Citations

Background Predictive models employing machine learning algorithms are increasingly being used in clinical decision making, and improperly calibrated models can result in systematic harm. We sought to investigate the impact of class imbalance correction, a commonly applied preprocessing step in machine learning model development, on calibration and modelled clinical decision making in a large real-world context. Methods A histogram boosted gradient classifier was trained on a highly imbalanced national dataset of >1.8 million patients undergoing surgery, to predict the risk of 90-day mortality and complications after surgery. Class imbalance correction strategies including random oversampling, synthetic minority oversampling technique, random under-sampling, and cost-sensitive learning were compared to the natural distribution ('natural'). Models were tested and compared with classification metrics, calibration plots, decision curve analysis, and simulated clinical impact analysis. Results The natural model demonstrated high performance (AUROC 0.94, 95% CI 0.94--0.95 for mortality; 0.84, 95% CI 0.84--0.85 for complications) and calibration (log loss 0.05, 95% CI 0.04--0.05 for mortality; 0.23, 95% CI 0.23--0.24 for complications). Class imbalance mitigation (CSL, ROS, RUS, and SMOTE) did not improve AUROC or AUPRC but increased recall and F1 scores at the expense of precision and accuracy. However, these methods severely compromised model calibration, leading to significant over-prediction of risks (up to a 62.8 % increase) as further evidenced by increased log loss across all mitigation techniques. Decision curve analysis and clinical scenario testing confirmed that the natural model provided the highest net benefit. Conclusion Class imbalance correction methods result in significant miscalibration, leading to possible harm whenused for clinical decision making.

57 Background: Cardiovascular disease (CVD) is the leading non-cancer cause of death among men with prostate cancer (PC), particularly during androgen-deprivation therapy. While CVD risk-prediction models are well established in the general population, barriers to integration within oncology workflows remain poorly defined. Guided by the PRISM and RE-AIM frameworks, this study evaluated multilevel determinants influencing adoption, feasibility, and sustainability of CVD risk-assessment tools in PC care. Methods: A national, cross-sectional survey of 45 physicians (medical oncology 49%, urology 38%, radiation oncology 13%) from 18 U.S. states was conducted. Respondents represented early- (27%), mid- (36%), and late-career (38%) clinicians across academic (87%) and community (13%) settings. Quantitative items mapped to PRISM/RE-AIM domains assessed adaptability, cost/resource feasibility, effectiveness, equity, and sustainability; open-ended responses explored contextual barriers and facilitators. Descriptive statistics summarized quantitative data. Results: Physicians reported moderate-to-high readiness to adopt a CVD risk-assessment tool, emphasizing workflow alignment and EHR integration over algorithmic complexity. Table 1 summarizes domain-specific findings. Ninety-three percent of respondents favored lab-independent models that minimize workflow disruption. Conclusions: Pre-implementation analysis highlighted that structural barriers, rather than algorithmic or statistical limitations, remain the primary obstacles to adopting CVD risk-assessment tools in prostate-cancer care. Physicians emphasized the need for EHR-integrated, non-laboratory-based, and time-efficient tools that align with routine oncology workflows and minimize cross-specialty friction. Addressing these practical and organizational determinants may be pivotal to achieving equitable, scalable, and sustainable cardio-oncology implementation across diverse practice settings. PRISM and RE-AIM constructs evaluation results. Construct Likely/Very Likely (%) Interpretation Perceived Effectiveness 87% Broad confidence in clinical utility Equity in Patient Reach 75% Strong perceived equity in reach Equity in Effectiveness 73% Expected to perform consistently across populations Adaptability to Workflow 71% High compatibility with existing clinic workflows Cost/Resource Feasibility 65% Feasible for resource-constrained settings Likelihood of Adoption 60% Moderate readiness for implementation Equity in Adoption 58% Equitable uptake possible despite resource differences Reach to Target Population 57% Moderate likelihood of reaching most patients Sustainability Over Time 51% Moderate long-term feasibility Consistency of Implementation 42% Need for standardized protocols

