Search Results (1,884)

Intensive dairying emits greenhouse gases and degrades waterways through nutrient runoff and leaching. Trees and shrubs on paddock margins may slow nutrient movement by raising infiltration and by immobilising nutrients in soil carbon. We measured soil carbon along transects running from pasture into three clusters on a pasture-based, intensively managed dairy farm near Lincoln, New Zealand: a mixed New Zealand native planting of nine species (Site MN), a hybrid willow, Salix matsudana × S. alba (Site W), and a planting of Kunzea robusta (kānuka) and Leptospermum scoparium (mānuka) (Site KM). Each planting type occurs at one site, so the study compared three case-study plantings, rather than native and exotic trees in general. Within these plantings, we tested four hypotheses: soil organic carbon (SOC) stock is higher under trees than under pasture (H1); particulate organic carbon (POC) concentration is higher under trees than under pasture (H2); SOC peaks at the dripline (H3); and SOC is higher under exotic trees than under New Zealand natives (H4). Carbon stocks over 0–0.3 m were significantly greater 12 m inside the mixed native planting, at 114 t C ha⁻¹, than at the dripline and in the adjacent pasture, 95 t C ha⁻¹, an increase of 20%. Stocks under willow and under kānuka/mānuka did not differ significantly from pasture. POC varied significantly with location under all three plantings at both 0–0.1 m and 0.1–0.3 m, and mean POC under the trees exceeded that under the adjacent pasture at every site and depth. In the 0–0.1 m layer POC averaged 0.22 to 0.35 g per 100 g of soil under the trees compared to 0.06 to 0.11 g per 100 g of soil under pasture. No planting had its highest carbon at the dripline, and the exotic willow had no more carbon than the natives, so H3 and H4 were not supported in these plantings. Tree clusters on paddock boundaries and marginal land add POC to agricultural landscapes at no cost to soil carbon. Future work should determine whether that carbon immobilises mobile nitrogen and phosphorus. Full article

Jujube is an important fruit tree and one of the five major economic forest tree species native to China, possessing high nutritional and medicinal value. It plays a key supporting role in the efficient utilization of marginal land resources—such as mountainous, sandy, saline-alkali, and drought-prone areas—as well as in the revitalization of rural industries. Molecular marker technology, with its advantages of stability, accuracy, and efficiency, has become an important tool in jujube genetic breeding research and is widely applied. This article summarizes molecular markers based on three generations of technological intergenerational systems: the first generation of hybridization-based markers (Restriction Fragment Length Polymorphism, RFLP), the second generation of PCR-based markers (represented by Simple Sequence Repeat, SSR), and the third generation of high-throughput sequencing markers (Single Nucleotide Polymorphism, SNP, Genotyping-by-Sequencing, GBS, etc.). A comprehensive classification system based on technical principles, polymorphism sources, and other dimensions is constructed to clearly explain the driving forces and development trends of molecular marker system evolution in jujube tree research, and to clarify the selection criteria and adaptation strategies of molecular markers. Research has found that the second-generation molecular marker SSR is the fundamental core tool for standardized identification of jujube germplasm resources and genetic analysis of low-budget populations. The third-generation molecular markers represented by SNPs and InDel are high-density maps, Genome-Wide Association Study(GWAS), Genomic selection, and other modern precision breeding core carriers, forming a layered complementary technology system, and are gradually becoming the mainstream of research. This paper summarizes the application progress of molecular markers in the precise identification of jujube germplasm resources, analysis of genetic diversity, determination of genetic relationships, construction of high-density genetic maps, genome-wide association analysis, functional gene mapping, and tracing of domestication and evolution. This article integrates all existing research using the unified scientific framework of “technology defect driven tagging iteration”, analyzes the internal logic of the evolution and replacement of different tagging systems, the inherent limitations of early tagging, the existing problems in current research, and discusses future research directions, aiming to provide a reference for the scientific and efficient application of molecular markers in jujube and to promote the improvement and upgrading of the jujube molecular marker-assisted breeding technology system. Full article

