One AI Training Fits All? Exploring Behavioral Personas in Rare Cancer Diagnosis— Fumarate Hydratase-Deficient Renal Cell Carcinoma

One AI Training Fits All? Exploring Behavioral Personas in Rare Cancer Diagnosis— Fumarate Hydratase-Deficient Renal Cell Carcinoma | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article One AI Training Fits All? Exploring Behavioral Personas in Rare Cancer Diagnosis— Fumarate Hydratase-Deficient Renal Cell Carcinoma Changhyun Park, Yong Il Lee, Jinew Seo, Ja-Min Park, Sun Young Yoon, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8343823/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 11 You are reading this latest preprint version Abstract Fumarate Hydratase-deficient Renal Cell Carcinoma (FHdRCC) is a rare (< 0.4% of RCCs), aggressive subtype with significant morphological overlap, posing diagnostic challenges. While artificial intelligence (AI) shows promise in common cancer diagnostics, its impact on pathologist decision-making in rare diseases—particularly concerning automation bias—remains poorly understood. We developed a deep learning model to classify FHdRCC. We conducted a crossover reader study with 21 pathologists (7 genitourinary (GU) specialists, 7 non-GU specialists, 7 residents) diagnosing 30 challenging cases (15 FHdRCC, 15 non-FHdRCC) with and without AI assistance. We analyzed diagnostic performance and performed an exploratory analysis of human-AI interaction by quantifying AI Acceptance Rate (AAR) and Automation Bias Rate (ABR)—the rates of following AI recommendations when correct or incorrect, respectively—leading to the identification of preliminary behavioral personas. AI assistance significantly improved diagnostic accuracy (60.0% to 73.3%, p = 0.012) and inter-rater reliability (Fleiss' κ from 0.311 to 0.482, p < 0.001). AI-driven gains were negatively correlated with baseline expertise (R=-0.66, p = 0.001), revealing independence from traditional training. Clustering identified two behavioral personas: Receptive (n = 14; high AAR/ABR) employing efficiency-focused strategies, and Resistant (n = 7; low AAR/ABR) using deliberation-focused approaches. While the Receptive group drove accuracy gains (p = 0.041), high automation bias neutralized improvements in overall optimal decision-making (p = 0.259). AI assistance enhances rare cancer diagnostic accuracy, but effectiveness is mediated by behavioral personas rather than traditional expertise. The performance paradox—accuracy gains offset by automation bias—suggests persona-tailored training is essential: Receptive users need critical evaluation skills; Resistant users need trust-building. These exploratory findings require validation in larger studies with sufficient power to characterize automation bias patterns. Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology Fumarate hydratase-deficient renal cell carcinoma Deep learning Human-AI interaction Automation bias Behavioral personas Digital pathology Figures Figure 1 Figure 2 Figure 3 Figure 4 1 Introduction Renal cell carcinoma (RCC), a significant malignancy accounting for 3–5% of adult cancers, comprises over 15 distinct histologic subtypes, making accurate classification critical for patient prognosis and personalized treatment strategies( 1 , 2 ). While the recent advent of digital pathology and artificial intelligence (AI) has opened new possibilities by improving diagnostic accuracy for common subtypes like clear cell RCC, its potential remains confined to a limited scope( 3 , 4 ). Fumarate hydratase-deficient renal cell carcinoma (FHdRCC) represents a particularly challenging diagnostic entity in genitourinary pathology. This rare (< 0.4% of all RCCs) and highly aggressive variant is caused by germline or somatic mutations in the fumarate hydratase (FH) gene and is often associated with hereditary leiomyomatosis and renal cell carcinoma (HLRCC) syndrome( 5 – 8 ). A greater challenge is its significant morphological overlap with other high-grade RCCs, particularly papillary RCC type 2, which presents a profound diagnostic difficulty that can lead to misdiagnosis( 5 , 8 – 11 ). Definitive diagnosis therefore requires ancillary testing, such as immunohistochemistry (IHC) for FH and 2-succinocysteine (2SC), and/or FH gene mutation analysis, making this "triple threat" of rarity, aggressiveness, and diagnostic difficulty an ideal challenge where AI assistance is most critically needed. However, while pathology AI research has made significant strides in common malignancies, several important gaps remain. First, most studies have understandably focused on high-prevalence diseases where large training datasets are readily available, leaving rare entities like FHdRCC relatively unexplored( 3 , 4 , 12 , 13 ). Second, much of the research has concentrated on evaluating the standalone performance of AI models, with fewer studies examining the actual impact of AI assistance on pathologists' diagnostic capabilities in real-world clinical contexts( 14 – 16 ). Finally, there remains limited empirical evidence on the complex dynamics of human-AI interaction behavioral patterns across expertise levels in a rare diseased pattern of pathology—specifically, how AI influences decision-making processes and how potential risks such as automation bias manifest in diagnostic practice( 14 , 16 , 17 ). To address these gaps, this study has three primary objectives: ( 1 ) to develop a deep learning model for classifying the rare and diagnostically challenging FHdRCC; ( 2 ) to evaluate the impact of this AI-assistive tool on the diagnostic performance of pathologists with varying expertise (genitourinary (GU) specialists, non-GU specialists, and residents); and ( 3 ) to conduct an exploratory analysis of AI's influence on the decision-making process, proposing a model of behavioral personas based on the trade-off between AI acceptance and automation bias. 2 Materials and methods 2.1 Study cohort construction and ground truth establishment This study was approved by the Institutional Review Board of Asan Medical Center (IRB No. 2023 − 0674). From our institution's archives (10,437 kidney tumor cases, 1989–2023), we assembled a diagnostically challenging cohort by screening for FHdRCC and its histological mimickers. We included cases based on criteria predictive of diagnostic difficulty: high grade (World Health Organization/International Society of Urological Pathology (WHO/ISUP) grade 3 or 4), advanced stage (pT3 or higher), or prior immunohistochemistry (IHC) for FH/2SC performed due to suspicion of FHdRCC. Ground truth diagnosis was established by two pathologists (Y.I.L. and Y.M.C.) who independently reviewed all hematoxylin and eosin (H&E) and IHC results according to 2022 WHO classification( 2 ), with discrepancies resolved by consensus. (Detailed tissue microarray (TMA) construction and IHC panel methods are provided in Supplementary Material S1 and Supplementary Table 1). 2.2 Digital slide acquisition and datasets Following ground truth establishment, available H&E slides were digitized into whole-slide images (WSIs). Manual annotations categorized regions of interest (ROIs) into three classes: FHdRCC, Others RCC (other RCC subtypes), and Normal Kidney (non-neoplastic tissue). These annotated WSIs were partitioned into an internal dataset (for model training and validation) and two independent sets, External Dataset 1 (technical validation) and External Dataset 2 (external validation). (Details in Figure. 1, Supplementary Material S1, and Supplementary Tables 2 and 3). 2.3 Model development and evaluation We selected the ResNeXt101_32x8d model as our baseline. The final fully connected layer was modified for a three-class classification task (FHdRCC, Others RCC, Normal Kidney) (Figure. 1B). For WSI-level analysis, the model aggregated patch predictions: it first filtered out all patches predicted as Normal Kidney, then calculated the relative proportion of FHdRCC patches versus 'Others RCC' patches to yield a final, binary WSI-level classification (FHdRCC vs. non-FHdRCC). Model performance was evaluated at both patch and WSI levels. (Complete details on model architecture, training, and evaluation are provided in Supplementary Materials S2–S4). 2.4 Reader study Our crossover reader study (Figure. 1D) included 21 pathologists (7 genitourinary (GU) specialists, 7 non-GU specialists, 7 residents) who evaluated 30 blinded cases (15 FHdRCCs, 15 non-FHdRCCs; Supplementary Table 4). Participants were randomly allocated into two groups (Group A, n = 11; Group B, n = 10). In two sessions separated by a 2-week washout period, participants evaluated the cases under AI-assisted and unassisted conditions. In the assisted session, AI provided two outputs (Supplementary Figure. 1): ( 1 ) a colormap visualizing suspected FHdRCC locations and ( 2 ) a final binary prediction (e.g., AI Prediction: FHdRCC). Primary performance endpoints were accuracy, recall, precision, and reading time. (Details on the reader study are described in Supplementary Materials S5). 2.5 Human-AI interaction behavioral analysis 2.5.1 Behavioral archetype classification To analyze the pathologist's decision-making process, each diagnostic interaction was classified into one of four mutually exclusive behavioral archetypes based on two axes: ( 1 ) influence (diagnosis changed: Yes/No) and ( 2 ) agreement (final diagnosis aligned with AI: Yes/No). This yielded: Confirmation (No change, Agree), Conversion (Change, Agree), Defiance (No change, Disagree), and Rejection (Change, Disagree). These archetypes were then stratified as either Beneficial (correct final diagnosis) or Detrimental (incorrect final diagnosis)( 18 , 19 ). 2.5.2 Analysis of contextual human-AI interaction rates We defined four distinct diagnostic scenarios based on the correctness of the pathologist's initial diagnosis and the AI's prediction, relative to the ground truth( 20 , 21 ) (Supplementary Table 5): Scenario S1 (Concordant Correct) : Initial diagnosis was correct, and the AI prediction was correct. The optimal decision is Beneficial Confirmation . Scenario S2 (AI Corrects Human) : Initial diagnosis was incorrect, but the AI prediction was correct. The optimal decision is Beneficial Conversion . Scenario S3 (Human Corrects AI) : Initial diagnosis was correct, but the AI prediction was incorrect. The optimal decision is Beneficial Defiance . Scenario S4 (Concordant Incorrect) : Initial diagnosis was incorrect, and the AI prediction was incorrect. The optimal decision is Beneficial Rejection . Based on these scenarios, we defined and calculated three context-dependent metrics for each pathologist: AI Acceptance Rate (AAR) : This metric measures a pathologist's ability to be correctly influenced by the AI, specifically in Scenario S2. $$\:\text{A}\text{I}\:\text{A}\text{c}\text{c}\text{e}\text{p}\text{t}\text{a}\text{n}\text{c}\text{e}\:\text{R}\text{a}\text{t}\text{e}\:\left(\text{A}\text{A}\text{R}\right)=\frac{\text{N}\text{u}\text{m}\text{b}\text{e}\text{r}\:\text{o}\text{f}\:\text{B}\text{e}\text{n}\text{e}\text{f}\text{i}\text{c}\text{i}\text{a}\text{l}\:\text{C}\text{o}\text{n}\text{v}\text{e}\text{r}\text{s}\text{i}\text{o}\text{n}\text{s}\:\text{i}\text{n}\:\text{S}2}{\text{T}\text{o}\text{t}\text{a}\text{l}\:\text{N}\text{u}\text{m}\text{b}\text{e}\text{r}\:\text{o}\text{f}\:\text{S}2\:\text{S}\text{c}\text{e}\text{n}\text{a}\text{r}\text{i}\text{o}\text{s}}$$ Automation Bias Rate (ABR) : These metric measures susceptibility to automation bias when the AI is incorrect, specifically in Scenarios S3 and S4. $$\:\text{A}\text{u}\text{t}\text{o}\text{m}\text{a}\text{t}\text{i}\text{o}\text{n}\:\text{B}\text{i}\text{a}\text{s}\:\text{R}\text{a}\text{t}\text{e}\:\left(\text{A}\text{B}\text{R}\right)=\frac{\text{N}\text{u}\text{m}\text{b}\text{e}\text{r}\:\text{o}\text{f}\:[\text{D}\text{e}\text{t}\text{r}\text{i}\text{m}\text{e}\text{n}\text{t}\text{a}\text{l}\:\text{C}\text{o}\text{n}\text{v}\text{e}\text{r}\text{s}\text{i}\text{o}\text{n}\text{s}\:\text{i}\text{n}\:\text{S}3+\text{D}\text{e}\text{t}\text{r}\text{i}\text{m}\text{e}\text{n}\text{t}\text{a}\text{l}\:\text{C}\text{o}\text{n}\text{f}\text{i}\text{r}\text{m}\text{a}\text{t}\text{i}\text{o}\text{n}\text{s}\:\text{i}\text{n}\:\text{S}4]}{\text{T}\text{o}\text{t}\text{a}\text{l}\:\text{N}\text{u}\text{m}\text{b}\text{e}\text{r}\:\text{o}\text{f}\:[\text{S}3+\text{S}4]\:\text{S}\text{c}\text{e}\text{n}\text{a}\text{r}\text{i}\text{o}\text{s}}$$ Optimal Decision Rate (ODR) : This holistic metric assesses the quality of the decision-making process itself, measuring the proportion of all 30 cases for which the pathologist made the single most beneficial decision given the context. $$\:\text{O}\text{p}\text{t}\text{i}\text{m}\text{a}\text{l}\:\text{D}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\:\text{R}\text{a}\text{t}\text{e}\:\left(\text{O}\text{D}\text{R}\right)=\frac{\text{N}\text{u}\text{m}\text{b}\text{e}\text{r}\:\text{o}\text{f}\:\text{O}\text{p}\text{t}\text{i}\text{m}\text{a}\text{l}\:\text{D}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\text{s}}{\text{T}\text{o}\text{t}\text{a}\text{l}\:\text{D}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\text{s}\:(\text{n}=30)}$$ We performed k-means clustering on the standardized (z-scored) AAR and ABR values from all 21 participants to explore potential data-driven behavioral personas( 22 ). The optimal number of clusters (k = 2) was confirmed using standard validation techniques( 23 ). In addition, We measured the AI-Assisted Reading Time (in seconds) to compare the cognitive load of Optimal decisions versus Suboptimal decisions within each of the four interaction scenarios (S1–S4). 