Ayub, Umair

5108 Background: We previously proposed a hybrid framework combining LLMs and rule based algorithm (RBA) for automating risk stratification in locPCa. Herein, we aim to validate the risk stratification agent (RSA) in a prospective cohort, develop & evaluate the treatment recommendation agent (TRA) based on NCCN guidelines, and develop an interactive interface to facilitate clinicians for accurate risk stratification and treatment recommendations at the point of care. Methods: This study included pts with locPCa (2004-2024) presenting at Mayo Clinic with at least 1 positive prostate biopsy and MRI report available. For RSA prospective validation, GPT4 extracted key phenotypic variables (PSA, T stage, prostate volume, number of cores, Gleason patterns, grade group) from unstructured MRI and biopsy reports using a zeroshot prompt. An RBA then classified pts into NCCN risk groups. The agent performance was compared with the treating clinician's documentation and evaluated against gold-standard labels manually annotated by two independent clinicians. For development of TRA, two experiments were performed using GPT4 with and without retrieval-augmented generation (RAG) to generate treatment plans. Generated treatment plans were evaluated using NCCN guideline-informed treatment decision tree based algorithm (DTA). Evaluation metrics, weighted – accuracy (acc) and F1, were computed. A clinician facing interface (lisr.org/risk) was developed to provide accurate risk stratification and treatment plans. Results: A total of 858 pts were included (500 for prospective validation, 358 for treatment recommendations). Prospective validation for RSA demonstrated a higher F1 score of 0.89 compared to the treating clinician (F1: 0.58). Treatment recommendation experiments showed that GPT4 with RAG achieved higher acc (64% full – all correct treatment options and 36% partial – at least 1 correct treatment option) compared to GPT4 alone (35% and 65%, respectively). Sensitivity analysis using DTA-informed GPT4 note generation achieved 94% full acc and 6% partial acc. GPT4 alone generated hallucinated treatment options in 71% of cases, while GPT4 with RAG reduced this to 32%. Conclusions: The multi-agent framework based on LLM and RBA achieves high accuracy in risk stratification and treatment recommendation for locPCa. A multi-agent framework with an interactive interface holds high promise to enable efficient, accurate decision-making, and improve locPCa management at the point of care Performance evaluation. RSA(acc; F1) % TRA(full acc; partial acc) % Clinician RSA* GPT4 RAG+GPT4 DTA+GPT4 Overall 60; 58 89; 89 29; 71 64; 36 94; 6 V. High 94; 33 96; 96 55; 45 69; 31 93; 7 High 46; 52 92; 89 - - - Int. Unf 80; 61 91; 90 5; 95 38; 62 93; 7 Int. Fav 57; 64 84; 87 0; 100 98; 2 100; 0 Low 57; 65 87; 80 100; 0 95; 5 90; 10 V. Low 8; 13 50; 67^ - - - *Statistical significance (p<0.01) except for ^V. Low.