Biochar amendment is widely used to enhance soil carbon storage and regulate greenhouse gas emissions, but field evidence from forest plantations remains limited. This 1.5-year field study investigated the responses of soil greenhouse gas (GHG) emissions and carbon storage to rice husk biochar applied at 15 t ha⁻¹ (LB) and 30 t ha⁻¹ (HB) in a subtropical poplar plantation. At 3.0 m from the tree stem, biochar did not significantly affect cumulative N₂O emissions but significantly increased cumulative CO₂ emissions and reduced CH₄ emissions by 49.36% (LB) and 21.80% (HB), resulting in no significant change in non-CO₂ GHG-based global warming potential. In contrast, biochar significantly enhanced measured soil organic carbon (SOC) storage in several soil layers, especially in the 0–20 cm layer, and increased mineral-associated organic carbon (MAOC) while decreasing the POC: MAOC ratio, indicating a greater allocation of SOC towards more recalcitrant fractions. However, when carbon storage was integrated over the 0–100 cm profile, a significant increase was observed only under the high biochar rate at D1. These findings suggest that biochar can enhance carbon storage in poplar plantation soils, although its GHG mitigation effect is limited and gas-specific. Additional measurements in biochar-amended plots showed higher N₂O emissions closer to the tree stem (1.5 m), suggesting that future studies should use full-factorial treatment × distance designs to evaluate tree-proximity effects on biochar responses in plantation soils. Full article

Forest landscapes used for recreation also provide climate-regulating services that are often absent from spatial planning. We mapped aboveground biomass density (AGBD) and aboveground carbon storage in the Postavaru-Piatra Mare mountain landscape near Brașov, Romania, using a cross-platform remote-sensing workflow. MLS point clouds from 16 plots served as an independent plot-level reference, UAV LiDAR acquired at 70, 120, and 150 m provided airborne comparisons, and GEDI L4A samples, Sentinel-2 predictors, and terrain variables were modelled in Google Earth Engine using Random Forest (RF) and Gradient Boosted Trees (GBT). GBT achieved R² = 0.34 and RMSE = 106.42 Mg ha⁻¹, whereas RF achieved R² = 0.33 and RMSE = 107.45 Mg ha⁻¹. These values indicate limited explanatory power; accordingly, the resulting 25 m maps are interpreted as landscape-scale screening products rather than operational stand inventories. Against MLS, RF produced an RMSE of 79.05 Mg ha⁻¹ and GBT 82.58 Mg ha⁻¹, while UAV estimates were positively biased at all tested flight heights. Mean carbon density peaked at 1200–1400 m and was highest in coniferous forest, whereas broadleaved and tree-cover classes contributed the largest total stock because of their larger area. The RF and GBT estimates differed by 2.3% in aggregated study-area carbon stock. Aboveground carbon storage was the only ecosystem service quantified. The resulting maps should be interpreted as landscape-scale screening layers and as ecological inputs for future studies integrating trail, visitor-use, participatory-mapping or perception data; they do not quantify cultural ecosystem services. Full article

With the escalating frequency of sophisticated zero-day attacks, overcoming the critical limitations of signature-based Intrusion Detection Systems (IDSs) has become paramount. This study proposes a hybrid multi-layered intrusion detection framework combining traditional machine learning, Deep Neural Architectures (DenseNN), and ensemble methods to evaluate zero-day resilience within cloud-level backend connectivity interfacing EV and V2X management ecosystems. Using the comprehensive CSE-CIC-IDS2018 benchmark as a surrogate environment, a code-executed Leave-One-Attack-Out (LOAO) cross-validation protocol across 13 distinct attack families was implemented to assess unseen-attack-family generalization within the benchmark to unseen threats. Furthermore, Explainable Artificial Intelligence (XAI) auditing, utilizing SHapley Additive exPlanations (SHAP) and Integrated Gradients, was integrated to inspect decision boundaries and resolve feature-attribution failure modes. Critically, the audit identified an artifact-driven data leakage caused by the Timestamp and identifier features, demonstrating that models learned temporal schedules rather than behavioral network signatures. Re-executing all experiments post-leakage removal quantified performance drops across all classifiers (e.g., Gaussian NB dropping by up to 20.88 percentage points in accuracy (at the 60% training ratio; 18.30 points at the 80% ratio)). Under standard binary classification metrics, tree ensembles (Random Forest and Extra Trees) achieved high in-distribution detection (F1 > 0.95) with rapid inference latency (≈0.05–−0.07 ms/sample). However, the rigorous LOAO evaluation revealed a substantial generalization penalty on truly unseen zero-day families (e.g., SQL Injection and Infiltration), where simpler linear models demonstrated broader generalization robustness (mean LOAO F1 = 0.397) compared with complex tree-ensemble models. By rectifying dataset leakage and benchmarking deployment trade-offs (training runtime, throughput, and memory footprint), this study delivers actionable, transparent guidelines for deployment-oriented IDS evaluation in dynamic network infrastructures. Full article