2.6 Statistical analysis We used the Wilcoxon signed-rank test for paired comparisons (With AI vs. Without AI), the Wilcoxon rank-sum (Mann-Whitney U) test for unpaired group comparisons (persona groups), and the Kruskal-Wallis test for expertise group comparisons. Pearson's correlation was used for correlation analyses. Generalized Linear Mixed-Effects Models (GLMM) were used to validate the crossover study design. p-values < 0.05 were considered significant. Reporting of the AI model development, validation, and diagnostic accuracy in this study was guided by the TRIPOD recommendations for prediction models. 3 Results 3.1 Clinicopathological characteristics of study cohort 3.1.1 FHdRCC cases The cohort consisted of 39 patients diagnosed with FHdRCC, with a mean age of 47.5 years and a male predominance of 72% (Table 1 ). Tumors showed aggressive features: mean size 6.8 cm, with 67% at an advanced pathological stage (pT3 or higher) and 39% with lymph node metastasis. Histologically, 95% of cases were high-grade (WHO/ISUP 3 or 4), with a classic perinucleolar halo in 92%. The most frequent histologic patterns were tubulocystic (74%) and papillary (72%). Among the initial referring diagnoses, only 36% of cases were suspected as FHdRCC, while the most common prior classification was papillary RCC, type 2 (38%). The 1-, 3-, and 5-year overall survival rates for the cohort were 97.2%, 71.6%, and 61.8%, respectively. Table 1 Clinicopathological Characteristics of FHdRCC Cases and Study Datasets Characteristic FHdRCC Cases (N = 39 patients) Internal Dataset (N = 243 WSIs) External 1 Dataset (N = 74 WSIs) External 2 Dataset (N = 18 WSIs) P-value* Clinical Characteristics Age 47.5 ± 15.9 years 56.0 (46.0–69.0) 53.0 (39.0–64.0) 52.0 (40.2–60.8) 0.071 Sex, n (%) 0.539 Male 28 (72%) 161 (66%) 44 (59%) 11 (61%) Female 11 (28%) 82 (34%) 30 (41%) 7 (39%) Diagnostic Classification WSI Diagnosis, n (%) 0.293 FHdRCC 39 (100%) 31 (13%) 5 (7%) 3 (17%) non-FHdRCC 0 (0%) 212 (87%) 69 (93%) 15 (83%) Specimen Characteristics Specimen Type, n (%) 0.001 Nephrectomy 36 (92%) 201 (83%) 74 (100%) 15 (83%) Needle Biopsy 3 (8%) 42 (17%) — 3 (17%) Tumor Size 6.8 ± 4.2 cm 4.5 (2.6–8.3) cm 8.5 (6.0–12.0) cm 6.9 (5.1–9.4) cm < 0.001 Laterality, n (%) — — — — Left 27 (69%) Right 12 (31%) Pathological Staging (AJCC 8th ed.)¹ pT Stage, n (%) < 0.001 pT1 — 84 (35%) 10 (14%) 1 (6%) pT2 — 8 (3%) 10 (14%) 2 (11%) pT3 — 105 (43%) 41 (55%) 11 (61%) pT4 — 4 (2%) 12 (16%) 0 (0%) pT1/pT2 (combined)² 12 (33%) — — — — pT3/pT4 (combined)² 24 (67%) — — — — NA — 42 (17%) 1 (1%) 4 (22%) pN Stage, n (%)² — — — — pN0 2 (6%) pN1 14 (39%) pNx (Not assessed) 20 (56%) pM Stage, n (%)² — — — — pM1 3 (8%) pMx (Not assessed) 33 (92%) Histologic Features WHO/ISUP Grade, n (%) 0.044 Grade 1 0 (0%) — — — Grade 2 2 (5%) 26 (11%) 1 (1%) 1 (6%) Grade 3 30 (77%) 140 (58%) 45 (61%) 10 (56%) Grade 4 7 (18%) 48 (20%) 24 (32%) 5 (28%) NA — 29 (12%) 4 (5%) 2 (11%) Macronucleolus with Perinucleolar Halo, n (%) 0.001 Present 36 (92%) 56 (23%) 5 (7%) 7 (39%) Absent 3 (8%) 187 (77%) 69 (93%) 11 (61%) Histologic Pattern, n (%)³ Cribriform/sieve-like 3 (8%) 7% 5% 6% 0.816 Papillary 28 (72%) 59% 53% 67% 0.478 Solid 7 (18%) 42% 53% 39% 0.239 Tubulocystic 29 (74%) 24% 16% 33% 0.198 Technical Variables Specimen Acquisition Year (median, IQR) — 2020.0 (2015.5–2022.0) 2004.0 (2000.0–2006.8) 2021.0 (2020.0–2021.8) < 0.001 Slide Preparation Year (median, IQR) — 2021.0 (2016.0–2022.0) 2023.0 (2023.0–2023.0) 2021.0 (2020.0–2021.8) < 0.001 Initial Diagnostic Accuracy and Outcome Data² Initial Diagnosis, n (%) — — — — Papillary RCC, type 2 15 (38%) Fumarate hydratase-deficient RCC 14 (36%) RCC, NOS 4 (10%) Papillary RCC 2 (5%) Papillary RCC, type 1 2 (5%) Collecting duct carcinoma 1 (3%) Granular cell carcinoma 1 (3%) Follow-up (mean ± SD, months) 44.2 ± 38.3 — — — — Overall Survival — — — — 1-year 97.2% 3-year 71.6% 5-year 61.8% *P-values compare the three datasets (Internal, External 1, External 2) and do not apply to FHdRCC Cases. ¹Staging assessed in nephrectomy specimens only. ²Data available only for FHdRCC Cases (39 patients), not for mixed datasets. ³Multiple patterns can be present in a single tumor. Abbreviations: SD, Standard Deviation; IQR, Interquartile Range; AJCC, American Joint Committee on Cancer; WHO, World Health Organization; ISUP, International Society of Urological Pathology; RCC, Renal Cell Carcinoma; NOS, Not Otherwise Specified; FHdRCC, Fumarate Hydratase-deficient Renal Cell Carcinoma; WSI, Whole Slide Image; NA, Not Available. Note: FHdRCC Cases column represents patient-level data (N=39 patients with confirmed FHdRCC). Dataset columns represent WSI-level data (N=335 total WSIs including both FHdRCC and non-FHdRCC cases). Age values for FHdRCC Cases are mean ± SD; for Datasets are median (IQR). 3.1.2 Datasets Datasets comprised 335 whole-slide images (WSIs) partitioned into an internal set (n = 243) and two external test sets, External 1 (n = 74) and External 2 (n = 18) (Table 1 ). Although the distribution of primary diagnostic classes and patient demographics were well-balanced across the datasets, significant differences in technical and pathological variables created a challenging and heterogeneous cohort. The External 1, composed of older archival blocks (median acquisition year 2004), had significantly larger tumors and a higher proportion of advanced pT stages (p < 0.001) than the more contemporary Internal and External 2 sets. Furthermore, statistically significant differences were observed across datasets in the distribution of WHO/ISUP grades (p = 0.023) and the presence of the perinucleolar halo (p < 0.001). These variations tested the model's generalizability and real-world robustness. 3.2 Overall model performance At the patch level, the model demonstrated strong performance on the internal dataset (FHdRCC area under curve (AUC): 0.951) but showed a clear performance gradient on external sets. For FHdRCC classification, precision dropped from 0.48 (internal) to 0.18 and 0.24 (external), while recall was maintained at 0.67 and 0.71. At WSI-level, internal AUC (0.968) decreased to 0.768 and 0.867 on external datasets. These slide-level external results were associated with wide confidence intervals, indicating limited statistical power due to small sample sizes. Detailed metrics for all datasets are provided in Supplementary Material S6 and Supplementary Tables 6 and 7. 3.3 Validation of reader study design To ensure findings were not confounded by the crossover study design, we conducted validation analyses. Mixed-effects modeling confirmed no significant carry-over effects from session period (p = 0.772) or group sequence (p = 0.741), no learning or fatigue effects across the 30-case sequence (p > 0.470), and no significant impact from dataset source (p > 0.220). Additionally, AI assistance improved inter-rater reliability across all groups, with overall Fleiss' kappa increasing from 0.311 to 0.482 (detailed results in Supplementary Materials S7 and Supplementary Tables 8–10). 3.4 Impact of AI assistance on diagnostic performance We first evaluated the overall impact of AI assistance. For all 21 pathologists, AI significantly improved diagnostic accuracy (60.0% to 73.3%, p = 0.012) and precision (66.7% to 77.8%, p = 0.003) without increasing interpretation time (p = 0.224. When stratified by expertise, this benefit was statistically significant only for the GU subspecialty group (Accuracy p = 0.022), while other groups showed only numerical improvements (Figure. 2A; Supplementary Table 11). AI also promoted greater diagnostic consensus. Inter-rater reliability for all 21 pathologists improved significantly, with Fleiss' kappa rising from 0.31 to 0.48 (Supplementary Figure. 2), indicating reduced diagnostic variability. 3.5 Decoupling of traditional expertise and AI-interaction proficiency The finding that GU specialists benefited most was counter-intuitive, prompting an individual-level (n = 21) analysis to find the true drivers. There are a strong negative correlation (R = − 0.66, p = 0.0012) between baseline accuracy and AI-driven accuracy gain (Supplementary Fig. 3A), no correlation (R = 0.27, p = 0.23) between baseline accuracy and AI-collaboration skill (ODR) (Supplementary Fig. 3B) and no significant difference between expertise groups regarding AAR (p = 0.594), ABR (p = 0.690), and ODR (p = 0.090) (Supplementary Figure. 5 and Supplementary Table 12). These findings show that ( 1 ) professional labels are a poor proxy for AI interaction, and ( 2 ) AI collaboration is a distinct capability decoupled from traditional diagnostic skill. 3.6 Exploratory identification of behavioral personas To explore these data-driven patterns, we performed k-means clustering on the n = 21 pathologists using their AAR (benefit-seeking) and ABR (risk-taking) scores( 24 ). Standard statistical methods (Elbow and Silhouette) robustly confirmed k = 2 as the optimal number of clusters (Supplementary Figure. 4). This analysis revealed two stable, mutually exclusive personas (Fig. 2 B): Resistant (n = 7): A skeptical group (low AAR, low ABR). Receptive (n = 14): An accepting group (high AAR, high ABR). Crucially, these data-driven personas were entirely independent of traditional expertise labels (χ 2 test, p = 0.807)( 18 ). Both the Resistant (2 GU, 2 non-GU, 3 residents) and Receptive (5 GU, 5 Non-GU, 4 Resident) groups contained a balanced mix of all expertise levels (Supplementary Table 13). This finding confirms that AI interaction style is a behavioral trait distinct from traditional training( 25 ). 3.7 Exploratory analysis of persona-based performance We analyzed the diagnostic outcomes of these proposed personas. All performance gains were driven by the 'Receptive' group, which showed a statistically significant improvement in Accuracy (p = 0.041) and Precision (p = 0.017). The 'Resistant' persona showed no significant improvement (Figure. 3A). However, this benefit appeared to be neutralized by an opposing risk profile. An exploratory analysis of the five AI-incorrect cases (n = 5), though limited by statistical power, suggested the 'Receptive' group was more susceptible to automation bias (ABR) (p < 0.001) (Figure. 3B and Supplementary Table 14). This finding suggests a performance paradox: the Receptive group's gains from AI's correct answers (S2) were cancelled out by their losses from AI's incorrect answers (S3/S4). Consequently, the overall Optimal Decision Rate (ODR) across all 30 cases showed no statistically significant difference between the two personas (p = 0.259) (Figure. 3B and Supplementary Table 14). 3.8 Cognitive load analysis of proposed behavioral personas Finally, we analyzed cognitive load (reading time) to further explore the potential mechanism behind this paradox( 26 ), revealing two different cognitive strategies. Participants had reduced cognitive load when they agreed with AI predictions, as shown in both expertise and persona analyses (Figure. 4, Supplementary Figure. 6, and Supplementary Tables 15 and 16)( 18 ). Focusing on the robust S2 scenario (n = 25 cases, AI = Correct), we found the 'Resistant' group employed a deliberation strategy. They showed no significant time difference between accepting (Optimal) or rejecting (Suboptimal) AI's correct advice (p = 0.111), remaining slow and skeptical even when AI was right (Figure. 4B). In contrast, the Receptive group employed an "Efficiency" strategy( 27 ), being significantly faster (p = 0.009) when accepting AI's correct advice versus rejecting it. This "path of least resistance" behavior explains their high AAR (Benefit). Exploratory analysis of S3/S4 scenarios (n = 5) suggests this same fast, low-effort strategy also explains their high ABR (Risk), as succumbing to bias was cognitively cheaper (Figure. 4C and D). 4 Discussion This study demonstrates that AI assistance can significantly improve diagnostic accuracy (60.0% to 73.3%, p = 0.012) for pathologists diagnosing the rare and challenging FHdRCC. This finding aligns with literature on AI-driven accuracy improvements in pathology but extends this benefit to the critical domain of rare diseases, an area of high uncertainty for pathologists. The observed improvement in inter-rater reliability (Fleiss' kappa 0.311 to 0.482) also suggests AI can help standardize diagnoses. A key finding is that this performance improvement is not correlated with traditional clinical expertise (GU specialist, non-GU, resident). Expertise showed no correlation with AAR (p = 0.594), ABR (p = 0.690), or ODR (p = 0.090). This expert paradox (Supplementary Figure. 3A) revealed that AI-driven accuracy gain was strongly negatively correlated (Pearson R = − 0.66) with the pathologist's unassisted baseline accuracy( 19 ). This suggests AI collaboration is a novel skill set, decoupled from traditional diagnostic competency( 18 ). Clinically, this implies AI can act as a safety net, particularly for non-specialists or when experts face challenging cases( 28 ). Our exploratory analysis identified two data-driven behavioral personas based on interaction patterns: a Receptive (n = 14; high AAR, high ABR) and a Resistant (n = 7; low AAR, low ABR) group (Figure. 