e17155 Background: Localized PCa is classified into six NCCN risk categories, each with specific treatment recommendations. However, treatment concordance across these categories and its variation by race in clinical practice remain unexplored. Therefore, we aim to describe patterns of guideline-concordant treatment receipt for localized PCa across risk groups and race. Methods: Surveillance, Epidemiology, and End Results (SEER) database (2018–2020) was queried to identify patients with localized PCa from 17 registries. Data regarding clinical T stage, Gleason patterns, PSA level, biopsy cores and initial treatment received (radiation, radical prostatectomy, systemic therapy) was obtained. Each patient was then stratified into one of six NCCN-based risk categories using a rule-based algorithm. Using the stratified risk category of each patient, treatment recommendations were obtained from the NCCN guidelines. Concordance was assessed by comparing the received treatment and guideline-recommended treatment. A focused bi-variate logistic regression analysis was performed adjusting for risk groups, and race. Odds ratio (OR) with 95% confidence intervals (CI) were computed to quantify the magnitude of guideline-discordant treatment receipt. Results: This analysis included 69,584 patients with localized PCa (White: 53,875 [77.4%]; Black: 11,265 [16.2%]; Asian/Pacific Islander: 4,133 [5.9%]; American Indian/Alaska Native: 311 0.5%]). In terms of risk groups, patients in higher risk categories had increased odds of discordance compared to the lower risk groups (Table). In terms of race, Black men had significantly higher odds of receiving guideline-discordant therapy (OR:1.67, 95% CI:1.58–1.76), while no statistically significant difference was observed with American Indian/Alaska Native men (OR:1.21, 95% CI: 0.92–1.61) and Asian/Pacific Islander men (OR:1.05, 95% CI:0.97–1.14) compared to White men. Conclusions: Population level data, based on limited sample size in some groups, suggests that treatment concordance with NCCN guidelines varies by risk group and race in localized PCa. Black patients, and those with higher risk groups were more likely to receive guideline-discordant therapy. These findings highlight the need for strategies and prospective studies at individual patient level to evaluate and enhance guideline adherence to promote personalized care for these patients. Variable Discordant Treatment (%) OR (95% CI) White 22168 (41.1) REF Black 5653 (47.5) 1.67 (1.58-1.76) * Asian/Pacific Islander 1822 (44.1) 1.05 (0.97-1.14) American Indian/Alaska Native 191 (61.4) 1.22 (0.92- 1.61) Very Low Risk 770 (9) REF Low Risk 146 (2) 0.16 (0.13-0.19) * Favorable Risk 35 (0.4) 0.03 (0.02-0.04) * Unfavorable Risk 15125 (83.2) 36.30 (33.72-39.25) * High/Very High Risk 13441 (50) 7.32 (6.83-7.86) * * p <0.05.

261 Background: Clinicians use next-generation sequencing (NGS) to tailor targeted treatments for PCa patients. However, the results are often available in unstructured formats and are saturated with technical jargon, making them inaccessible for patients. Therefore, we aimed to assess the efficacy of LLMs to accurately extract genomic characteristics from unstructured reports and generate patient-friendly summaries for ease of variant interpretation in PCa patients. Methods: This retrospective study included patients with PCa who underwent NGS from 2023-2024. The GPT4 model was utilized using structured zero-shot prompts to extract genomic characteristics from unstructured reports in a non-decomposed manner. Prompt development was conducted iteratively using a 2% random sample of the dataset. Genomic characteristics ascertained by LLM were assessed against held out vendor-curated test dataset for evaluation. Performance was assessed using weighted evaluation metrics (precision, recall, and F1-score). Additionally, the extracted variables were organized using rule-based criterion and presented as a separate prompt to GPT4 to generate summaries for interpretation at the level of each biologically relevant variant. Mean Flesch-Kincaid scores and Vocd-D statistic with standard deviations were computed to assess readability and lexical diversity of LLM-generated variant interpretation summaries. Results: A total of 331 patients (370 NGS reports) were included in the study. Evaluation on the held-out test dataset (Table) showed that the weighted average precision ranged from 80% for extracting tumor mutational burden to 100% for extracting the CtDNA fraction from unstructured reports. The weighted average recall ranged from 85% for extracting CtDNA tumor fraction to 99% for ascertaining variant/alteration. The weighted average F1-scores varied from 0.83 for extracting tumor mutational burden to 1.0 for extracting specimen site from NGS reports. The mean Flesch-Kincaid score was 8.29 ± 1 indicating average/standard readability ease. The mean Vocd-D statistic was 7.55 ± 1 indicating focused summaries, however, with limited lexical diversity. Conclusions: This study suggests that large language models can effectively extract genomic characteristics from unstructured reports and can potentially improve genomic literacy among prostate cancer patients by providing easy-to-interpret variant summaries. However, targeted prompting may be required to increase lexical diversity for improved engagement. Performance across different genomic characteristics. Category Precision Recall F1-Score Altered gene 0.86 0.97 0.89 Variant/alteration type 0.98 0.99 0.98 Variant allele fraction 0.96 0.96 0.96 Test name/type 1 0.99 1 Specimen site 1 0.99 1 Tumor mutational burden 0.8 0.93 0.83 CtDNA tumor fraction 1 0.85 0.92 MSI status 0.99 0.97 0.98