Discrete-Variable Quantum Key Distribution (DV-QKD) provides a mechanism for establishing secret keys between legitimate parties using quantum-state transmission and authenticated classical post-processing. In this work, the underlying quantum layer follows a biased-basis decoy-state BB84 model using phase-randomized weak coherent pulses, while additional implementation-level mechanisms are integrated to support Cyber-Physical System (CPS) communication. Exponential Double Delta Encoding-based Merkle Tree (EDDE-MT) is employed as a receiver-side detection-record integrity mechanism for detecting deletion, insertion, reordering, or modification of records relative to an authenticated committed detection-event batch. It does not establish the completeness of the original TCSPC acquisition, detect records omitted before commitment, detect physical photon loss, or increase the information-theoretic secrecy of the QKD key. Time-Correlated Single Photon Counting (TCSPC) is used for detection-event and timing acquisition, while 2’s Complement Cyclic Redundancy Check-based Low-Density Parity Check (2CCRC-LDPC) supports error reconciliation. Following privacy amplification, the legitimate parties retain matching copies of the distilled QKD key locally. Discrete Variable Quantum Mellin Transform Cryptography (DVQMTC) uses fresh, non-reused segments of this privacy-amplified key for application-layer payload protection; the Mellin-transform component is treated only as implementation-level preprocessing and not as a cryptographic key-generation mechanism. Side-channel monitoring is performed using Gini Cramer’s V Correlation-Stationary Wavelet Transform (GCVC-SWT), Helical Valley-Principal Component Analysis (HV-PCA), and an Entmax-based hyperbolic Tangent exponential Linear Unit-Bidirectional Long Short-Term Memory (E-TeLU-Bi-LSTM) classifier. On the AES-HD benchmark, E-TeLU-Bi-LSTM achieved 99.24% classification accuracy; this value represents benchmark-level classification performance and is not interpreted as experimental validation of physical side-channel protection in a deployed DV-QKD system. Frequency Division Multiple Access (FDMA) and the Halton Quasi-Sequence-Invasive Weed Optimization Algorithm (HQS-IWOA) are further incorporated as classical network-resource segmentation and load-management mechanisms and do not modify the composable QKD security bound. The contribution of the work is therefore positioned as a system-level engineering integration of QKD key establishment, detection-record integrity, reconciliation, application-layer data protection, side-channel monitoring, and network-resource management for CPS. The information-theoretic secrecy claim remains restricted to the underlying finite-key decoy-state BB84 procedure under the stated security assumptions; no new QKD security theorem, formally new cryptographic primitive, or experimentally validated physical quantum communication capability is claimed. Full article

Transformer lifecycle data across organizations are typically vertically partitioned among material suppliers, manufacturers, logistics service providers, testing agencies, and operation and maintenance units. This study presents a reproducible multi-layer vertical federated learning (VFL) prototype that integrates salted-hash identifier matching, additive sharing of local score vectors over finite fields, and a local public key infrastructure with a signature-based audit verification mechanism. A deterministic synthetic dataset is first constructed, comprising 5200 aligned records and 31 predictor variables, which are partitioned among five participants with varying numbers of features per participant. Second, across five validation runs, the VFL models under both the standard block-wise and score-sharing paths achieved an AUC of 0.8825 ± 0.0119, an F1 score of 0.7367 ± 0.0249, and an accuracy of 0.8102 ± 0.0183 on the test set. The classification results of both paths were fully consistent with the centralized gradient-descent logistic regression baseline. Notably, the score-sharing path exhibited a maximum log-odds deviation of only 2.22 × 10⁻⁸ on the test set, with no prediction discrepancies observed. Third, across 10 independently generated synthetic populations, the nonlinear output mechanism highlights the limitations of linear models: the AUC of vertical federated learning (VFL) drops to 0.6457 ± 0.0171, while Extra Trees and HistGradientBoosting achieve 0.7731 ± 0.0149 and 0.7743 ± 0.0139, respectively. Finally, in a separate residual-sharing diagnostic test, when 1 to 4 participants jointly shared the residuals, the label inference AUC remained around 0.499–0.500; however, when all five participants shared or plaintext residuals were used, the labels could be fully recovered. Both simple membership inference diagnostic tests yielded results close to random. The local signature log verifier rejected all 700 injected faults and accepted the 400 clean control log events. These results validate the feasibility of numerical reproducibility and local audit functionality under synthetic data and single-process conditions, yet they are insufficient to demonstrate end-to-end label privacy protection, malicious security, effectiveness on real data, or real-time ledger performance. Full article

