Data were obtained from the University of Florida Health Integrated Data Repository, which houses its clinical content in the Observational Medical Outcomes Partnership Common Data Model. The structured dataset included demographics, diagnoses, medications, procedures, and laboratory results, while unstructured data comprised CT radiology report narratives. We retrospectively identified over 60,000 patients with renal-related conditions observed from December 2011 to August 2024. The study cohort included patients who met all of the following criteria: (1) had at least 2 distinct renal tumor diagnoses, recorded at different time points, based on ICD (International Classification of Diseases) codes, with the earliest diagnosis designated as the index date; (2) experienced at most 1 change in tumor classification (benign to malignant or malignant to benign), with the final status serving as the outcome label (patients with multiple reversals were excluded); and (3) had at least 1 renal CT report dated before the index date. A schematic illustrating the temporal ordering of imaging, diagnosis, outcome labeling, and surgical intervention is provided in Figure S9 in Multimedia Appendix 1.

Benign vs malignant tumor status was derived from longitudinal EHR diagnosis codes, as pathology reports were not uniformly available. Patients were classified as benign or malignant if they had at least 2 corresponding ICD diagnosis codes recorded on distinct dates, with the final observed status used as the outcome label. To reduce label instability, patients with multiple benign-malignant reversals were excluded, and a 1-direction transition rule was applied.

To assess the robustness of the ICD-based outcome definition, sensitivity analyses were conducted in patient subsets with higher-specificity reference signals. These included (1) a pathology-confirmed subset identified from unstructured surgical pathology reports and (2) a nephrectomy subset, in which renal tumor diagnosis codes recorded after surgery were used as proxy reference labels. These analyses were used to evaluate the consistency of outcome labeling under more stringent definitions.

We integrated structured EHR data with CT radiology report analyses to generate three feature sets for modeling: (1) structured EHR variables; (2) kidney-specific findings extracted from radiology reports; and (3) tumor characteristics derived from these findings. The detailed workflow for feature extraction from radiology reports is shown in Figure 1B.

Structured clinical variables were retrieved from tables containing patient demographics, diagnoses, procedures, medications, vital signs, and laboratory results. The observation window extended from the earliest available record up to the index date. Demographic variables (eg, age, sex, and race/ethnicity) were categorized, diagnoses were mapped using PheWAS (Phenome-Wide Association Study) Phecode groupings [36], and medications were aggregated at the ingredient level using Anatomical Therapeutic Chemical classification codes [37]. For vital signs and laboratory tests, the most recent values preceding the index date were selected. For blood pressure, the highest and lowest systolic and diastolic values recorded on the most recent measurement day were used as continuous summary features. Dialysis procedures were encoded as binary indicators. Categorical features were one-hot encoded, and continuous variables were standardized prior to modeling.

To isolate kidney-specific findings, the most recent radiology report within the observation window was selected. Given that abdominal CT reports often include findings for multiple organs, the LLM Qwen2.5-32B [38] was then used to isolate the paragraph exclusively detailing kidney-specific findings, yielding a consolidated narrative for downstream analysis. The model was hosted on premises to ensure patient confidentiality and maintain data integrity.

Abnormality characteristics were extracted from the kidney-specific narratives using an LLM to capture detailed lesion descriptors absent from structured data. Extracted features included abnormality presence, type (eg, cyst, mass, or tumor), lesion size, anatomical position (eg, left kidney or right kidney), exophytic status, CT attenuation, and contrast enhancement. For patients with multiple preindex renal CT examinations, tumor growth rate was calculated using only lesion size measurements obtained prior to the index date, defined as the earliest recorded renal tumor diagnosis. Specifically, when at least 2 preindex imaging studies were available, growth rate was computed as the slope of lesion size change (cm per day) between the earliest and latest available preindex imaging studies.