3B). These personas were independent of traditional expertise (p = 0.807)( 18 ). This revealed a persona paradox: all significant accuracy gains were driven by the Receptive group (Figure. 4), but this was neutralized by their high susceptibility to automation bias (ABR) (Figure. 3B). The gains from AI's correct answers were offset by losses from its incorrect ones( 20 ). Consequently, the overall Optimal Decision Rate (ODR) was not significantly different between the groups (p = 0.259) (Figure. 3B). This finding moves beyond general warnings about automation bias( 29 ) to empirically link it to a specific behavioral profile, suggesting that improving statistical accuracy does not automatically improve overall decision quality. This paradox appears driven by different cognitive strategies (Figure. 4). The Receptive group used a fast, efficiency-focused strategy, engaging in cognitive offloading( 30 ). They were faster when accepting correct AI advice (S2, p = 0.009)( 18 ), a low-effort strategy that also explains their high ABR. The Resistant group used a slow, deliberation-focused strategy, showing no time difference (p = 0.111) as they verified the AI's input. This protected them from bias (low ABR) but limited their gains (low AAR), reflecting a true human-in-the-loop oversight role. These exploratory findings suggest AI can serve as a safety net for non-specialists or on rare cases (the expert paradox). However, the persona paradox hypothesizes that a one-size-fits-all deployment is insufficient. Persona-tailored training may be required: Receptive users may need training in healthy skepticism (Beneficial Defiance), while Resistant users may need help building appropriate trust (Beneficial Conversion). These findings are exploratory. The primary limitation is statistical power. The automation bias analysis was based on only five AI-incorrect cases (the n = 5 problem), rendering the ABR metric statistically unstable. Therefore, the persona paradox is a hypothesis-generating model requiring validation. Other limitations include the artificial reader study environment, generalizability from a single rare disease, and an analysis limited to the AI's binary output, not visual colormap interaction. Future research must validate this model in larger cohorts designed with a robust number of AI-incorrect cases. Additionally, prospective studies evaluating persona-tailored AI systems could determine whether adaptive interfaces improve decision quality. This study confirms AI's utility in improving diagnostic accuracy for rare cancers. Critically, it provides a preliminary, exploratory model suggesting interaction is dictated not by traditional expertise but by data-driven behavioral personas—a decoupling. This model reveals a persona paradox driven by a trade-off between efficiency-focused cognitive offloading and deliberation-focused verification. Future research is essential to validate these exploratory personas and determine if this paradox represents a generalizable challenge in human-AI collaboration. Declarations Consent to Participate This study was approved by the Institutional Review Board of Asan Medical Center (IRB No. 2023-0674) and conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the IRB for the retrospective analysis of de-identified archival specimens. Written informed consent was obtained from all participating pathologists in the reader study. Data Availability De-identified data, whole-slide images, and model weights are available to qualified researchers for 5 years post-publication upon reasonable request. Requests require a research proposal directed to the corresponding authors ( [email protected] or [email protected] ), subject to data use agreement and IRB approval (Asan Medical Center IRB No. 2023-0674). The datasets generated and/or analysed during the current study are not publicly available due to patient privacy restrictions but are available from the corresponding authors upon reasonable request with appropriate ethical approvals. Code Availability The source code used to train and evaluate the deep learning model, along with scripts for statistical analysis, is publicly available at https://github.com/brody9512/FhdRCC_AHI. The code can be accessed without restrictions. Competing Interests All authors declare no financial or non-financial competing interests. Author Contributions CP and YIL contributed equally as co-first authors; NK and YMC contributed equally as co-corresponding authors. CP and YIL drafted the manuscript. CP developed the software and methodology and performed formal analysis. CP and YIL conducted formal analysis. JS, J-MP, SYY, and BA contributed to investigation, data curation, and resources. NK and YMC provided supervision and funding acquisition. All authors reviewed and approved the final manuscript. CP, YIL, and NK verified the underlying data. All authors had full access to all the data in the study and accept responsibility for the decision to submit for publication. Funding This work was supported by the Korea Health Industry Development Institute (HR20C0026); National Research Foundation of Korea (RS-2025-00514209); and Asan Institute for Life Sciences, Asan Medical Center (2024IT0001-1). References Ahn B, Jeong J, Lee YI, Park J-M, Yoon SY, Song C, et al. Pathologic Diagnosis of Renal Cell Carcinoma in the Era of the 2022 World Health Organization Classification: Key Points for Clinicians. J Urol Oncol. 2024;22(2):115–27. WHO Classification of Tumours Editorial Board. Renal cell tumours. Urinary and male genital tumours. WHO Classification of Tumours. 8. 5th ed. Lyon, France: International Agency for Research on Cancer; 2022. p. 32–42. Li M-Y, Pan Y, Lv Y, Ma H, Sun P-L, Gao H-W. Digital pathology and artificial intelligence in renal cell carcinoma focusing on feature extraction: a literature review. Frontiers in Oncology. 2025;Volume 15–2025. Moon SW, Kim J, Kim YJ, Kim SH, An CS, Kim KG, et al. Leveraging explainable AI and large-scale datasets for comprehensive classification of renal histologic types. Scientific Reports. 2025;15(1):1745. Lau HD, Chan E, Fan AC, Kunder CA, Williamson SR, Zhou M, et al. A Clinicopathologic and Molecular Analysis of Fumarate Hydratase-deficient Renal Cell Carcinoma in 32 Patients. The American Journal of Surgical Pathology. 2020;44(1). Sun G, Zhang X, Liang J, Pan X, Zhu S, Liu Z, et al. Integrated Molecular Characterization of Fumarate Hydratase–deficient Renal Cell Carcinoma. Clinical Cancer Research. 2021;27(6):1734–43. Trpkov K, Hes O, Agaimy A, Bonert M, Martinek P, Magi-Galluzzi C, et al. Fumarate Hydratase–deficient Renal Cell Carcinoma Is Strongly Correlated With Fumarate Hydratase Mutation and Hereditary Leiomyomatosis and Renal Cell Carcinoma Syndrome. The American Journal of Surgical Pathology. 2016;40(7). Chen DA, Virk RK. Fumarate Hydratase–Deficient Renal Cell Carcinoma: A Review. AJSP: Reviews & Reports. 2020;25(6):280–3. Sung JW, Lee YI, Kim Y, Song C, Park J-M, Yoon SY, et al. Papillary renal cell carcinoma revisited: impact of the World Health Organization 2022 classification on prognostication. BJU International. 2025;135(3):510–6. Yang X, Liu Y, Wang H, Xu Y, Zhang H, Zhao M, et al. Fumarate Hydratase–Deficient Renal Cell Carcinoma With Predominant Tubulocystic Features Mimics Tubulocystic Renal Cell Carcinoma. Archives of Pathology & Laboratory Medicine. 2024;148(12):1358–64. Ohe C, Smith SC, Sirohi D, Divatia M, de Peralta-Venturina M, Paner GP, et al. Reappraisal of Morphologic Differences Between Renal Medullary Carcinoma, Collecting Duct Carcinoma, and Fumarate Hydratase–deficient Renal Cell Carcinoma. The American Journal of Surgical Pathology. 2018;42(3):279–92. Aggarwal A, Bharadwaj S, Corredor G, Pathak T, Badve S, Madabhushi A. Artificial intelligence in digital pathology — time for a reality check. Nature Reviews Clinical Oncology. 2025;22(4):283–91. Distante A, Marandino L, Bertolo R, Ingels A, Pavan N, Pecoraro A, et al. Artificial Intelligence in Renal Cell Carcinoma Histopathology: Current Applications and Future Perspectives. Diagnostics. 2023;13(13):2294. Asif A, Rajpoot K, Graham S, Snead D, Minhas F, Rajpoot N. Unleashing the potential of AI for pathology: challenges and recommendations. The Journal of Pathology. 2023;260(5):564–77. Bodén ACS, Molin J, Garvin S, West RA, Lundström C, Treanor D. The human-in-the-loop: an evaluation of pathologists’ interaction with artificial intelligence in clinical practice. Histopathology. 2021;79(2):210–8. Wekenborg MK, Gilbert S, Kather JN. Examining human-AI interaction in real-world healthcare beyond the laboratory. npj Digital Medicine. 2025;8(1):169. Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2011;19(1):121–7. Gaube S, Suresh H, Raue M, Merritt A, Berkowitz SJ, Lermer E, et al. Do as AI say: susceptibility in deployment of clinical decision-aids. npj Digital Medicine. 2021;4(1):31. Tschandl P, Rinner C, Apalla Z, Argenziano G, Codella N, Halpern A, et al. Human–computer collaboration for skin cancer recognition. Nature Medicine. 2020;26(8):1229–34. Bansal G, Wu T, Zhou J, Fok R, Nushi B, Kamar E, et al., editors. Does the whole exceed its parts? the effect of ai explanations on complementary team performance. Proceedings of the 2021 CHI conference on human factors in computing systems; 2021. Rainey C, Bond R, McConnell J, Gill A, Hughes C, Kumar D, et al. The impact of AI feedback on the accuracy of diagnosis, decision switching and trust in radiography. PLOS ONE. 2025;20(5):e0322051. Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2012;19(1):121–7. Kodinariya TM, Makwana PR. Review on determining number of Cluster in K-Means Clustering. International Journal. 2013;1(6):90–5. Rahwan I, Cebrian M, Obradovich N, Bongard J, Bonnefon J-F, Breazeal C, et al. Machine behaviour. Nature. 2019;568(7753):477–86. Jacobs M, Pradier MF, McCoy TH, Perlis RH, Doshi-Velez F, Gajos KZ. How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection. Translational Psychiatry. 2021;11(1):108. Sweller J. Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science. 1988;12(2):257–85. Kahneman D. Fast and slow thinking. Allen Lane and Penguin Books, New York. 2011. El-Khoury R, Zaatari G. The Rise of AI-Assisted Diagnosis: Will Pathologists Be Partners or Bystanders? Diagnostics. 2025;15(18):2308. Rosbach E, Ganz J, Ammeling J, Riener A, Aubreville M, editors. Automation Bias in AI-assisted Medical Decision-making under Time Pressure in Computational Pathology. BVM Workshop; 2025: Springer. Buçinca Z, Malaya MB, Gajos KZ. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proc ACM Hum-Comput Interact. 2021;5(CSCW1):Article 188. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterial.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 21 Jan, 2026 Reviews received at journal 19 Jan, 2026 Reviews received at journal 19 Jan, 2026 Reviews received at journal 07 Jan, 2026 Reviewers agreed at journal 06 Jan, 2026 Reviewers agreed at journal 05 Jan, 2026 Reviewers agreed at journal 05 Jan, 2026 Reviewers invited by journal 05 Jan, 2026 Editor assigned by journal 17 Dec, 2025 Submission checks completed at journal 17 Dec, 2025 First submitted to journal 12 Dec, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8343823","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":570374850,"identity":"28b63942-dc3a-4aaa-b0af-eaa452ae6c8b","order_by":0,"name":"Changhyun Park","email":"","orcid":"","institution":"College of Medicine, Our Lady of Fatima University","correspondingAuthor":false,"prefix":"","firstName":"Changhyun","middleName":"","lastName":"Park","suffix":""},{"id":570374851,"identity":"876d80c9-c6a9-4384-aeb9-65464276e22d","order_by":1,"name":"Yong Il Lee","email":"","orcid":"","institution":"Asan Medical Center, University of Ulsan College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yong","middleName":"Il","lastName":"Lee","suffix":""},{"id":570374852,"identity":"4fd8a8e7-0682-4217-8ca0-e575ab6caf72","order_by":2,"name":"Jinew Seo","email":"","orcid":"","institution":"Asan Medical Center, University of Ulsan College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Jinew","middleName":"","lastName":"Seo","suffix":""},{"id":570374853,"identity":"c3087001-6c71-42fd-a888-60864ffe96e9","order_by":3,"name":"Ja-Min Park","email":"","orcid":"","institution":"Asan Institute for Life Sciences, Asan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Ja-Min","middleName":"","lastName":"Park","suffix":""},{"id":570374854,"identity":"1c77fea0-d163-48fc-ab27-d327fae04953","order_by":4,"name":"Sun Young Yoon","email":"","orcid":"","institution":"Asan Institute for Life Sciences, Asan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Sun","middleName":"Young","lastName":"Yoon","suffix":""},{"id":570374855,"identity":"b924d3b5-b328-4288-9785-50063e26d459","order_by":5,"name":"Bokyung Ahn","email":"","orcid":"","institution":"Asan Medical Center, University of Ulsan College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Bokyung","middleName":"","lastName":"Ahn","suffix":""},{"id":570374856,"identity":"f288daae-27ac-45ee-81d9-349b4fcb2f17","order_by":6,"name":"Yong Mee Cho","email":"","orcid":"","institution":"Asan Medical