Passive Ultra-High Frequency (UHF) Radio Frequency Identification (RFID) systems deployed in metallic commercial aircraft cabins suffer from severe multipath fading, non-stationary channel dynamics, and operator gait-induced signal jitter. Addressing these challenges without physical airframe modifications or regulatory recertification remains a critical operational bottleneck. This paper presents an edge-native, software-defined framework that integrates micro-electromechanical system (MEMS) inertial measurements with an IMU-assisted Adaptive Kalman Filter (AKF) and a distilled surrogate decision tree. The proposed algorithm extracts localized motion energy ($E_{\mathrm{IMU}}$) to dynamically scale the measurement noise covariance ($R_k$) prior to physical-layer signal corruption, thereby eliminating phase lag and power hunting. For deterministic edge execution on COTS handheld devices, surrogate model distillation compresses a parent Random Forest ensemble into an $8.2\mathrm{KB}$ 13-leaf decision tree (depth 5) yielding $0.12\mathrm{ms}$ inference latency. Empirical validation across 17 operational sessions in Airbus A320, Boeing 737, and Airbus A321 cabins (10,720 valid reads) demonstrates a $99.45\%$ mean RSSI jitter reduction ($95\%\mathrm{CI}:[99.21\%,99.63\%]$) and a $7.30\times$ suppression of transmit power oscillations. Statistically, asset detection completeness is fully preserved ($0.791$ vs. $0.795$ baseline, $z=0.281,p=0.779$). Operating entirely within standard handheld software runtimes, this approach bypasses Supplemental Type Certificate (STC) requirements while ensuring robust aerospace asset visibility. Full article

Accurate indoor localization is a fundamental enabling technology for modern smart environments and Location-Based Services (LBS). Among the various indoor positioning technologies, Bluetooth Low Energy (BLE) has emerged as a popular solution due to its low cost and low power consumption. However, BLE signals suffer from severe environmental noise, while continuously executing complex positioning models can quickly deplete the battery resources of mobile devices. To address both problems jointly, this paper proposes a robust and sustainable machine learning framework for indoor localization on mobile devices. The framework first applies a discrete-time Kalman Filter that suppresses noise induced by walls and moving human bodies, and then benchmarks 12 machine learning models (including a Neural Network baseline) on a real-world dataset of 15,000 samples from 10 smartphones under a $5\times 3$ repeated cross-validation (CV) protocol. To identify the best model for mobile deployment, we introduce the Green Efficiency Index (GEI), which balances positioning accuracy against software-estimated energy consumption in Joules. Results show that the evaluated Multi-Layer Perceptron (MLP) baseline struggles with the noisy BLE data, producing a Mean Absolute Error (MAE) of 1.81 m, whereas the evaluated tree-based models map indoor spatial patterns far more accurately. K-Nearest Neighbors (KNN) achieves the lowest MAE at 1.175 m but requires substantial memory, making it unsuitable for sustainable mobile deployment. Extra Trees therefore emerges as the optimal solution, achieving an MAE of 1.347 m with low energy consumption and a compact memory footprint. The framework also normalizes hardware differences across all 10 tested smartphones, providing an accurate, energy-efficient, and generalizable solution for indoor localization. Full article

The environmental and economic impacts of nitrate leaching from agricultural soils highlight the need for accurate prediction to support effective nitrogen management. This study evaluated six machine learning (ML) models—Multiple Linear Regression, Elastic Net, K-Nearest Neighbors, Decision Trees, Extra Trees, and Multi-Layer Perceptron—using 2993 field observations to predict nitrate leaching. The models incorporated crop sequence, manure and fertilizer inputs, soil, and percolation and are benchmarked against the empirical NLES5 model used in Danish nitrogen regulation. Four ML models achieved higher predictive accuracy than NLES5 when evaluated on independent test data, with the Extra Trees model achieving the best performance (R² = 0.63, RMSE = 23.3 kg N ha⁻¹), exceeding NLES5 (R² = 0.39, RMSE = 29.8 kg N ha⁻¹). Model interpretability analyses identified winter percolation, winter vegetation cover, and soil type as key drivers of nitrate leaching. The Extra Trees model was further evaluated using scenario analyses of long-term leaching trends, marginal responses to spring-applied mineral nitrogen, and spatial patterns within the Bolbro Bæk catchment. Findings highlight the potential of ML models to improve nitrate leaching predictions in ungauged areas. Future research could incorporate additional variables, including crop yield and tillage practices, to enhance model accuracy and support sustainable agricultural practices that maintain productivity while reducing nitrate leaching. Full article