Missing values in structured EHR variables were handled using modality-specific strategies. Before imputation, laboratory variables with greater than 50% missingness were excluded from the structured clinical feature set. Continuous features (ie, laboratory measurements and vital signs) were imputed using mean values calculated from the full cohort, and categorical variables (ie, diagnosis and medication records) were processed via standard presence-absence encoding, with missing entries encoded as 0 (eg, absence of a diagnosis or medication record). For radiology-derived abnormality characteristics, categorical attributes were treated as “unknown” when not explicitly mentioned and encoded as a separate category. Continuous abnormality variables (eg, lesion size and tumor growth rate) were encoded using a 2-part sentinel representation. For cases where size information was unavailable, or where fewer than 2 preindex longitudinal imaging studies were present to compute a growth rate, the numeric value was imputed with a placeholder of zero. This value was paired with a corresponding binary missingness indicator to ensure the model could mathematically distinguish between missing measurements and true physiologic stability (eg, a measured growth rate of zero). Both the numeric placeholder and the binary indicator were included as independent features in the model input. Kidney-specific text findings were encoded using pretrained biomedical transformer models, producing fixed-length embeddings of 768 dimensions. More details can be found in Table S7 in Multimedia Appendix 1.

To evaluate the performance of LLMs in extracting these abnormality characteristics, we manually annotated 500 independent radiology reports to establish a gold-standard dataset. The gold-standard annotations were created by 2 independent nonexpert annotators with biomedical research training, following detailed annotation guidelines developed in consultation with board-certified radiologists. Annotators independently labeled each of the 500 radiology reports, blinded to model outputs. Interrater reliability was assessed using Cohen κ for each abnormality attribute, with disagreements resolved through consensus discussion guided by a radiologist.

For automated extraction, the pipeline identifies kidney-related sentences, extracts lesions with their attributes as structured entries, validates outputs against a predefined schema, and ranks lesions by severity, prioritizing tumors over masses and cysts, and ordering within categories by size. Four locally deployed LLMs (Qwen2.5-7B [38], Qwen2.5-32B [38], LLaMA3-8B [39], and LLaMA3-70B [39]) were evaluated in inference-only mode under institutional data privacy constraints, without any task-specific model training.

To quantify extraction performance, 2 components of the pipeline were assessed. Kidney-specific paragraph retrieval was evaluated by comparing automatically retrieved paragraphs with manually annotated reference paragraphs using the Bilingual Evaluation Understudy-4 (BLEU-4) metric [40], which measures n-gram overlap between predicted and gold-standard text. Performance was summarized as the average of the BLEU-4 score and the percentage of samples achieving a BLEU-4 score greater than 0.5 across reports. A BLEU-4 score above 0.5 indicates good alignment between predicted and reference text in terms of vocabulary, phrase structure, and semantic content, and was considered acceptable for practical downstream use. Abnormality characteristic extraction was evaluated by computing per-attribute accuracy based on an exact match between model outputs and gold-standard annotations. All performance metrics were calculated on the same set of 500 annotated radiology reports.

Model performance was assessed using a stratified 80/20 train-test split. The 80% training partition was used for model development with 5-fold cross-validation. In each iteration, 1 fold was used for internal validation, and the remaining 4 folds were used for training. The resulting 5-fold–specific models were each evaluated on the same independent 20% held-out test set, which was not used for training, validation, resampling, or hyperparameter tuning. For the machine learning classifiers, hyperparameters were tuned via grid search, with model selection guided by the minimization of negative log-loss on the validation folds. In parallel, deep neural networks were trained using the AdamW optimizer and a linear learning rate scheduler with an initial warm-up phase, with cross-entropy loss serving as the selection criterion on validation subsets.

The primary evaluation metric was the AUC. Secondary metrics included overall accuracy, balanced accuracy, sensitivity (recall for the malignant class), specificity, precision (positive predictive value), and F₁-score, all reported as the mean and SD across 5-fold–specific models evaluated on the independent held-out test set. Balanced accuracy was included to provide a more representative assessment of the model’s discriminative performance beyond class prevalence by equally weighting sensitivity and specificity in the presence of class imbalance. Decision thresholds for probability-based models were determined by maximizing the Youden J statistic. To address class imbalance, we compared 3 approaches: no resampling, random undersampling, and random oversampling. To prevent data leakage, all resampling procedures were performed strictly within the training partition of each cross-validation fold; validation folds and the held-out test set were preserved in their original class distributions.

We assessed the contribution of each modality to the best-performing multimodal fusion model using 2 complementary approaches.

The study has been approved, and the requirement to obtain any informed consent has been waived by the University of Florida Institutional Review Board (protocol number IRB202100401). The research does not involve greater than minimal risk for participation. Analyses only involve the secondary analysis of data that are either limited data sets or deidentified. Our research team has no direct contact with human subjects. All methods were carried out in accordance with relevant guidelines and regulations.

