Trevor Joseph Royce

PURPOSE: Metastatic hormone-sensitive prostate cancer (mHSPC) is a heterogeneous disease state with multiple treatment options. Biomarkers are needed for risk stratification and to guide treatment intensification or deintensification. We used a computational histology artificial intelligence–based platform (CHAI) to develop and validate a digital image–only prognostic biomarker in mHSPC with participant-level data from two prospective, phase III randomized controlled trials. METHODS: The CHAI platform extracted quantitative histomorphologic features from whole-slide images of hematoxylin and eosin–stained diagnostic specimens. Data from CHAARTED were used to construct a signature of features associated with overall survival (OS). A continuous risk score was dichotomized into favorable- and unfavorable-risk groups. The performance of the locked model was assessed in ENZAMET as an independent validation cohort using Kaplan-Meier methods and multivariable Cox proportional hazards models. Institutional review board approval was obtained for each participating data set. Given all patient information was deidentified, the study was considered institutional review board–exempt and consent waived. RESULTS: Overall, 1,191 participants were included: 507 in development (CHAARTED) and 684 in validation (ENZAMET). In the validation cohort, CHAI unfavorable-risk participants had worse OS (hazard ratio [HR], 2.6 [95% CI, 2.0 to 3.4]; P < .001) and progression-free survival (PFS; HR, 2.2 [95% CI, 1.7 to 2.7]; P < .001). Five-year OS was 39% for unfavorable-risk versus 69% for favorable-risk participants. In adjusting for clinical variables, the biomarker remained prognostic for OS and PFS ( P < .001). Among those who received doublet (androgen-deprivation therapy [ADT] plus androgen receptor pathway inhibitor [ARPI]) versus triplet systemic therapy (ADT plus docetaxel plus ARPI), there was a significant interaction between biomarker and treatment intensification for OS ( P = .003) and PFS ( P < .001), suggesting that patients at higher risk may benefit more from triplet therapy. CONCLUSION: We developed and validated an image-only AI-based biomarker associated with clinical outcomes and potential benefit from treatment escalation in mHSPC independent of conventional clinicopathologic risk factors.

Demographic and clinical characteristics of the UPCA referral cohort.

Multimodal artificial intelligence (MMAI) biomarker score distribution by A) NCCN and B) CAPRA risk groups.

Univariable analyses of multimodal artificial intelligence (MMAI) score for biochemical recurrence (BCR) and adverse pathology (AP) at radical prostatectomy within NCCN Subgroup.

Comparison of multimodal artificial intelligence (MMAI) score to individual clinical variables, NCCN risk groups, and CAPRA score.

PURPOSE: Predictive biomarkers to guide selection of first-line chemotherapy for advanced pancreatic ductal adenocarcinoma (PDAC) are an unmet clinical need. This study used the Computational Histology Artificial Intelligence (CHAI) platform to develop and validate a histomorphology-based G-chemo versus F-chemo (GvF) biomarker that predicts benefit from first-line fluoropyrimidine-based (F-chemo) versus gemcitabine-based (G-chemo) regimens. METHODS: The CHAI platform extracted quantitative histomorphologic features from whole-slide images of hematoxylin and eosin–stained diagnostic biopsies. In a multi-institutional development cohort, features associated with differential outcomes as measured by time to next treatment or death (TNTD) between F-chemo–treated and G-chemo–treated patients produced continuous biomarker scores, which were dichotomized into G-pref or F-pref results. The biomarker and threshold were locked. An independent validation cohort from the prospective COMPASS and Know Your Tumor studies assessed differential treatment outcomes by TNTD and overall survival (OS). RESULTS: There were 477 patients (development: 178; validation: 299). In validation, among 173 F-pref patients, those treated with F-chemo had significantly better outcomes than G-chemo for both TNTD ( P = .035; median TNTD: F-chemo 8.6 months; G-chemo 7.5 months) and OS ( P = .003; median OS: F-chemo 14.4 months; G-chemo 11.7 months). Among 126 G-pref patients, G-chemo had significantly superior TNTD ( P = .038; median TNTD: F-chemo 7.2 months; G-chemo 9.6 months), but no difference in OS ( P = .5; median OS: F-chemo 12.4 months; G-chemo 14.3 months). In propensity score–weighted analysis, the biomarker predicted treatment effect (biomarker-treatment interaction TNTD P < .001; OS P = .005). RNA subtypes were associated with TNTD and OS but did not predict differential treatment effects ( P = .3). CONCLUSION: The histomorphology-based GvF biomarker predicted differential treatment benefit of first-line GvF. This biomarker can guide optimal treatment selection for first-line therapy in advanced PDAC.

PURPOSE: Large language models (LLMs) show promise in assisting knowledge-intensive fields such as oncology, where up-to-date information and multidisciplinary expertise are critical. Traditional LLMs risk hallucinations and reliance on static, possibly outdated data that lack domain-specific context. Retrieval-augmented generation (RAG) has emerged as a strategy to address these issues by incorporating domain-specific information from external knowledge repositories. METHODS: We evaluated 15 LLMs, including Meta Llama-2/3, generative pretrained transformer (GPT)-3.5/4/4o variants, Claude-3, Gemini-2.0, and DeepSeek-R1. In a zero-shot workflow, each LLM answered 298 scorable questions from the 2021 American College of Radiology in-training examination. We implemented a RAG pipeline (Iridium Model) that transforms user prompts into vector embeddings, queries a specialized radiation oncology database, and merges relevant text with the original prompt to form an augmented query. We compared zero-shot versus RAG-augmented performance. RESULTS: Larger-parameter LLMs had higher zero-shot accuracy, with six models outscoring graduating residents ( P < .01). Top scorers were reasoning models GPT-4o1, o3-mini, and DeepSeek-R1, which achieved 91.6%, 86.6%, and 91.6% without RAG, respectively. Gemini-2.0 improved 6.7% (to 79.2%), Llama-3-70b 8.4% (to 75.8%), and GPT-4o 5.7% (to 85.6%) with RAG. Top scoring reasoning models surpassed graduating resident averages by 17.7%-20% ( P < .01), but had no improvement or detriment with RAG. Domain-specific gains occurred in clinical, biology, and physics. Majority voting boosted aggregate accuracy when individual model performance exceeded 50%. RAG workflows and reasoning models incurred higher computational costs. CONCLUSION: Radiation-oncology–specific retrieval-augmented generation pipeline enhances nonreasoning LLM performance in radiation oncology by integrating domain-specific evidence, whereas it does not improve performance of reasoning models. These findings demonstrate that RAG can elevate clinical decision support by enabling simpler, cost-effective nonreasoning models to tackle complex tasks through retrieval capabilities—an efficient alternative to extensive model training that also yields citable, evidence-based explanations.