Blockchain-based electronic voting has yet to reach the electoral environments that could benefit from it most. Existing systems are designed for well-connected, grid-powered urban settings and fail precisely where digital voting could widen participation the most; this gap motivates a framework engineered from the ground up for low-infrastructure conditions rather than one adapted to them. Across much of Africa, four deployment barriers stand in the way: intermittent connectivity that excludes an estimated 43% of the population, identity documents that remain valid for up to ten years and degrade conventional face recognition from 99% to below 80%, power demands of 200–250 W that rule out battery-powered operation, and cryptographic designs locked to a single algorithm regardless of operating context. This paper presents AdaptVote, a five-layer framework built on the premise that these barriers must be removed jointly rather than in isolation. At its core, the AI-FOLM algorithm anchors INT8-quantised ArcFace embeddings to age-stable craniofacial ratios on a Xilinx Zynq-7020 FPGA, reaching a 96.7% True Accept Rate at 0.1% False Accept Rate over 6–10 year age gaps—2.2 percentage points beyond the state of the art. A nullifier-based protocol allows ballots to be cast entirely offline with Merkle-tree integrity, an adaptive ECDSA/EdDSA/BLS layer cuts Ethereum settlement costs by 60–86%, and the complete polling station runs at 4.2 W peak power and 28 Wh per 12 h session, a 99.3% energy reduction over GPU baselines. A twelve-hour field deployment with 300 voters at Mohammed First University, Nador, Morocco, conducted under 35% network availability, recorded a 95.7% first-attempt authentication rate (95% CI: 92.7–97.5%), 99.2% system uptime, and a 53.3% reduction in average voting time relative to paper ballots ($p<0.001$). Together, these results suggest that trustworthy electronic voting can be engineered for, rather than merely adapted to, low-infrastructure electoral settings. Full article

This research presents an integrated techno-energetic framework to decouple electric vehicle (EV) fleet growth from urban grid instability, using Bogotá, Colombia, as a case study. As emerging megacities confront rising charging demands, conventional reactive grid reinforcements are becoming technically and economically constrained. To address this, we develop a multi-layered optimization model that transforms urban Voronoi polygons into operational energy catchment units. Utilizing a Monte Carlo Tree Search (MCTS) algorithm, the framework determines infrastructure deployment sequences under two operational thresholds: a target high-resilience EV-to-charger ratio and a conservative scenario. Each localized node is technically dimensioned as a Representative Charging Station (RCS) equipped with a photovoltaic array and a Battery Energy Storage System (BESS). The results reveal spatial heterogeneity; high-density polygons in specific commercial and residential districts exhibit elevated infrastructure utilization alongside a stable solar resource baseline. Furthermore, the model demonstrates that this distributed architecture alleviates transformer thermal stress during the peak nocturnal charging period, mitigating localized overload risks and supporting distribution grid operational reliability. This study provides a scalable decision-making tool that bridges geospatial urban planning with renewable energy engineering to support the transition of constrained electrical networks. Full article

Urban London plane (Platanus × acerifolia) street trees require efficient and interpretable methods for identifying individuals that warrant closer inspection within large roadside inventories. This study developed a two-layer screening framework for 2022 London plane trees along six urban roads in Nanjing, China. First, wind-tunnel tests of five standardized crown geometries were used to derive geometry-specific drag-moment coefficients and calculate aerodynamic crown-loading potential (ACLP), a relative indicator of crown-related aerodynamic demand. ACLP was calculated only for trees within the tested crown height-to-width range. Second, measurements from 19 windthrown London plane trees after Typhoon Bebinca were used to derive a local lower-quartile root-plate reference and calculate a root-plate reference aperture ratio (RPAR) from recorded tree-pit width. Of the 2022 inventoried trees, 1273 (63.0%) fell within the tested crown-geometry range, and 128 were identified as high-ACLP candidates using a 90th-percentile screening threshold. These candidates were strongly concentrated on two roads, and the principal road-level pattern remained stable across alternative percentile thresholds. The post-storm analysis provided a separate local reference for identifying tree-pit apertures that were comparatively small relative to observed root-plate dimensions. ACLP and RPAR should therefore be interpreted as distinct first-stage screening indicators rather than as predictors of windthrow probability or anchorage capacity. The framework provides a transparent basis for prioritizing follow-up field inspection and allocating limited urban tree-management resources. Full article