A total of 967 patients with renal tumors met the inclusion criteria. The mean age at initial diagnosis, whether benign or malignant, was 69 (SD 12.93) years, with approximately 60% of the cohort being male. Among these patients, 712 (73.6%) were diagnosed with malignant tumors. Comorbidities were common: over 90% (874/967) had disorders of the kidney and ureters, and 74% (719/967) had a documented history of hypertension. Smoking was reported in 15.51% (150/967) of patients, and the mean BMI was 30.07 (SD 7.08). Detailed baseline characteristics and comorbidity profiles are summarized in Table S1 in Multimedia Appendix 1. The median observation period was 34.43 (IQR 8.53-73.92) months. Surgical intervention was performed in 401 patients (41.47% of the cohort). Among these, 294 (30.40%) patients underwent partial nephrectomy and 250 (25.85%) patients underwent radical nephrectomy; 143 (14.79%) patients received both procedures. Tumor growth rate was available for 22.2% (215/967) of patients, reflecting those with at least 2 preindex surveillance CT examinations; the remaining cases were treated as missing for this feature.

Surgical pathology reports containing renal tumor–related diagnostic statements were available for 89 patients (from 97 reports), of whom 81 (91.0%) were classified as malignant based on pathology text. In this pathology-confirmed subset, the primary ICD-based outcome labels were fully concordant with pathology-derived malignancy status. In addition, 401 patients underwent nephrectomy, and among the 383 with postprocedure renal tumor diagnosis codes, 347 (90.6%) were malignant; outcome labels derived from ICD codes were fully concordant with postnephrectomy diagnostic coding.

Interrater reliability for the manually annotated gold-standard dataset was substantial to near-perfect across abnormality attributes (Cohen κ=0.95-1.00) (Table S10 in Multimedia Appendix 1). Table 1 summarizes the performance of LLMs in kidney-specific paragraph retrieval and abnormality characteristic extraction from radiology reports. As shown in Table 1, Qwen2.5-32B and LLaMA3-70B demonstrated comparable performance in abnormality characteristic extraction. For kidney-specific paragraph retrieval, both models achieved high BLEU-4 scores, with more than 95% (475/500) of reports exceeding a BLEU-4 threshold of 0.5, indicating accurate identification of the relevant anatomical sections. However, accuracy alone does not adequately reflect model usability. Qwen2.5-7B and LLaMA3-8B failed to extract 10.4% (52/500) and 18.2% (91/500) cases, respectively, despite repeated attempts. Additionally, the standard LLaMA3 release was limited by its 8192-token context window, which hindered few-shot prompting with multiple examples. In contrast, Qwen2.5-32B demonstrated more reliable kidney-specific paragraph retrieval across long, heterogeneous radiology reports and supported stable local deployment under institutional data-privacy constraints. Considering paragraph retrieval fidelity, abnormality extraction accuracy, and robustness, Qwen2.5-32B was selected as the primary model for downstream lesion feature extraction.

The consolidated performance results for all modeling approaches are presented in Table 2. Table 2 reports performance on the independent 20% held-out test set. Values are summarized as mean and SD across 5-fold–specific models, each trained during cross-validation within the 80% training partition and then evaluated on the same held-out test set. In the unimodal analyses, the SVM model applied to structured clinical variables achieved the highest descriptive AUC among the 6 traditional algorithms, achieving an AUC of 0.758 and a balanced accuracy of 0.717, whereas the RF model achieved the highest F₁-score (0.833). For the engineered abnormality characteristics, logistic regression led the same set of models with an AUC of 0.716 and an F₁-score of 0.766. Ablation studies comparing undersampling, oversampling, and no-sampling strategies are provided in Table S2 in Multimedia Appendix 1 for the clinical variable models and Table S3 in Multimedia Appendix 1 for the LLM-engineered abnormality characteristic models. Among the 4 medically pretrained BERT variants evaluated on kidney-specific report findings, RadBERT achieved the highest descriptive performance, yielding an AUC of 0.746 and an F₁-score of 0.719. In addition to AUC and F₁-score, balanced accuracy was reported for all models to account for class imbalance (Table 2). Among fusion models, early fusion achieved the highest balanced accuracy (mean 0.775, SD 0.010), followed by late fusion (mean 0.749, SD 0.013), indicating consistent performance across malignant and benign classes rather than dominance by the majority malignant class.