Center, University of Ulsan College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yong","middleName":"Mee","lastName":"Cho","suffix":""},{"id":570374857,"identity":"267b0531-01e7-470a-8e16-bc9e6a6e4403","order_by":7,"name":"Namkug Kim","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1UlEQVRIiWNgGAWjYHACZiBO4GFgbwDSBhakaOE5ANIiQbwWBgaJBBCHCC3m7KeTjXl+pcnoznx+dcOPAgkG/vbuBLxaLHtyNyfz9uXwmN3OKbvZA3SYxJmzG/BqMTiQu/kwb08FSEvaDR6gFgOJXAJazr+Farl5Ju3mH6K03AA6jOcH0GE32I/dJs6WG283G85tSOMxO5PDdlvGQIKHsF/O526WePMn2d7s+PFnN9/8sZHjb+/FrwUMGNtAJI8BmCSsHAz+gAj2B0SqHgWjYBSMgpEGAIlxSjpUKrClAAAAAElFTkSuQmCC","orcid":"","institution":"Asan Medical Center, University of Ulsan College of Medicine","correspondingAuthor":true,"prefix":"","firstName":"Namkug","middleName":"","lastName":"Kim","suffix":""}],"badges":[],"createdAt":"2025-12-12 09:08:02","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8343823/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8343823/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":99814867,"identity":"d821336a-19dc-4a4a-aa7e-cf8d7edda5b1","added_by":"auto","created_at":"2026-01-08 14:43:00","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":86966,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/2451065f384bb1441a66b8eb.docx"},{"id":99815122,"identity":"d93acba3-aaea-495f-b83f-beadb50394f2","added_by":"auto","created_at":"2026-01-08 14:43:31","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":34583,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/a16c8e115a6e723c63e2ce07.docx"},{"id":99814810,"identity":"99115178-d63b-40bf-a876-9943e491d754","added_by":"auto","created_at":"2026-01-08 14:42:54","extension":"json","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10331,"visible":true,"origin":"","legend":"","description":"","filename":"eb4eb11fcf68453cb1a41805c9350fb9.json","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/4257d78533edeaf3ebd21284.json"},{"id":99815009,"identity":"b2ac5501-70c9-4826-9cda-89f204c82fbe","added_by":"auto","created_at":"2026-01-08 14:43:19","extension":"docx","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1623886,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/4f27e9f1db0b4661e955782b.docx"},{"id":99814620,"identity":"dc8d1ea3-e113-4d40-89ae-838290ed1058","added_by":"auto","created_at":"2026-01-08 14:42:21","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":143053,"visible":true,"origin":"","legend":"","description":"","filename":"eb4eb11fcf68453cb1a41805c9350fb91enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/1e8e07cbec048356da6b5cd3.xml"},{"id":99814709,"identity":"063959c3-a2e6-46bd-8d3a-a0c86b3bd39c","added_by":"auto","created_at":"2026-01-08 14:42:37","extension":"pptx","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":42081879,"visible":true,"origin":"","legend":"","description":"","filename":"Figure.1.pptx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/5fba6e93addd6c81b965885e.pptx"},{"id":99814743,"identity":"875df874-102b-4697-9cea-afb9165fb69d","added_by":"auto","created_at":"2026-01-08 14:42:42","extension":"pptx","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":185879,"visible":true,"origin":"","legend":"","description":"","filename":"Figure.2.pptx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/bd194ece6e8de82a49f5d7c5.pptx"},{"id":99814843,"identity":"a5bd839d-59a3-4943-8a7b-36456af5697f","added_by":"auto","created_at":"2026-01-08 14:42:58","extension":"pptx","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":227509,"visible":true,"origin":"","legend":"","description":"","filename":"Figure.3.pptx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/696685071d3ea3c1ca8ec35f.pptx"},{"id":99814707,"identity":"d0cfa62f-1532-417e-bfba-038352156d6b","added_by":"auto","created_at":"2026-01-08 14:42:37","extension":"pptx","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":336340,"visible":true,"origin":"","legend":"","description":"","filename":"Figure.4.pptx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/b6a4cb704bd291bad841f6e6.pptx"},{"id":99814622,"identity":"6e506c09-1426-428b-bbf1-dc7b4845d6a7","added_by":"auto","created_at":"2026-01-08 14:42:22","extension":"xml","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":146048,"visible":true,"origin":"","legend":"","description":"","filename":"eb4eb11fcf68453cb1a41805c9350fb91structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/70321b28ea1684512bc20ca9.xml"},{"id":99814886,"identity":"c195849b-3301-4dd6-a32a-605c03c36e3d","added_by":"auto","created_at":"2026-01-08 14:43:04","extension":"html","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":157391,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/ad4a8868bf027f28dd585bff.html"},{"id":99815102,"identity":"9030cb2e-fbf8-4531-97d5-f65756c2087c","added_by":"auto","created_at":"2026-01-08 14:43:31","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":143211,"visible":true,"origin":"","legend":"\u003cp\u003eDeep learning model development and validation workflow for fumarate hydratase-deficient renal cell carcinoma (FHdRCC) classification. (A) Multi-institutional dataset comprising internal training data from Asan Medical Center and two external validation sets were used. (B) Box-like region of interests (ROIs) including FHdRCC (red), others renal cell carcinoma (blue), and normal kidney were annotated by experts. Patch preprocessing pipeline including color normalization, data augmentation was performed and ResNeXt101_32x8d model was trained for three-class classification. (C) Using WSI-level inference using sliding windowing, relative proportion and binary classification output (FHdRCC vs non-FHdRCC) were evaluated among RCC patches. (D) A crossover reader study with 21 pathologists across three expertise levels was performed for evaluating 30 cases with and without AI assistance after 2-week washout period. Abbreviation: GU, genitourinary specialists; non-GU, non-genitourinary specialists.\u003c/p\u003e","description":"","filename":"Figure.1.png","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/7598fa1ac7758a6a1495e4b4.png"},{"id":99814809,"identity":"d99596a7-61eb-44f5-80f9-b8136efd3036","added_by":"auto","created_at":"2026-01-08 14:42:54","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":410458,"visible":true,"origin":"","legend":"\u003cp\u003eDiagnostic performance and behavioral persona analysis. (A) Diagnostic performance by expertise, with and without AI. Box plots comparing performance metrics and reading time without (red) and with (blue) AI assistance, stratified by all pathologists (All) and by expertise group (GU, non-GU, Resident). (B) Behavioral persona clustering of pathologists. K-means clustering (k=2) of 21 pathologists based on AI Acceptance Rate (AAR) (x-axis) and Automation Bias Rate (ABR) (y-axis). This identified two personas: Resistant (n=7; low AAR/ABR) and Receptive (n=14; high AAR/ABR). Abbreviations: GU, genitourinary specialists; non-GU, non-genitourinary specialists.\u003c/p\u003e","description":"","filename":"Figure.2.png","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/6ea6fae86da435688ed3006c.png"},{"id":99814735,"identity":"eeb0f8d9-e4ca-4c54-ad88-c9f437ac50c2","added_by":"auto","created_at":"2026-01-08 14:42:41","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":153069,"visible":true,"origin":"","legend":"\u003cp\u003eBehavioral persona-based performance analysis reveals the persona paradox. (A) Impact of AI assistance on diagnostic performance by behavioral persona. Box plots comparing performance without AI (red) and with AI (blue) across all pathologists and by persona. All significant gains in Accuracy (p=0.041) and Precision (p=0.017) were driven solely by the Receptive group. The Resistant group showed no significant changes. (B) Analysis of interaction metrics reveals the persona paradox. Box plots comparing key interaction metrics between the Resistant and Receptive personas: (i) The Receptive group had a significantly higher AI Acceptance Rate (AAR) (p=0.007). (ii) Conversely, they also showed a significantly higher Automation Bias Rate (ABR) (p \u0026lt; 0.001). (iii) This risk neutralized their performance gains, resulting in no significant difference in the overall Optimal Decision Rate (ODR) (p=0.259).\u003c/p\u003e","description":"","filename":"Figure.3.png","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/0b3870d349b4b9e2c914c99d.png"},{"id":99814878,"identity":"1af2baf5-ab79-4027-894e-5f08613afd5e","added_by":"auto","created_at":"2026-01-08 14:43:02","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":382879,"visible":true,"origin":"","legend":"\u003cp\u003eCognitive load analysis reveals distinct persona-based decision strategies.\u003c/p\u003e\n\u003cp\u003eBox plots comparing AI-assisted reading time (cognitive load) for Optimal (blue) vs. Suboptimal (red) decisions by persona across four scenarios (S1-S4). Resistant: No significant time difference between accepting or rejecting AI's correct advice (S2, p=0.111), indicating a deliberation strategy. Receptive: Was significantly faster when accepting correct advice (Optimal in S2, p=0.009), indicating an efficiency strategy.\u003c/p\u003e","description":"","filename":"Figure.4.png","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/eb83fc7f0f773699823c5bb2.png"},{"id":99816918,"identity":"21f8606e-6769-4ba3-ac38-c6e0c6c55946","added_by":"auto","created_at":"2026-01-08 14:48:07","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2478023,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/c85acee4-5927-4250-a0eb-205be5cd6d82.pdf"},{"id":99815007,"identity":"dfdb13f7-0ccb-40e6-bf59-4565324ba6ff","added_by":"auto","created_at":"2026-01-08 14:43:19","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1623886,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-8343823/v1/5044d483a44ddce1ae492d6b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"One AI Training Fits All? Exploring Behavioral Personas in Rare Cancer Diagnosis— Fumarate Hydratase-Deficient Renal Cell Carcinoma","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eRenal cell carcinoma (RCC), a significant malignancy accounting for 3\u0026ndash;5% of adult cancers, comprises over 15 distinct histologic subtypes, making accurate classification critical for patient prognosis and personalized treatment strategies(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). While the recent advent of digital pathology and artificial intelligence (AI) has opened new possibilities by improving diagnostic accuracy for common subtypes like clear cell RCC, its potential remains confined to a limited scope(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFumarate hydratase-deficient renal cell carcinoma (FHdRCC) represents a particularly challenging diagnostic entity in genitourinary pathology. This rare (\u0026lt;\u0026thinsp;0.4% of all RCCs) and highly aggressive variant is caused by germline or somatic mutations in the fumarate hydratase \u003cem\u003e(FH)\u003c/em\u003e gene and is often associated with hereditary leiomyomatosis and renal cell carcinoma (HLRCC) syndrome(\u003cspan additionalcitationids=\"CR6 CR7\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). A greater challenge is its significant morphological overlap with other high-grade RCCs, particularly papillary RCC type 2, which presents a profound diagnostic difficulty that can lead to misdiagnosis(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan additionalcitationids=\"CR9 CR10\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). Definitive diagnosis therefore requires ancillary testing, such as immunohistochemistry (IHC) for FH and 2-succinocysteine (2SC), and/or \u003cem\u003eFH\u003c/em\u003e gene mutation analysis, making this \"triple threat\" of rarity, aggressiveness, and diagnostic difficulty an ideal challenge where AI assistance is most critically needed.\u003c/p\u003e \u003cp\u003eHowever, while pathology AI research has made significant strides in common malignancies, several important gaps remain. First, most studies have understandably focused on high-prevalence diseases where large training datasets are readily available, leaving rare entities like FHdRCC relatively unexplored(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Second, much of the research has concentrated on evaluating the standalone performance of AI models, with fewer studies examining the actual impact of AI assistance on pathologists' diagnostic capabilities in real-world clinical contexts(\u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). Finally, there remains limited empirical evidence on the complex dynamics of human-AI interaction behavioral patterns across expertise levels in a rare diseased pattern of pathology\u0026mdash;specifically, how AI influences decision-making processes and how potential risks such as automation bias manifest in diagnostic practice(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTo address these gaps, this study has three primary objectives: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) to develop a deep learning model for classifying the rare and diagnostically challenging FHdRCC; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) to evaluate the impact of this AI-assistive tool on the diagnostic performance of pathologists with varying expertise (genitourinary (GU) specialists, non-GU specialists, and residents); and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) to conduct an exploratory analysis of AI's influence on the decision-making process, proposing a model of behavioral personas based on the trade-off between AI acceptance and automation bias.