476 Background: Lack of standardized reporting and collection of patient-reported outcomes (PROs) limits cross-trial comparisons of first-line IO-based regimens for mRCC. Prospectively collected real world data may fill this important knowledge gap. Pts with stable/improved QoL at 3 mo versus baseline have longer median OS than pts with worse/unobserved HRQoL versus baseline (Cella, 2024). Methods: ODYSSEY is a multi-site, prospective observational study of 468 US pts with mRCC. Pts must have mRCC (any histology), no prior systemic therapy (ST), and follow up at a PCORnet study site. Exclusion criteria include treatment for cancer except mRCC. Descriptive statistics are used to show PRO score changes from baseline. We evaluate longitudinal change from baseline in PRO scores using least squares mean (LSM). Minimally important differences (MID) are 3 points for FKSI-19 total score, 1 point for the FKSI-Disease Related Symptoms (DRS) subscale and 7 points for FACT-G. Higher score indicates better QOL. The primary objective is to determine patterns of change in QOL and symptom burden of pts with mRCC. Results: As of 10/1/25, 360 pts were managed with ST, including 266 pts on IO-based combinations: 35 cabozantinib + nivolumab (C+N), 49 lenvatinib + pembrolizumab (L+P), 40 axitinib + pembrolizumab (A+P), and 142 ipilimumab + nivolumab (I+N). Overall median follow-up is 12.1 mos (IQR 6, 24). Age, sex and race were similar across regimens. Prior nephrectomy (50%) was lower than in trials. C+N and L+P pts had less clear cell (56 and 65%) versus A+P and I+N pts (94 and 82%). A+P pts had more favorable IMDC risk (19%) and less poor risk (14%). Baseline QOL was variable across regimens, with L+P generally the lowest. Absolute changes in PROs from enrollment to mo 3 (early change) by regimen are shown in the Table. LSM change from baseline in FKSI-19 and FKSI-DRS favored IO-IO versus IO-TKI at 3 mos but not at 6 mos (not shown). Conclusions: While early change in PROs did not meet the MID threshold between the IO-TKI and I+N cohorts, different patterns of early change are apparent among specific IO-TKI regimens. IO-TKI regimens exhibited higher variation in early PRO change vs I+N. As the data matures, we will associate baseline PROs and early PRO changes with early death and other outcomes. Clinical trial information: NCT04919122 . 3 month change. Instrument C+N (N=35) L+P (N=49) A+P (N=40) All IO-TKI(N=124) I+N(N=142) FKSI-19, N (%) 26 (74) 33 (67) 33 (83) 92 (74) 104 (73) Mean (SD) -0.3 (11) 2.8 (13) -3.2 (12) -0.2 (12) 0.7 (10) Median (IQR) 1.4 (-7, 6) 5.0 (-7, 11) -1.0 (-14, 5) 1.4 (-10, 8) -0.3 (-5, 6) FKSI-DRS, N (%) 26 (74) 33 (67) 33 (83) 92 (74) 104 (73) Mean (SD) -0.8 (5) 1.5 (7) -1.8 (6) -0.3 (6) 0.8 (5) Median (IQR) -0.5 (-5, 3) 2.0 (-5, 5) -1.0 (-5, 3) 0.0 (-5, 4) 0.0 (-2, 4) FACT-G, N (%) 27 (77) 32 (65) 29 (73) 88 (71) 102 (72) Mean (SD) 2.1 (14) 5.2 (16) -3.5 (15) 1.4 (15) 0.3 (13) Median (IQR) 4.0 (-4, 10) 1.2 (-6, 13) -6.4 (-11, 4) 1.6 (-8, 10) 1.6 (-7, 8)

Abstract Pancreatic ductal adenocarcinoma (PDAC) is unresponsive to immune checkpoint therapies largely due to immunosuppression mediated by myeloid and lymphoid cells. Effective treatment strategies require, in addition to immune stimulation, the depletion of these immunosuppressive cells and their interactions. Here, we demonstrate for the first time, that a single verteporfin-based photodynamic therapy (PDT) treatment triples median survival in a KPC-PDAC model when combined with anti-PD1 therapy. Mechanistically, apart from reducing CK19 + tumor cells, PDT depletes immunosuppressive myeloid (CD11c+) and lymphoid (FOXP3+) cell populations, by ~ 4- and ~ 18.4-folds, respectively. PDT further disrupts interactions between these immunosuppressive cells, attenuating a key immune-tolerance axis. At the systemic level, PDT reduces splenic immunosuppression through modulation of tumor-derived cytokines. Within this reprogrammed immune-landscape, anti–PD-1 therapy elicits robust CD8 + T-cell responses, reflected by an ~ 6.7-fold increase in the intratumoral CD8+/Treg ratio. Collectively, these findings establish verteporfin-PDT as an immune-priming strategy that complements anti–PD-1 therapy by alleviating immunosuppression in PDAC.

An understanding of regions of endemicity and the spectrum of organ system involvement can improve recognition of the imaging features of coccidioidomycosis, which can overlap with those of other diseases.

