastasis of colorectal cancer, we integrated 124 clinicopathological variables to refine our predictive models. Beyond standard preprocessing—including standardization and median imputation—we rigorously compared XGBoost and LightGBM against Ridge, Lasso, and linear regression via five-fold cross-validation. To specifically address right-censoring, an XGBoost Cox model was implemented and validated using Harrell’s C-index, with SHAP and LIME providing essential model interpretability. Results: Boosting models consistently outperformed linear alternatives, which struggled with high error rates and negative R2 values. Specifically, XGBoost achieved an MAE of 475 ± 60 and an RMSE of 585 ± 88. The XGBoost Cox model reached a C-index of 0.64 ± 0.06. SHAP analysis highlighted inflammatory markers and peritoneal disease extent as the most influential prognostic drivers. Conclusions: While boosting models offer a clear accuracy advantage over linear methods, their prognostic power remains moderate. These findings underscore the potential of ensemble learning in oncology, yet mandate external validation before these tools can be integrated into clinical decision-making. Full article

Background/Objectives: Post-thrombotic syndrome (PTS) following unprovoked deep vein thrombosis (DVT) lacks readily available, calibrated risk estimates at defined follow-up horizons. Building on signals that thrombus burden, care processes, and a form of metabolic–inflammatory tone influence outcomes, we prospectively evaluated survival machine-learning models, explicitly including hyperuricemia while excluding what we consider major inflammatory confounders. Methods: Adults with first-episode unprovoked lower-extremity DVT were enrolled at two centers (July 2024–September 2025). PTS (Villalta) was assessed at 3, 6, 9, and 12 months. The cohort was split 70/30 into training and test sets. Eight learners (RSF, GBM, LASSO + Cox, CoxBoost, survivalsvm, XGBoost-Cox, superpc, and plsRcox) were tuned using 10-fold cross-validation in training and once evaluated in the independent test set. Performance metrics included all time-dependent AUCs, fixed-time ROC AUCs with bootstrap 95% CIs, C-index, various forms of calibration, decision-curve analysis, and simple Kaplan–Meier risk group separation. Results: 193 patients were analyzed (PTS in 64%). High 9-month AUCs were seen in training: GBM (0.992) and RSF (0.982) being the strongest; by 12 months, both remained near constant. Test set performance followed a similar pattern, with RSF again favored (AUC 0.948) and XGBoost/GBM close behind. Calibration was satisfactory, net benefit from decision curves positive, and to a large extent, risk groups were separated as expected. Conclusions: Survival machine-learning models, at least in this dual-center prospective cohort, produced a clinically useful risk of PTS. Hyperuricemia, or any metabolically based signal, is a valuable addition to the “anatomy and care” of DVT. External validation is still required. Full article

Concrete compressive strength is a fundamental indicator of the mechanical properties of High-Performance Concrete (HPC) with multiple components. Traditionally, it is measured through laboratory tests, which are time-consuming and resource-intensive. Therefore, this study develops a machine learning-based regression framework to predict compressive strength, aiming to reduce experimental costs and resource usage. Under three different data preprocessing strategies—raw data, standard score, and Box–Cox transformation—a selected set of high-performance ensemble models demonstrates excellent predictive capacity, with both the coefficient of determination (${\mathrm{R}}^2$) and explained variance score (EVS) exceeding 90% across all datasets, indicating high accuracy in compressive strength prediction. In particular, stacking ensemble (${\mathrm{R}}^2$-$0.920$, $\mathrm{EVS}$-$0.920$), XGBoost regression (${\mathrm{R}}^2$-$0.920$, $\mathrm{EVS}$-$0.920$), and HistGradientBoosting regression (${\mathrm{R}}^2$-$0.913$, $\mathrm{EVS}$-$0.914$) based on Box–Cox transformation data show strong generalization capability and stability. Additionally, tree-based and boosting methods demonstrate high effectiveness in capturing complex feature interactions. Furthermore, this study presents an analytical workflow that enhances feature interpretability through visualization techniques—including Partial Dependence Plots (PDP), Individual Conditional Expectation (ICE), and SHapley Additive exPlanations (SHAP). These methods clarify the contribution of each feature and quantify the direction and magnitude of its impact on predictions. Overall, this approach supports automated concrete quality control, optimized mixture proportioning, and more sustainable construction practices. Full article

Background/Objectives: Diffuse large B-cell lymphoma (DLBCL) is a biologically and clinically heterogeneous malignancy with variable outcomes. Accurate risk prediction at diagnosis remains essential to guide treatment and follow-up strategies. In this retrospective study we aimed to assess the performance of multiple modeling approaches to predict death by 26 months of follow-up in patients with DLBCL using data available in the diagnostic stage. Methods: In this study we included 412 patients with DLBCL who were evaluated, treated, and followed-up at the Regional Institute of Oncology in Iasi, Romania, between 2015 and 2023. Clinical and paraclinical data determined at baseline examination was used to train and test six machine learning models (logistic regression, random forest—RF, support vector machine with a radial-basis kernel—SVM-RBF, multilayer perceptron neural network—MLP, random survival forest—RSF, and extreme gradient boosting—XGBoost) and to compare their performance to the Cox proportional hazards model. Results: Among the models, RF achieved the highest discrimination (AUC = 0.9060), with balanced performance (accuracy = 0.833; F1 = 0.902), followed by XGBoost (AUC = 0.8335) and MLP (AUC = 0.7861; accuracy = 0.849). RF and logistic regression demonstrated the best calibration (Brier = 0.360 and 0.377). The Cox model achieved moderate discrimination (time-dependent AUC = 0.5561; C-index = 0.55). Conclusions: Our findings align with contemporary reports showing that machine learning frameworks can outperform classical prediction approaches. Full article

The health of electricity meters directly affects measurement accuracy and the interests of users. Traditional evaluation methods for electricity meters are limited by static error detection and manual calibration, and are unable to capture dynamic operating conditions or the complex influence of the power environment. To address this issue, this paper proposes an enhanced Cox proportional hazard (CoxPH) model based on Transformer for evaluating the health of electricity meters through a data-driven approach. This model integrates the data collected by the terminal (such as three-phase voltage, current, power, etc.) and operation and maintenance records. After data preprocessing, key covariates were extracted, including the average values of three-phase voltage and current fluctuations, current polarity reversal, and measurement error. The Transformer-based Cox proportional hazard (Trans CoxPH) model overcomes the linear assumption of the traditional CoxPH model by utilizing the self-attention and multi-head attention mechanisms of Transformer, and is able to capture the nonlinear relationships and time dependencies in time-series power data. Experimental results show that the performance of the Trans CoxPH model is superior to the traditional CoxPH model, temporal convolutional network-based Cox proportional hazard (TCN-CoxPH) model, extreme gradient boosting-based Cox proportional hazard (XGBoost CoxPH) model, and DeepSurvival long short-term memory (DeepSurvival LSTM) model. On the validation set, its concordance index (C-index) reaches 0.7827 with a Brier score of only 0.0501, significantly improving prediction accuracy and generalization ability. This model can effectively identify complex patterns and provides a reliable tool for the intelligent operation and maintenance of a power metering system. Full article