\u003c/p\u003e"},{"header":"2 Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study cohort construction and ground truth establishment\u003c/h2\u003e \u003cp\u003eThis study was approved by the Institutional Review Board of Asan Medical Center (IRB No. 2023\u0026thinsp;\u0026minus;\u0026thinsp;0674). From our institution's archives (10,437 kidney tumor cases, 1989\u0026ndash;2023), we assembled a diagnostically challenging cohort by screening for FHdRCC and its histological mimickers. We included cases based on criteria predictive of diagnostic difficulty: high grade (World Health Organization/International Society of Urological Pathology (WHO/ISUP) grade 3 or 4), advanced stage (pT3 or higher), or prior immunohistochemistry (IHC) for FH/2SC performed due to suspicion of FHdRCC. Ground truth diagnosis was established by two pathologists (Y.I.L. and Y.M.C.) who independently reviewed all hematoxylin and eosin (H\u0026amp;E) and IHC results according to 2022 WHO classification(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e), with discrepancies resolved by consensus. (Detailed tissue microarray (TMA) construction and IHC panel methods are provided in Supplementary Material S1 and Supplementary Table\u0026nbsp;1).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Digital slide acquisition and datasets\u003c/h2\u003e \u003cp\u003eFollowing ground truth establishment, available H\u0026amp;E slides were digitized into whole-slide images (WSIs). Manual annotations categorized regions of interest (ROIs) into three classes: FHdRCC, Others RCC (other RCC subtypes), and Normal Kidney (non-neoplastic tissue). These annotated WSIs were partitioned into an internal dataset (for model training and validation) and two independent sets, External Dataset 1 (technical validation) and External Dataset 2 (external validation). (Details in Figure. 1, Supplementary Material S1, and Supplementary Tables\u0026nbsp;2 and 3).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Model development and evaluation\u003c/h2\u003e \u003cp\u003eWe selected the ResNeXt101_32x8d model as our baseline. The final fully connected layer was modified for a three-class classification task (FHdRCC, Others RCC, Normal Kidney) (Figure. 1B). For WSI-level analysis, the model aggregated patch predictions: it first filtered out all patches predicted as Normal Kidney, then calculated the relative proportion of FHdRCC patches versus 'Others RCC' patches to yield a final, binary WSI-level classification (FHdRCC vs. non-FHdRCC). Model performance was evaluated at both patch and WSI levels. (Complete details on model architecture, training, and evaluation are provided in Supplementary Materials S2\u0026ndash;S4).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Reader study\u003c/h2\u003e \u003cp\u003eOur crossover reader study (Figure. 1D) included 21 pathologists (7 genitourinary (GU) specialists, 7 non-GU specialists, 7 residents) who evaluated 30 blinded cases (15 FHdRCCs, 15 non-FHdRCCs; Supplementary Table\u0026nbsp;4). Participants were randomly allocated into two groups (Group A, n\u0026thinsp;=\u0026thinsp;11; Group B, n\u0026thinsp;=\u0026thinsp;10). In two sessions separated by a 2-week washout period, participants evaluated the cases under AI-assisted and unassisted conditions. In the assisted session, AI provided two outputs (Supplementary Figure. 1): (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) a colormap visualizing suspected FHdRCC locations and (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) a final binary prediction (e.g., AI Prediction: FHdRCC). Primary performance endpoints were accuracy, recall, precision, and reading time. (Details on the reader study are described in Supplementary Materials S5).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Human-AI interaction behavioral analysis\u003c/h2\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.5.1 Behavioral archetype classification\u003c/h2\u003e \u003cp\u003eTo analyze the pathologist's decision-making process, each diagnostic interaction was classified into one of four mutually exclusive behavioral archetypes based on two axes: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) influence (diagnosis changed: Yes/No) and (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) agreement (final diagnosis aligned with AI: Yes/No). This yielded: Confirmation (No change, Agree), Conversion (Change, Agree), Defiance (No change, Disagree), and Rejection (Change, Disagree). These archetypes were then stratified as either Beneficial (correct final diagnosis) or Detrimental (incorrect final diagnosis)(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.5.2 Analysis of contextual human-AI interaction rates\u003c/h2\u003e \u003cp\u003eWe defined four distinct diagnostic scenarios based on the correctness of the pathologist's initial diagnosis and the AI's prediction, relative to the ground truth(\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e) (Supplementary Table\u0026nbsp;5):\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eScenario S1 (Concordant Correct)\u003c/b\u003e: Initial diagnosis was correct, and the AI prediction was correct. The optimal decision is \u003cb\u003eBeneficial Confirmation\u003c/b\u003e.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eScenario S2 (AI Corrects Human)\u003c/b\u003e: Initial diagnosis was incorrect, but the AI prediction was correct. The optimal decision is \u003cb\u003eBeneficial Conversion\u003c/b\u003e.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eScenario S3 (Human Corrects AI)\u003c/b\u003e: Initial diagnosis was correct, but the AI prediction was incorrect. The optimal decision is \u003cb\u003eBeneficial Defiance\u003c/b\u003e.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eScenario S4 (Concordant Incorrect)\u003c/b\u003e: Initial diagnosis was incorrect, and the AI prediction was incorrect. The optimal decision is \u003cb\u003eBeneficial Rejection\u003c/b\u003e.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eBased on these scenarios, we defined and calculated three context-dependent metrics for each pathologist:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAI Acceptance Rate (AAR)\u003c/b\u003e: This metric measures a pathologist's ability to be correctly influenced by the AI, specifically in Scenario S2.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv id=\"Equa\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\text{A}\\text{I}\\:\\text{A}\\text{c}\\text{c}\\text{e}\\text{p}\\text{t}\\text{a}\\text{n}\\text{c}\\text{e}\\:\\text{R}\\text{a}\\text{t}\\text{e}\\:\\left(\\text{A}\\text{A}\\text{R}\\right)=\\frac{\\text{N}\\text{u}\\text{m}\\text{b}\\text{e}\\text{r}\\:\\text{o}\\text{f}\\:\\text{B}\\text{e}\\text{n}\\text{e}\\text{f}\\text{i}\\text{c}\\text{i}\\text{a}\\text{l}\\:\\text{C}\\text{o}\\text{n}\\text{v}\\text{e}\\text{r}\\text{s}\\text{i}\\text{o}\\text{n}\\text{s}\\:\\text{i}\\text{n}\\:\\text{S}2}{\\text{T}\\text{o}\\text{t}\\text{a}\\text{l}\\:\\text{N}\\text{u}\\text{m}\\text{b}\\text{e}\\text{r}\\:\\text{o}\\text{f}\\:\\text{S}2\\:\\text{S}\\text{c}\\text{e}\\text{n}\\text{a}\\text{r}\\text{i}\\text{o}\\text{s}}$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAutomation Bias Rate (ABR)\u003c/b\u003e: These metric measures susceptibility to automation bias when the AI is incorrect, specifically in Scenarios S3 and S4.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv id=\"Equb\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\text{A}\\text{u}\\text{t}\\text{o}\\text{m}\\text{a}\\text{t}\\text{i}\\text{o}\\text{n}\\:\\text{B}\\text{i}\\text{a}\\text{s}\\:\\text{R}\\text{a}\\text{t}\\text{e}\\:\\left(\\text{A}\\text{B}\\text{R}\\right)=\\frac{\\text{N}\\text{u}\\text{m}\\text{b}\\text{e}\\text{r}\\:\\text{o}\\text{f}\\:[\\text{D}\\text{e}\\text{t}\\text{r}\\text{i}\\text{m}\\text{e}\\text{n}\\text{t}\\text{a}\\text{l}\\:\\text{C}\\text{o}\\text{n}\\text{v}\\text{e}\\text{r}\\text{s}\\text{i}\\text{o}\\text{n}\\text{s}\\:\\text{i}\\text{n}\\:\\text{S}3+\\text{D}\\text{e}\\text{t}\\text{r}\\text{i}\\text{m}\\text{e}\\text{n}\\text{t}\\text{a}\\text{l}\\:\\text{C}\\text{o}\\text{n}\\text{f}\\text{i}\\text{r}\\text{m}\\text{a}\\text{t}\\text{i}\\text{o}\\text{n}\\text{s}\\:\\text{i}\\text{n}\\:\\text{S}4]}{\\text{T}\\text{o}\\text{t}\\text{a}\\text{l}\\:\\text{N}\\text{u}\\text{m}\\text{b}\\text{e}\\text{r}\\:\\text{o}\\text{f}\\:[\\text{S}3+\\text{S}4]\\:\\text{S}\\text{c}\\text{e}\\text{n}\\text{a}\\text{r}\\text{i}\\text{o}\\text{s}}$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eOptimal Decision Rate (ODR)\u003c/b\u003e: This holistic metric assesses the quality of the decision-making process itself, measuring the proportion of all 30 cases for which the pathologist made the single most beneficial decision given the context.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv id=\"Equc\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:\\text{O}\\text{p}\\text{t}\\text{i}\\text{m}\\text{a}\\text{l}\\:\\text{D}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}\\:\\text{R}\\text{a}\\text{t}\\text{e}\\:\\left(\\text{O}\\text{D}\\text{R}\\right)=\\frac{\\text{N}\\text{u}\\text{m}\\text{b}\\text{e}\\text{r}\\:\\text{o}\\text{f}\\:\\text{O}\\text{p}\\text{t}\\text{i}\\text{m}\\text{a}\\text{l}\\:\\text{D}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}\\text{s}}{\\text{T}\\text{o}\\text{t}\\text{a}\\text{l}\\:\\text{D}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}\\text{s}\\:(\\text{n}=30)}$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003eWe performed k-means clustering on the standardized (z-scored) AAR and ABR values from all 21 participants to explore potential data-driven behavioral personas(\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). The optimal number of clusters (k\u0026thinsp;=\u0026thinsp;2) was confirmed using standard validation techniques(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn addition, We measured the AI-Assisted Reading Time (in seconds) to compare the cognitive load of Optimal decisions versus Suboptimal decisions within each of the four interaction scenarios (S1\u0026ndash;S4).\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Statistical analysis\u003c/h2\u003e \u003cp\u003eWe used the Wilcoxon signed-rank test for paired comparisons (With AI vs. Without AI), the Wilcoxon rank-sum (Mann-Whitney U) test for unpaired group comparisons (persona groups), and the Kruskal-Wallis test for expertise group comparisons. Pearson's correlation was used for correlation analyses. Generalized Linear Mixed-Effects Models (GLMM) were used to validate the crossover study design. p-values\u0026thinsp;\u0026lt;\u0026thinsp;0.05 were considered significant.\u003c/p\u003e \u003cp\u003eReporting of the AI model development, validation, and diagnostic accuracy in this study was guided by the TRIPOD recommendations for prediction models.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Clinicopathological characteristics of study cohort\u003c/h2\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1 FHdRCC cases\u003c/h2\u003e \u003cp\u003eThe cohort consisted of 39 patients diagnosed with FHdRCC, with a mean age of 47.5 years and a male predominance of 72% (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Tumors showed aggressive features: mean size 6.8 cm, with 67% at an advanced pathological stage (pT3 or higher) and 39% with lymph node metastasis. Histologically, 95% of cases were high-grade (WHO/ISUP 3 or 4), with a classic perinucleolar halo in 92%. The most frequent histologic patterns were tubulocystic (74%) and papillary (72%). Among the initial referring diagnoses, only 36% of cases were suspected as FHdRCC, while the most common prior classification was papillary RCC, type 2 (38%). The 1-, 3-, and 5-year overall survival rates for the cohort were 97.2%, 71.6%, and 61.8%, respectively.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClinicopathological Characteristics of FHdRCC Cases and Study Datasets\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFHdRCC Cases\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;39 patients)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInternal Dataset\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;243 WSIs)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExternal 1 Dataset\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;74 WSIs)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eExternal 2 Dataset\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;18 WSIs)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eP-value*\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClinical Characteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e47.5\u0026thinsp;\u0026plusmn;\u0026thinsp;15.9 years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e56.0 (46.0\u0026ndash;69.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e53.0 (39.0\u0026ndash;64.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e52.0 (40.2\u0026ndash;60.