Introduction: Optimizing patient selection and time-to-intervention will be crucial to the success of future clinical trials seeking to implement therapies and improve intracerebral hemorrhage (ICH) patient outcomes. Automated and accurate ICH and perihematomal edema (PHE) volumetry carries significant potential to strengthen trials targeting ICH and PHE expansion. Aims: The Quantification of Hematoma and Perihematomal Edema Volumes in Intracerebral Hemorrhage (QUANTUM) trial was designed with aims of (1) validating a fully automated artificial intelligence (AI) method for supratentorial, spontaneous ICH and PHE volumetry against the manual and semiautomated quantification methods and (2) establishing a model framework for the translation of AI technology through conduct of a clinical trial under the equivalence hypothesis paradigm. Methods: This trial was a prospective, randomized, clinical study with blinded outcome assessment. Six independent raters were randomized in a 1:1 ratio to ICH and PHE volumetry from non-contrast CT scans in the Virtual International Stroke Trials Archive with either the semiautomated or manual quantification method. The primary outcomes were equivalence between (1)ICH volume measurements on admission (<24h from symptom onset) for the manual versus the fully automated or the sem-automated quantification methods, and (2)PHE volume measurements on day 3 (72±12h from symptom onset) for the manual versus the fully automated or the semiautomated quantification methods. The required sample size was 126 ICH and 126 PHE measurements, based upon a significance level of 0.025 to account for the two equivalence tests. Results: The geometric mean relative discrepancies between the fully automated and the manual ICH and PHE volume measurements did not cross the pre-specified 10% equivalence margin. The geometric mean relative discrepancies between the semiautomated and the manual ICH and PHE volume measurements did cross the pre-specified 10% equivalence margin. The fully automated method (mean 10.0±2.7s/scan) was significantly faster than the manual (mean 189±199 s/scan; P <0.001) and semiautomated (mean 79±68s/scan; P <0.001) methods. Conclusions: ICH and PHE volumes quantified by AI could replace clinician measurements within a 10% error margin and with substantially greater efficiency. The semiautomated method could not replace manual measurements. This trial demonstrated feasibility of the equivalence hypothesis framework for AI validation studies.

Abstract Background Ten years after the original publication, the Centiloid framework is now broadly used to harmonize amyloid‐PET quantification, facilitate data sharing and comparison across cohorts, and even assist visual interpretation in clinical settings. We evaluated the global implementation of Centiloids by comparing their distribution and corresponding positivity thresholds across cohorts. Methods We gathered data from publicly available cohorts and reached out to investigators across the world to collect cross‐sectional Centiloids, demographic and clinical information, and visual read data. Gaussian mixture models (GMM, k=2) were fitted to Centiloid values for each cohort and cutoffs were calculated as mean + 2SD of the lower Gaussian. When visual reads were available, we determined Centiloid cutoffs that maximized correspondence with visual reads (Cohen's kappa). Data was combined across cohorts using random effects meta‐analyses. Results As of January 2025, we included 37 cohorts ( n = 41,678 participants) with heterogeneous pipelines, radiotracers, and clinical and demographic characteristics (Table‐1). The low Gaussian peaks ranged from ‐9 to 10CL; the meta‐analysis identified a common peak at 1CL. The second peaks were more heterogeneous (range=38‐102CL; meta‐analysis outcome=64CL). Across cohorts, the proportion of cognitively unimpaired versus impaired participants impacted the position of both peaks, with better separation in cohorts enriched in impaired individuals (Figure 1C). The meta‐analysis indicated a GMM‐based cutoff of 19CL (95%CI: 16‐21CL, Figure 1B); subgroup analyses showed no evidence of significant effect between single versus multicenter settings (17 versus 20CL, p = 0.30), MRI‐based or PET‐only processing (18 versus 19CL, p = 0.88), and no evidence of difference across radiotracers (Flutemetamol: 16CL; PIB: 17CL, Flutafuranol: 18CL, Florbetaben: 19CL, Florbetapir: 20CL, p = 0.78). In a subset of 29,496 participants with visual reads available, binary visual reads corresponded well to Centiloids (common kappa=0.86, Figure 2A). The visual read‐based cutoff of 24CL (95%CI: 21‐27CL, Figure 2B) maximized correspondence between visual read and quantification and was slightly higher than the GMM‐based cutoff. All meta‐analysis models showed high non‐random heterogeneity (I 2 >80%) across studies, suggesting non‐random differences in peaks and cutoffs. Conclusions Meta‐analysis‐based cutoffs align well with thresholds from the existing literature. High heterogeneity among studies underscores the need to investigate contributing factors, raising concerns about applying common cutoffs.