Variability across cross-validation folds was observed in some configurations using ClinicalBERT and BioBERT embeddings. Because pretrained transformer embeddings are deterministic for fixed inputs, this variability reflects sensitivity of the downstream fusion classifier to specific training-validation splits in a high-dimensional setting rather than the instability of the embedding models themselves. For example, ClinicalBERT-based early fusion exhibited higher fold-level variance (overall accuracy: mean 0.661, SD 0.226; Table S4 in Multimedia Appendix 1), and BioBERT exhibited variability in middle fusion performance (F₁-score: mean 0.608, SD 0.344; Table S5 in Multimedia Appendix 1), indicating inconsistent fold-level behavior. In contrast, RadBERT-based models demonstrated notably more stable performance across both early and middle fusion strategies.

Multimodal integration was associated with higher descriptive predictive discrimination than unimodal modeling. The middle fusion framework achieved an AUC of 0.782, which was higher than the best unimodal baseline AUC, while the early fusion configuration reached the highest AUC of 0.813 and the highest F₁-score of 0.809. The late fusion strategy, which combined predicted probabilities from top-performing unimodal models through a meta-learner, yielded a competitive AUC of 0.805, with precision and sensitivity values of 0.902 and 0.722, respectively. These results suggest that integrating structured and unstructured data modalities may provide complementary information for renal tumor malignancy prediction.

To quantify the contribution of each data modality, we performed an ablation analysis on each fusion architecture by systematically omitting one modality and measuring the resulting impact on model performance. For the 3 modalities, all pairwise combinations were evaluated and their metrics compared against those obtained using the full 3-modality setup. Detailed results are presented in Table 3. Mean (SD) of each evaluation metric across 5-fold cross-validation is reported. Balanced accuracy was also reported in ablation analyses (Table 3) to ensure that observed performance differences across modality combinations were not driven by class prevalence effects. Input modalities include kidney-specific findings, abnormality characteristics, and clinical variables. Because RadBERT achieved the best overall performance, all ablation results reported in Table 3 are based on models incorporating RadBERT (ie, late fusion combining RadBERT with RF). Ablation analyses for the other 3 BERT variants are provided in Table S4 in Multimedia Appendix 1 (ie, early fusion missing-modality) and Table S5 in Multimedia Appendix 1 (ie, middle fusion missing-modality).

In late fusion experiments, all 24 combinations of 6 traditional machine learning classifiers and 4 pretrained BERT models were evaluated. Among these, RadBERT paired with RF emerged as the top-performing configuration (see Table S6 in Multimedia Appendix 1).

Omitting kidney-specific findings led to the greatest performance decline in both middle and late fusion configurations. In contrast, within the early fusion scheme, excluding clinical variables caused the most pronounced drop. Removal of abnormality characteristics had only a modest effect across all 3 fusion strategies, with early and late fusion approaches exhibiting robustness to their absence.

Among the unimodal approaches, structured clinical variables showed the highest descriptive predictive performance. Specifically, an SVM classifier with undersampling achieved an AUC of 0.758 (Table 2; Table S2 in Multimedia Appendix 1). By comparison, logistic regression on abnormality characteristics, enhanced via oversampling, yielded an AUC of 0.716 (Table S3 in Multimedia Appendix 1), likely limited by extraction fidelity and feature granularity. Text embeddings from complete paragraph-level kidney-specific findings slightly outperformed the abnormality descriptors, achieving an AUC of 0.746 (Table 2) when encoded with RadBERT [47] (initialized from BioBERT [49] and fine-tuned on radiology reports). PubMedBERT [50], pretrained from scratch on PubMed abstracts and full-text articles, showed comparatively lower performance, likely due to its limited alignment with radiology-report semantics. All 3 multimodal fusion strategies demonstrated higher descriptive AUC values than the unimodal models (Table 2). Early fusion, concatenating feature representations from each modality prior to classification, produced the highest AUC (0.813) and F₁-score (0.809), potentially because it preserves complementary information across modalities. Middle fusion, which first projects each modality into a lower-dimensional latent space, performed slightly worse (AUC 0.782; F₁-score=0.791), possibly due to information loss during dimensionality reduction. Late fusion remained competitive, achieving an AUC of 0.805 and an F₁-score of 0.800, but did not surpass early fusion on any major metric. Overall, these findings indicate that early integration of multimodal information achieved the highest descriptive performance in this study, while late fusion still offered stable and competitive performance through an ensemble-based combination of unimodal predictors.