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.071\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSex, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.539\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28 (72%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e161 (66%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e44 (59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e11 (61%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11 (28%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e82 (34%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e30 (41%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7 (39%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDiagnostic Classification\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWSI Diagnosis, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.293\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFHdRCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e39 (100%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31 (13%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5 (7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003enon-FHdRCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e212 (87%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e69 (93%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e15 (83%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSpecimen Characteristics\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecimen Type, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNephrectomy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e36 (92%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e201 (83%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e74 (100%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e15 (83%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeedle Biopsy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3 (8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e42 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTumor Size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.8\u0026thinsp;\u0026plusmn;\u0026thinsp;4.2 cm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.5 (2.6\u0026ndash;8.3) cm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8.5 (6.0\u0026ndash;12.0) cm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.9 (5.1\u0026ndash;9.4) cm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLaterality, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLeft\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e27 (69%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12 (31%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePathological Staging (AJCC 8th ed.)\u0026sup1;\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epT Stage, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epT1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e84 (35%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10 (14%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epT2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10 (14%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 (11%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epT3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e105 (43%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e41 (55%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e11 (61%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epT4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4 (2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12 (16%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0 (0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epT1/pT2 (combined)\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12 (33%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epT3/pT4 (combined)\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24 (67%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e42 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4 (22%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epN Stage, n (%)\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epN0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epN1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14 (39%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epNx (Not assessed)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e20 (56%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epM Stage, n (%)\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epM1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3 (8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003epMx (Not assessed)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e33 (92%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eHistologic Features\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWHO/ISUP Grade, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.044\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrade 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrade 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e26 (11%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1 (6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrade 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e30 (77%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e140 (58%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e45 (61%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e10 (56%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGrade 4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7 (18%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e48 (20%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24 (32%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5 (28%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e29 (12%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4 (5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 (11%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacronucleolus with Perinucleolar Halo, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e36 (92%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e56 (23%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5 (7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e7 (39%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3 (8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e187 (77%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e69 (93%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e11 (61%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHistologic Pattern, n (%)\u0026sup3;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCribriform/sieve-like\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3 (8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.816\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePapillary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28 (72%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e59%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e53%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e67%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.478\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSolid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7 (18%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e42%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e53%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e39%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.239\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTubulocystic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29 (74%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e16%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e33%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.198\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTechnical Variables\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecimen Acquisition Year (median, IQR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2020.0 (2015.5\u0026ndash;2022.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2004.0 (2000.0\u0026ndash;2006.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2021.0 (2020.0\u0026ndash;2021.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSlide Preparation Year (median, IQR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2021.0 (2016.0\u0026ndash;2022.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2023.0 (2023.0\u0026ndash;2023.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2021.0 (2020.0\u0026ndash;2021.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInitial Diagnostic Accuracy and Outcome Data\u0026sup2;\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInitial Diagnosis, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePapillary RCC, type 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15 (38%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFumarate hydratase-deficient RCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14 (36%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRCC, NOS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4 (10%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePapillary RCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePapillary RCC, type 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCollecting duct carcinoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1 (3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGranular cell carcinoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1 (3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFollow-up (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD, months)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e44.2\u0026thinsp;\u0026plusmn;\u0026thinsp;38.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOverall Survival\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1-year\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e97.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3-year\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e71.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5-year\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e61.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e*P-values compare the three datasets (Internal, External 1, External 2) and do not apply to FHdRCC Cases.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e\u0026sup1;Staging assessed in nephrectomy specimens only.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e\u0026sup2;Data available only for FHdRCC Cases (39 patients), not for mixed datasets.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e\u0026sup3;Multiple patterns can be present in a single tumor.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAbbreviations: SD, Standard Deviation; IQR, Interquartile Range; AJCC, American Joint Committee on Cancer; WHO, World Health Organization; ISUP, International Society of Urological Pathology; RCC, Renal Cell Carcinoma; NOS, Not Otherwise Specified; FHdRCC, Fumarate Hydratase-deficient Renal Cell Carcinoma; WSI, Whole Slide Image; NA, Not Available.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote: FHdRCC Cases column represents patient-level data (N=39 patients with confirmed FHdRCC). Dataset columns represent WSI-level data (N=335 total WSIs including both FHdRCC and non-FHdRCC cases). Age values for FHdRCC Cases are mean \u0026plusmn; SD; for Datasets are median (IQR).\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2 Datasets\u003c/h2\u003e \u003cp\u003eDatasets comprised 335 whole-slide images (WSIs) partitioned into an internal set (n\u0026thinsp;=\u0026thinsp;243) and two external test sets, External 1 (n\u0026thinsp;=\u0026thinsp;74) and External 2 (n\u0026thinsp;=\u0026thinsp;18) (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Although the distribution of primary diagnostic classes and patient demographics were well-balanced across the datasets, significant differences in technical and pathological variables created a challenging and heterogeneous cohort. The External 1, composed of older archival blocks (median acquisition year 2004), had significantly larger tumors and a higher proportion of advanced pT stages (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) than the more contemporary Internal and External 2 sets. Furthermore, statistically significant differences were observed across datasets in the distribution of WHO/ISUP grades (p\u0026thinsp;=\u0026thinsp;0.023) and the presence of the perinucleolar halo (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). These variations tested the model's generalizability and real-world robustness.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Overall model performance\u003c/h2\u003e \u003cp\u003eAt the patch level, the model demonstrated strong performance on the internal dataset (FHdRCC area under curve (AUC): 0.951) but showed a clear performance gradient on external sets. For FHdRCC classification, precision dropped from 0.48 (internal) to 0.18 and 0.24 (external), while recall was maintained at 0.67 and 0.71. At WSI-level, internal AUC (0.968) decreased to 0.768 and 0.867 on external datasets. These slide-level external results were associated with wide confidence intervals, indicating limited statistical power due to small sample sizes. Detailed metrics for all datasets are provided in Supplementary Material S6 and Supplementary Tables\u0026nbsp;6 and 7.