Bio-based phase change materials (PCMs) can increase transient heat storage in lightweight building envelopes, but their performance depends on the climate, transition properties, layer design, and assumptions used to translate thermal loads into carbon and cost indicators. Although PCM optimization, machine learning surrogates, and lifecycle assessment have each been studied extensively, comparatively few studies combine them while explicitly separating simulation-derived thermal outputs from scenario-dependent environmental and economic post-processing and benchmarking bio-based candidates against paraffin on a common wall area basis. This study develops a simulation-based screening framework for a south-facing office wall model using 18,000 EnergyPlus cases, climate-specific machine learning surrogates, TreeSHAP interpretation, NSGA-II optimization, and scenario-based lifecycle carbon and cost accounting. XGBoost achieved pooled held-out R² values of 0.974 for occupied discomfort degree-hours and 0.978 for total annual thermal demand. For Palermo, the directly re-simulated balanced configuration (T_m = 24.8 °C, L_h = 178 kJ/kg, 22 mm thickness, intermediate position) reduced occupant discomfort by 42.6% and the modeled single-zone total thermal demand by 6.0%. Under the central all-electric scenario (SCOP = SEER = 3.0, grid factor = 0.233 kg CO₂eq/kWh, 25 years), scenario-based net lifecycle carbon was −22.1 kg CO₂eq/m² for the analyzed south wall with an 8.0-year environmental payback, compared with −7.4 kg CO₂eq/m² and 19.2 years for RT28 paraffin. Energy savings did not recover the additional investment; the incremental lifecycle cost was +24.1 EUR/m². The theoretical contribution is a transparent, climate-dependent screening logic that couples surrogate interpretation with explicit evidence boundaries; the applied outcome is a palmitic–capric target-property region prioritized for laboratory validation rather than a deployment-ready product. Only directly re-simulated configurations are used for quantitative applied thermal claims; surrogate-only Pareto points are retained as exploratory screening candidates and are not interpreted as validated optima. The numerical results are specific to the modeled south-wall, single-zone boundary; whole-building and cross-regional application requires local recalibration and direct validation. By linking passive comfort, carbon accounting, material innovation, and responsible pre-experimental selection, the workflow is relevant to the decarbonization objectives represented by SDGs 7, 9, 11, 12, and 13. Full article

Background/Objectives: Radiation-induced hypothyroidism (RIHT) is a frequent complication in head and neck cancer (HNC) patients following radiotherapy, significantly impacting quality of life. This study aimed to develop and validate an interpretable machine learning model for predicting RIHT, with a focus on providing individualized risk assessment. Methods: The study included a development cohort (n = 256) from the Second Affiliated Hospital of Harbin Medical University and an independent external validation cohort (n = 296) from the First Affiliated Hospital of Kunming Medical University. Using 18 predictive features, six machine learning algorithms—Random Forest, Support Vector Machine, Multi-Layer Perceptron, Logistic Regression, XGBoost, and LightGBM—were constructed. Model performance was evaluated using AUC, balanced accuracy, F1 score, PR-AUC, Brier score, and Brier skill score. SHapley Additive exPlanations (SHAP) was applied for model interpretability. Results: In internal validation, LightGBM achieved the highest performance (AUC = 0.872, balanced accuracy = 0.703, PR-AUC = 0.712, and Brier skill score = 0.371), outperforming logistic regression. In external validation, Random Forest achieved the highest AUC (0.769), while XGBoost and LightGBM achieved AUCs of 0.721 and 0.716, respectively. SHAP analysis identified pre-treatment thyroid volume, T-stage, tumor site, mean thyroid dose (Dmean), and V50 as the most influential features for model prediction. Notably, dose–volume parameters (V30–V60) were highly correlated (r = 0.75–0.98); therefore, they should be interpreted as a cluster rather than as independent predictors. We caution against overinterpreting individual Vx thresholds as rigid clinical cutoffs. Conclusions: Machine learning models, particularly tree-based ensemble methods (LightGBM and XGBoost), demonstrate good predictive performance for RIHT. Combined with SHAP analysis, these models provide transparent, individualized risk assessments. However, the dose–response relationship for RIHT is continuous, and no single dose threshold should be used as a rigid clinical cutoff. Model code is available from the corresponding author upon reasonable request. Future prospective studies with time-to-event analysis are needed to further validate clinical utility. Full article