Although AUC values in the range of 0.8 are insufficient to support automated clinical decision-making, they may offer incremental value for preoperative risk stratification. For patients with indeterminate renal masses, a multimodal model could complement radiologic interpretation by providing an additional quantitative estimate of malignancy risk to inform multidisciplinary discussion and clinical judgment. Such estimates are not intended to direct specific interventions but rather to contextualize imaging findings and patient characteristics within existing diagnostic workflows. Higher specificity, if confirmed in future validation studies, could potentially help reduce unnecessary interventions for benign lesions, an ongoing challenge in renal mass management. Overall, incremental gains in predictive performance may translate into practical support for clinical risk assessment when used alongside established evaluation processes.

To assess the contribution of each modality within the fusion architectures, we performed systematic ablation experiments. In the early fusion model, omitting structured clinical variables was associated with the largest decrease in AUC, from 0.813 to 0.755, indicating that clinical variables contributed the strongest predictive signal in this configuration. Removing kidney-specific findings led to a smaller decline in AUC to 0.779, whereas excluding radiologic abnormality characteristics had only a modest effect, with the AUC remaining 0.796. In the middle fusion model, the pattern differed slightly: omitting kidney-specific findings caused the greatest reduction in AUC, from 0.782 to 0.731, while excluding abnormality characteristics or clinical variables resulted in smaller decreases, to 0.749 and 0.747, respectively. In the late fusion model, removing kidney-specific findings caused the largest reduction in AUC, from 0.805 to 0.752, closely followed by omission of clinical variables, which reduced the AUC to 0.758; excluding abnormality characteristics resulted in only a modest decline, with the AUC remaining 0.794. Overall, these ablation results suggest that structured clinical variables were the dominant source of predictive information in the early fusion scheme, whereas kidney-specific text findings contributed most strongly in the middle and late fusion models. Across all 3 fusion strategies, removal of abnormality characteristics produced the smallest or among the smallest performance changes, suggesting that part of the radiologic information captured by these handcrafted features may already be represented in the kidney-specific text embeddings.

Furthermore, we conducted an interpretability analysis on the optimal early fusion model by applying Captum’s layer-integrated gradients to the kidney-specific text and Shapley additive explanations (SHAP) values to the abnormality characteristics and clinical variables. As illustrated in the horizontal bar plot visualization (Figure S11 in Multimedia Appendix 1) and exemplified in Table 4, tokens such as “huge,” “irregularly,” and “greater” attained relatively high integrated gradients attributions, suggesting a stronger contribution to malignancy predictions. This observation is in line with prior imaging research reporting that irregular tumor margins are associated with higher-grade clear cell RCC [51].

Quantitative SHAP analysis of clinical variables (Figure 3A) showed that higher systolic blood pressure, diastolic blood pressure, and BMI were associated with increased predicted probabilities of malignancy. These associations are consistent with previous epidemiological studies: individuals with diastolic pressure ≥100 mm Hg face more than a twofold higher risk of RCC compared to those with <80 mm Hg [52], and a dose-response meta-analysis demonstrates that each 1 kg/m² increase in BMI is linked to an approximately 6% higher incidence of KC [53]. For abnormality characteristics (Figure 3B), SHAP analysis identified tumor size as the most influential imaging feature, with contrast‐enhancement also contributing meaningfully to the model’s predictions. This finding aligns with prior evidence showing that each 1 cm increase in tumor diameter is associated with a ~30% higher likelihood of malignancy (effect size=1.3, 95% CI 1.22-1.43 per cm) [54]. Furthermore, established imaging guidelines indicate that postcontrast attenuation gains greater than 15 Hounsfield units may aid in distinguishing malignant renal lesions [55]. Collectively, these interpretability results suggest that our multimodal fusion framework captures and integrates clinically validated markers across multiple modalities.