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Validation of reader study design\u003c/h2\u003e \u003cp\u003eTo ensure findings were not confounded by the crossover study design, we conducted validation analyses. Mixed-effects modeling confirmed no significant carry-over effects from session period (p\u0026thinsp;=\u0026thinsp;0.772) or group sequence (p\u0026thinsp;=\u0026thinsp;0.741), no learning or fatigue effects across the 30-case sequence (p\u0026thinsp;\u0026gt;\u0026thinsp;0.470), and no significant impact from dataset source (p\u0026thinsp;\u0026gt;\u0026thinsp;0.220). Additionally, AI assistance improved inter-rater reliability across all groups, with overall Fleiss' kappa increasing from 0.311 to 0.482 (detailed results in Supplementary Materials S7 and Supplementary Tables\u0026nbsp;8\u0026ndash;10).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Impact of AI assistance on diagnostic performance\u003c/h2\u003e \u003cp\u003eWe first evaluated the overall impact of AI assistance. For all 21 pathologists, AI significantly improved diagnostic accuracy (60.0% to 73.3%, p\u0026thinsp;=\u0026thinsp;0.012) and precision (66.7% to 77.8%, p\u0026thinsp;=\u0026thinsp;0.003) without increasing interpretation time (p\u0026thinsp;=\u0026thinsp;0.224. When stratified by expertise, this benefit was statistically significant only for the GU subspecialty group (Accuracy p\u0026thinsp;=\u0026thinsp;0.022), while other groups showed only numerical improvements (Figure. 2A; Supplementary Table\u0026nbsp;11).\u003c/p\u003e \u003cp\u003eAI also promoted greater diagnostic consensus. Inter-rater reliability for all 21 pathologists improved significantly, with Fleiss' kappa rising from 0.31 to 0.48 (Supplementary Figure. 2), indicating reduced diagnostic variability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Decoupling of traditional expertise and AI-interaction proficiency\u003c/h2\u003e \u003cp\u003eThe finding that GU specialists benefited most was counter-intuitive, prompting an individual-level (n\u0026thinsp;=\u0026thinsp;21) analysis to find the true drivers.\u003c/p\u003e \u003cp\u003eThere are a strong negative correlation (R = \u0026minus;\u0026thinsp;0.66, p\u0026thinsp;=\u0026thinsp;0.0012) between baseline accuracy and AI-driven accuracy gain (Supplementary Fig.\u0026nbsp;3A), no correlation (R\u0026thinsp;=\u0026thinsp;0.27, p\u0026thinsp;=\u0026thinsp;0.23) between baseline accuracy and AI-collaboration skill (ODR) (Supplementary Fig.\u0026nbsp;3B) and no significant difference between expertise groups regarding AAR (p\u0026thinsp;=\u0026thinsp;0.594), ABR (p\u0026thinsp;=\u0026thinsp;0.690), and ODR (p\u0026thinsp;=\u0026thinsp;0.090) (Supplementary Figure. 5 and Supplementary Table\u0026nbsp;12).\u003c/p\u003e \u003cp\u003eThese findings show that (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) professional labels are a poor proxy for AI interaction, and (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) AI collaboration is a distinct capability decoupled from traditional diagnostic skill.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Exploratory identification of behavioral personas\u003c/h2\u003e \u003cp\u003eTo explore these data-driven patterns, we performed k-means clustering on the n\u0026thinsp;=\u0026thinsp;21 pathologists using their AAR (benefit-seeking) and ABR (risk-taking) scores(\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). Standard statistical methods (Elbow and Silhouette) robustly confirmed k\u0026thinsp;=\u0026thinsp;2 as the optimal number of clusters (Supplementary Figure. 4). This analysis revealed two stable, mutually exclusive personas (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB):\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eResistant\u003c/b\u003e (n\u0026thinsp;=\u0026thinsp;7): A skeptical group (low AAR, low ABR).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eReceptive\u003c/b\u003e (n\u0026thinsp;=\u0026thinsp;14): An accepting group (high AAR, high ABR).\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eCrucially, these data-driven personas were entirely independent of traditional expertise labels (χ\u003csup\u003e2\u003c/sup\u003e test, p\u0026thinsp;=\u0026thinsp;0.807)(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). Both the Resistant (2 GU, 2 non-GU, 3 residents) and Receptive (5 GU, 5 Non-GU, 4 Resident) groups contained a balanced mix of all expertise levels (Supplementary Table\u0026nbsp;13). This finding confirms that AI interaction style is a behavioral trait distinct from traditional training(\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e3.7 Exploratory analysis of persona-based performance\u003c/h2\u003e \u003cp\u003eWe analyzed the diagnostic outcomes of these proposed personas. All performance gains were driven by the 'Receptive' group, which showed a statistically significant improvement in Accuracy (p\u0026thinsp;=\u0026thinsp;0.041) and Precision (p\u0026thinsp;=\u0026thinsp;0.017). The 'Resistant' persona showed no significant improvement (Figure. 3A).\u003c/p\u003e \u003cp\u003eHowever, this benefit appeared to be neutralized by an opposing risk profile. An exploratory analysis of the five AI-incorrect cases (n\u0026thinsp;=\u0026thinsp;5), though limited by statistical power, suggested the 'Receptive' group was more susceptible to automation bias (ABR) (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) (Figure. 3B and Supplementary Table\u0026nbsp;14).\u003c/p\u003e \u003cp\u003eThis finding suggests a performance paradox: the Receptive group's gains from AI's correct answers (S2) were cancelled out by their losses from AI's incorrect answers (S3/S4). Consequently, the overall Optimal Decision Rate (ODR) across all 30 cases showed no statistically significant difference between the two personas (p\u0026thinsp;=\u0026thinsp;0.259) (Figure. 3B and Supplementary Table\u0026nbsp;14).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e3.8 Cognitive load analysis of proposed behavioral personas\u003c/h2\u003e \u003cp\u003eFinally, we analyzed cognitive load (reading time) to further explore the potential mechanism behind this paradox(\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e), revealing two different cognitive strategies. Participants had reduced cognitive load when they agreed with AI predictions, as shown in both expertise and persona analyses (Figure. 4, Supplementary Figure. 6, and Supplementary Tables\u0026nbsp;15 and 16)(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFocusing on the robust S2 scenario (n\u0026thinsp;=\u0026thinsp;25 cases, AI\u0026thinsp;=\u0026thinsp;Correct), we found the 'Resistant' group employed a deliberation strategy. They showed no significant time difference between accepting (Optimal) or rejecting (Suboptimal) AI's correct advice (p\u0026thinsp;=\u0026thinsp;0.111), remaining slow and skeptical even when AI was right (Figure. 4B). In contrast, the Receptive group employed an \"Efficiency\" strategy(\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e), being significantly faster (p\u0026thinsp;=\u0026thinsp;0.009) when accepting AI's correct advice versus rejecting it. This \"path of least resistance\" behavior explains their high AAR (Benefit). Exploratory analysis of S3/S4 scenarios (n\u0026thinsp;=\u0026thinsp;5) suggests this same fast, low-effort strategy also explains their high ABR (Risk), as succumbing to bias was cognitively cheaper (Figure. 4C and D).\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eThis study demonstrates that AI assistance can significantly improve diagnostic accuracy (60.0% to 73.3%, p\u0026thinsp;=\u0026thinsp;0.012) for pathologists diagnosing the rare and challenging FHdRCC. This finding aligns with literature on AI-driven accuracy improvements in pathology but extends this benefit to the critical domain of rare diseases, an area of high uncertainty for pathologists. The observed improvement in inter-rater reliability (Fleiss' kappa 0.311 to 0.482) also suggests AI can help standardize diagnoses.\u003c/p\u003e \u003cp\u003eA key finding is that this performance improvement is not correlated with traditional clinical expertise (GU specialist, non-GU, resident). Expertise showed no correlation with AAR (p\u0026thinsp;=\u0026thinsp;0.594), ABR (p\u0026thinsp;=\u0026thinsp;0.690), or ODR (p\u0026thinsp;=\u0026thinsp;0.090). This expert paradox (Supplementary Figure. 3A) revealed that AI-driven accuracy gain was strongly negatively correlated (Pearson R = \u0026minus;\u0026thinsp;0.66) with the pathologist's unassisted baseline accuracy(\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). This suggests AI collaboration is a novel skill set, decoupled from traditional diagnostic competency(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). Clinically, this implies AI can act as a safety net, particularly for non-specialists or when experts face challenging cases(\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOur exploratory analysis identified two data-driven behavioral personas based on interaction patterns: a Receptive (n\u0026thinsp;=\u0026thinsp;14; high AAR, high ABR) and a Resistant (n\u0026thinsp;=\u0026thinsp;7; low AAR, low ABR) group (Figure. 3B). These personas were independent of traditional expertise (p\u0026thinsp;=\u0026thinsp;0.807)(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). This revealed a persona paradox: all significant accuracy gains were driven by the Receptive group (Figure. 4), but this was neutralized by their high susceptibility to automation bias (ABR) (Figure. 3B). The gains from AI's correct answers were offset by losses from its incorrect ones(\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). Consequently, the overall Optimal Decision Rate (ODR) was not significantly different between the groups (p\u0026thinsp;=\u0026thinsp;0.259) (Figure. 3B). This finding moves beyond general warnings about automation bias(\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e) to empirically link it to a specific behavioral profile, suggesting that improving statistical accuracy does not automatically improve overall decision quality.\u003c/p\u003e \u003cp\u003eThis paradox appears driven by different cognitive strategies (Figure. 4). The Receptive group used a fast, efficiency-focused strategy, engaging in cognitive offloading(\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e). They were faster when accepting correct AI advice (S2, p\u0026thinsp;=\u0026thinsp;0.009)(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e), a low-effort strategy that also explains their high ABR. The Resistant group used a slow, deliberation-focused strategy, showing no time difference (p\u0026thinsp;=\u0026thinsp;0.111) as they verified the AI's input. This protected them from bias (low ABR) but limited their gains (low AAR), reflecting a true human-in-the-loop oversight role.\u003c/p\u003e \u003cp\u003eThese exploratory findings suggest AI can serve as a safety net for non-specialists or on rare cases (the expert paradox). However, the persona paradox hypothesizes that a one-size-fits-all deployment is insufficient. Persona-tailored training may be required: Receptive users may need training in healthy skepticism (Beneficial Defiance), while Resistant users may need help building appropriate trust (Beneficial Conversion).\u003c/p\u003e \u003cp\u003eThese findings are exploratory. The primary limitation is statistical power. The automation bias analysis was based on only five AI-incorrect cases (the n\u0026thinsp;=\u0026thinsp;5 problem), rendering the ABR metric statistically unstable. Therefore, the persona paradox is a hypothesis-generating model requiring validation. Other limitations include the artificial reader study environment, generalizability from a single rare disease, and an analysis limited to the AI's binary output, not visual colormap interaction. Future research must validate this model in larger cohorts designed with a robust number of AI-incorrect cases. Additionally, prospective studies evaluating persona-tailored AI systems could determine whether adaptive interfaces improve decision quality.\u003c/p\u003e \u003cp\u003eThis study confirms AI's utility in improving diagnostic accuracy for rare cancers. Critically, it provides a preliminary, exploratory model suggesting interaction is dictated not by traditional expertise but by data-driven behavioral personas\u0026mdash;a decoupling. This model reveals a persona paradox driven by a trade-off between efficiency-focused cognitive offloading and deliberation-focused verification. Future research is essential to validate these exploratory personas and determine if this paradox represents a generalizable challenge in human-AI collaboration.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConsent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Institutional Review Board of Asan Medical Center (IRB No. 2023-0674) and conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the IRB for the retrospective analysis of de-identified archival specimens. Written informed consent was obtained from all participating pathologists in the reader study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDe-identified data, whole-slide images, and model weights are available to qualified researchers for 5 years post-publication upon reasonable request. Requests require a research proposal directed to the corresponding authors ([email protected] or [email protected]), subject to data use agreement and IRB approval (Asan Medical Center IRB No. 2023-0674). The datasets generated and/or analysed during the current study are not publicly available due to patient privacy restrictions but are available from the corresponding authors upon reasonable request with appropriate ethical approvals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe source code used to train and evaluate the deep learning model, along with scripts for statistical analysis, is publicly available at https://github.com/brody9512/FhdRCC_AHI. The code can be accessed without restrictions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors declare no financial or non-financial competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCP and YIL contributed equally as co-first authors; NK and YMC contributed equally as co-corresponding authors. CP and YIL drafted the manuscript. CP developed the software and methodology and performed formal analysis. CP and YIL conducted formal analysis. JS, J-MP, SYY, and BA contributed to investigation, data curation, and resources. NK and YMC provided supervision and funding acquisition. All authors reviewed and approved the final manuscript. CP, YIL, and NK verified the underlying data. All authors had full access to all the data in the study and accept responsibility for the decision to submit for publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Korea Health Industry Development Institute (HR20C0026); National Research Foundation of Korea (RS-2025-00514209); and Asan Institute for Life Sciences, Asan Medical Center (2024IT0001-1).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAhn B, Jeong J, Lee YI, Park J-M, Yoon SY, Song C, et al. Pathologic Diagnosis of Renal Cell Carcinoma in the Era of the 2022 World Health Organization Classification: Key Points for Clinicians. J Urol Oncol. 2024;22(2):115\u0026ndash;27.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWHO Classification of Tumours Editorial Board. Renal cell tumours. Urinary and male genital tumours. WHO Classification of Tumours. 8. 5th ed. Lyon, France: International Agency for Research on Cancer; 2022. p. 32\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi M-Y, Pan Y, Lv Y, Ma H, Sun P-L, Gao H-W. Digital pathology and artificial intelligence in renal cell carcinoma focusing on feature extraction: a literature review. Frontiers in Oncology. 2025;Volume 15\u0026ndash;2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoon SW, Kim J, Kim YJ, Kim SH, An CS, Kim KG, et al. Leveraging explainable AI and large-scale datasets for comprehensive classification of renal histologic types. Scientific Reports. 2025;15(1):1745.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLau HD, Chan E, Fan AC, Kunder CA, Williamson SR, Zhou M, et al. A Clinicopathologic and Molecular Analysis of Fumarate Hydratase-deficient Renal Cell Carcinoma in 32 Patients. The American Journal of Surgical Pathology. 2020;44(1).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun G, Zhang X, Liang J, Pan X, Zhu S, Liu Z, et al. Integrated Molecular Characterization of Fumarate Hydratase\u0026ndash;deficient Renal Cell Carcinoma. Clinical Cancer Research. 2021;27(6):1734\u0026ndash;43.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTrpkov K, Hes O, Agaimy A, Bonert M, Martinek P, Magi-Galluzzi C, et al. Fumarate Hydratase\u0026ndash;deficient Renal Cell Carcinoma Is Strongly Correlated With Fumarate Hydratase Mutation and Hereditary Leiomyomatosis and Renal Cell Carcinoma Syndrome. The American Journal of Surgical Pathology. 2016;40(7).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen DA, Virk RK. Fumarate Hydratase\u0026ndash;Deficient Renal Cell Carcinoma: A Review. AJSP: Reviews \u0026amp; Reports. 2020;25(6):280\u0026ndash;3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSung JW, Lee YI, Kim Y, Song C, Park J-M, Yoon SY, et al. Papillary renal cell carcinoma revisited: impact of the World Health Organization 2022 classification on prognostication. BJU International. 2025;135(3):510\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang X, Liu Y, Wang H, Xu Y, Zhang H, Zhao M, et al. Fumarate Hydratase\u0026ndash;Deficient Renal Cell Carcinoma With Predominant Tubulocystic Features Mimics Tubulocystic Renal Cell Carcinoma. Archives of Pathology \u0026amp; Laboratory Medicine. 2024;148(12):1358\u0026ndash;64.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOhe C, Smith SC, Sirohi D, Divatia M, de Peralta-Venturina M, Paner GP, et al. Reappraisal of Morphologic Differences Between Renal Medullary Carcinoma, Collecting Duct Carcinoma, and Fumarate Hydratase\u0026ndash;deficient Renal Cell Carcinoma. The American Journal of Surgical Pathology. 2018;42(3):279\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAggarwal A, Bharadwaj S, Corredor G, Pathak T, Badve S, Madabhushi A. Artificial intelligence in digital pathology \u0026mdash; time for a reality check. Nature Reviews Clinical Oncology. 2025;22(4):283\u0026ndash;91.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDistante A, Marandino L, Bertolo R, Ingels A, Pavan N, Pecoraro A, et al. Artificial Intelligence in Renal Cell Carcinoma Histopathology: Current Applications and Future Perspectives. Diagnostics. 2023;13(13):2294.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsif A, Rajpoot K, Graham S, Snead D, Minhas F, Rajpoot N. Unleashing the potential of AI for pathology: challenges and recommendations. The Journal of Pathology. 2023;260(5):564\u0026ndash;77.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBod\u0026eacute;n ACS, Molin J, Garvin S, West RA, Lundstr\u0026ouml;m C, Treanor D. The human-in-the-loop: an evaluation of pathologists\u0026rsquo; interaction with artificial intelligence in clinical practice. Histopathology. 2021;79(2):210\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWekenborg MK, Gilbert S, Kather JN. Examining human-AI interaction in real-world healthcare beyond the laboratory. npj Digital Medicine. 2025;8(1):169.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2011;19(1):121\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaube S, Suresh H, Raue M, Merritt A, Berkowitz SJ, Lermer E, et al. Do as AI say: susceptibility in deployment of clinical decision-aids. npj Digital Medicine. 2021;4(1):31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTschandl P, Rinner C, Apalla Z, Argenziano G, Codella N, Halpern A, et al. Human\u0026ndash;computer collaboration for skin cancer recognition. Nature Medicine. 2020;26(8):1229\u0026ndash;34.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBansal G, Wu T, Zhou J, Fok R, Nushi B, Kamar E, et al., editors. Does the whole exceed its parts? the effect of ai explanations on complementary team performance. Proceedings of the 2021 CHI conference on human factors in computing systems; 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRainey C, Bond R, McConnell J, Gill A, Hughes C, Kumar D, et al. The impact of AI feedback on the accuracy of diagnosis, decision switching and trust in radiography. PLOS ONE. 2025;20(5):e0322051.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2012;19(1):121\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKodinariya TM, Makwana PR. Review on determining number of Cluster in K-Means Clustering. International Journal. 2013;1(6):90\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRahwan I, Cebrian M, Obradovich N, Bongard J, Bonnefon J-F, Breazeal C, et al. Machine behaviour. Nature. 2019;568(7753):477\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJacobs M, Pradier MF, McCoy TH, Perlis RH, Doshi-Velez F, Gajos KZ. How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection. Translational Psychiatry. 2021;11(1):108.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSweller J. Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science. 1988;12(2):257\u0026ndash;85.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKahneman D. Fast and slow thinking. Allen Lane and Penguin Books, New York. 2011.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEl-Khoury R, Zaatari G. The Rise of AI-Assisted Diagnosis: Will Pathologists Be Partners or Bystanders? Diagnostics. 2025;15(18):2308.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRosbach E, Ganz J, Ammeling J, Riener A, Aubreville M, editors. Automation Bias in AI-assisted Medical Decision-making under Time Pressure in Computational Pathology. BVM Workshop; 2025: Springer.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBu\u0026ccedil;inca Z, Malaya MB, Gajos KZ. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proc ACM Hum-Comput Interact. 2021;5(CSCW1):Article 188.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Fumarate hydratase-deficient renal cell carcinoma, Deep learning, Human-AI interaction, Automation bias, Behavioral personas, Digital pathology","lastPublishedDoi":"10.21203/rs.3.rs-8343823/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8343823/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eFumarate Hydratase-deficient Renal Cell Carcinoma (FHdRCC) is a rare (\u0026lt;\u0026thinsp;0.4% of RCCs), aggressive subtype with significant morphological overlap, posing diagnostic challenges. While artificial intelligence (AI) shows promise in common cancer diagnostics, its impact on pathologist decision-making in rare diseases\u0026mdash;particularly concerning automation bias\u0026mdash;remains poorly understood.\u003c/p\u003e \u003cp\u003eWe developed a deep learning model to classify FHdRCC. We conducted a crossover reader study with 21 pathologists (7 genitourinary (GU) specialists, 7 non-GU specialists, 7 residents) diagnosing 30 challenging cases (15 FHdRCC, 15 non-FHdRCC) with and without AI assistance. We analyzed diagnostic performance and performed an exploratory analysis of human-AI interaction by quantifying AI Acceptance Rate (AAR) and Automation Bias Rate (ABR)\u0026mdash;the rates of following AI recommendations when correct or incorrect, respectively\u0026mdash;leading to the identification of preliminary behavioral personas.\u003c/p\u003e \u003cp\u003eAI assistance significantly improved diagnostic accuracy (60.0% to 73.3%, p\u0026thinsp;=\u0026thinsp;0.012) and inter-rater reliability (Fleiss' κ from 0.311 to 0.482, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). AI-driven gains were negatively correlated with baseline expertise (R=-0.66, p\u0026thinsp;=\u0026thinsp;0.001), revealing independence from traditional training. Clustering identified two behavioral personas: Receptive (n\u0026thinsp;=\u0026thinsp;14; high AAR/ABR) employing efficiency-focused strategies, and Resistant (n\u0026thinsp;=\u0026thinsp;7; low AAR/ABR) using deliberation-focused approaches. While the Receptive group drove accuracy gains (p\u0026thinsp;=\u0026thinsp;0.041), high automation bias neutralized improvements in overall optimal decision-making (p\u0026thinsp;=\u0026thinsp;0.259).\u003c/p\u003e \u003cp\u003eAI assistance enhances rare cancer diagnostic accuracy, but effectiveness is mediated by behavioral personas rather than traditional expertise. The performance paradox\u0026mdash;accuracy gains offset by automation bias\u0026mdash;suggests persona-tailored training is essential: Receptive users need critical evaluation skills; Resistant users need trust-building. These exploratory findings require validation in larger studies with sufficient power to characterize automation bias patterns.\u003c/p\u003e","manuscriptTitle":"One AI Training Fits All? Exploring Behavioral Personas in Rare Cancer Diagnosis— Fumarate Hydratase-Deficient Renal Cell Carcinoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-08 14:18:12","doi":"10.21203/rs.3.rs-8343823/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-21T16:57:45+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-19T22:53:16+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-19T20:41:07+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-07T09:19:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"43295420930407644974883125245889004599","date":"2026-01-07T01:09:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"286281287802425919178575954790052778072","date":"2026-01-06T04:20:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"158538076427706558052314251729987688663","date":"2026-01-05T18:07:29+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-05T17:59:03+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-12-18T01:01:27+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-12-17T11:11:38+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Digital Medicine","date":"2025-12-12T08:44:11+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"cca4e8f7-df95-4b33-bef1-b9fb859c4787","owner":[],"postedDate":"January 8th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":60693754,"name":"Biological sciences/Cancer"},{"id":60693755,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":60693756,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-05-13T02:25:53+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-08 14:18:12","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8343823","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8343823","identity":"rs-8343823","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}