Despite these findings, several limitations should be acknowledged. First, this was a single-institution study, and no external validation was performed. Consequently, model performance may reflect institution-specific characteristics, including documentation practices, patient populations, and radiology reporting styles unique to our health care system. This is particularly relevant for NLP-based feature extraction, which can be sensitive to local language conventions. Second, the retrospective use of EHR data introduces inherent biases related to incomplete or inaccurately recorded information. Mean imputation was applied strictly to continuous variables, specifically laboratory measurements and vital signs. To reduce instability from extreme sparsity, laboratory variables with greater than 50% missingness were excluded before imputation. However, some retained variables, including cholesterol and high-density lipoprotein cholesterol, had missingness rates close to this threshold and remained subject to mean imputation. The missingness percentages for these variables are reported in Table S8 in Multimedia Appendix 1 to provide transparency regarding the extent of missing data. We acknowledge that this simplified strategy may attenuate physiological variance and fail to preserve correlations between related clinical measures, such as metabolic markers, potentially affecting model calibration, particularly for non–tree-based models used in the fusion strategies. In addition, continuous structured variables were imputed and standardized using cohort-level parameters rather than parameters estimated separately within each cross-validation training fold, which may introduce limited information leakage. Third, ground-truth tumor labels were not uniformly confirmed by surgical pathology. For the unoperated majority of the cohort, outcome labels were derived from longitudinal clinical diagnosis codes that may reflect radiologic interpretation rather than definitive tumor biology. This introduces a risk of diagnostic circularity, where the imaging features used for prediction overlap with those informing clinical labeling. While we mitigated this by requiring repeated concordant codes on distinct dates and performing sensitivity analyses on the pathology-confirmed subset, residual misclassification and circularity may persist. To further assess potential temporal leakage from treatment-related terminology, we conducted a temporal audit of all radiology reports used for feature extraction. Among 1338 reports, only 17 (1.27%) contained the term “surgically,” and none of these patients had nephrectomy or other renal procedural codes recorded prior to the index date, suggesting minimal risk of leakage from completed interventions. However, because explicit diagnostic terminology was not systematically removed from the radiology text, some NLP-derived features may partially reflect radiologists’ diagnostic impressions rather than purely morphological descriptors, introducing a potential source of label leakage in the text-derived features. Although the updated interpretability analysis primarily highlighted descriptive terms such as “greater,” “irregularly,” and “huge,” the broader possibility of diagnostic-language leakage remains a limitation of the current approach. Finally, the LLM-based extraction pipeline incurs nontrivial computational requirements. While models were used in inference-only mode, larger architectures like Qwen2.5-32B require high-memory GPU infrastructure for local deployment. Because processing a single radiology report requires several seconds in batch mode, the proposed framework is currently best suited for asynchronous deployment within institutional analytics pipelines rather than real-time clinical use in resource-constrained environments.

Future work should focus on validating this multimodal approach across multiple health care institutions using larger and more diverse patient populations to ensure broad applicability. Although internal validation demonstrated consistent performance, multicenter external validation across heterogeneous health care systems is required to assess generalizability prior to clinical implementation. Prospective studies are also needed to evaluate clinical use and real-world effectiveness in live diagnostic workflows. Incorporating more robust missing data handling strategies, such as iterative imputation (eg, Multivariate Imputation by Chained Equations [56]), could more dynamically capture physiological dependencies among continuous clinical variables. Furthermore, integrating longitudinal patient data may allow the model to provide dynamic risk assessments as new clinical information becomes available over time. Additional research could explore deeper integration with imaging modalities, including advanced radiomic features or the direct use of raw imaging data, potentially enhancing diagnostic accuracy. Finally, improving scalability and computational efficiency through model compression techniques or cloud-based solutions could facilitate wider clinical adoption in resource-constrained clinical settings.

This study developed and evaluated a multimodal malignancy prediction pipeline that integrates structured EHR variables with radiology report–derived features, including LLM-extracted abnormality characteristics and transformer-based embeddings of kidney-specific findings. Across systematic comparisons of early, middle, and late fusion, multimodal models achieved higher descriptive performance than unimodal baselines, with early fusion yielding the highest AUC and F₁-score. Ablation and interpretability analyses indicated that structured clinical variables and kidney-specific report embeddings appeared to provide complementary predictive value: clinical variables provided the strongest signal in the early fusion setting, whereas kidney-specific text embeddings contributed most strongly in the middle and late fusion models. Together, these findings suggest that multimodal EHR modeling is a promising and scalable approach for supporting preoperative renal mass risk stratification, although external validation and formal statistical comparisons are needed before claims of superiority or clinical effectiveness can be made.