136 Background: The ECOG performance status (PS) is a critical measure for evaluating functional capacity and determining clinical trial eligibility in patients with advanced prostate cancer. We aim to assess the capability of large language models (LLMs) to accurately infer ECOG PS from unstructured oncology notes. Methods: This retrospective study included patients with metastatic castration-resistant prostate cancer (mCRPC) receiving Lutetium-177–PSMA-617 ( 177 Lu) at the Mayo Clinic (2022-2023). Baseline ECOG PS scores were manually curated by a trained clinician from clinical notes recorded within four weeks prior to 177 Lu initiation. Structured zero-shot prompts were then used to prompt GPT-4 to infer ECOG PS scores from unstructured oncology notes. Total dataset was randomly sampled to prompt development (20%) and held-out test set (80%). Three experimental setups were employed to assess model performance: (i) utilizing unmasked clinical notes, (ii) clinical notes with only numerical ECOG PS scores masked while retaining narrative descriptive/assessment by oncologist of ECOG score, and (iii) clinical notes with both numerical scores and narrative ECOG descriptors masked. Final performance was assessed using evaluation metrics (precision, recall, F1 score) on the held-out test set. Results: A total of 240 patients (n: 50 prompt development; n: 190 test set) were included in the analysis. The median age was 70 (IQR: 66-76); a majority of patients were White (n: 229; 95%) and non-Hispanic (n: 233; 97%). The most prevalent ECOG PS score was 0 (n: 140; 58%), followed by 1 (n: 88; 37%), and 2 (n: 12; 5%). The performance was numerically higher when using unmasked notes (weighted average precision: 0.89; recall: 0.89; F1: 0.89), compared to using notes with masked scores only (weighted average precision: 0.76; recall: 0.79; F1: 0.75) and notes with both scores and narrative masked (weighted average precision: 0.57; recall: 0.53; F1: 0.48). Metrics by each ECOG PS score is outlined (Table). Conclusions: Large language models demonstrate significant potential in precisely extracting ECOG performance status from unstructured oncology notes, offering valuable support for treatment decision-making and clinical trial enrollment. However, model performance declines as reliance on implicit reporting within clinical notes increases. Future efforts focusing on prediction of ECOG using a series of notes over time will be the next step. ECOG Unmasked oncology notes ECOG Score masked only ECOG Score + ECOG narrative marked Precision Recall F1 Score Precision Recall F1 Score Precision Recall F1 Score 0 0.89 0.92 0.91 0.75 0.93 0.83 0.59 0.74 0.66 1 0.88 0.86 0.87 0.76 0.74 0.75 0.19 0.47 0.27 2 1.00 0.82 0.90 0.78 0.33 0.47 0.78 0.16 0.26

341 Background: Specialist clinicians (oncologists, urologists, radiation oncologists) read through free text reports of prostate MRI and biopsy reports to risk stratify patients during initial consultation of localized PCa. This laborious and time-consuming process is repeated for each day during the clinic. Herein, we propose a hybrid stratification framework utilizing a combination of large language model (LLM) and rule-based programming to automate this process. Methods: This retrospective study included patients with localized PCa (2004-2024) presenting to care at Mayo Clinic with at least one positive prostate biopsy and MRI report available. The state-of-the-art LLM – GPT4 – was utilized using a structured zeroshot prompt to extract relevant phenotypic variables (clinical T stage, Gleason patterns, grade group, extent of specimen/cores involved, prostate size/volume) from unstructured MRI and biopsy reports. Using a rule-based algorithm, the extracted variables and PSA at PCa diagnosis were used to categorize each patient into one of the six NCCN risk categories (very low [VL], low [L], favorable intermediate [FInt], unfavorable intermediate [UFInt], high risk (HR), very high risk [VHR]). Prompts were iteratively developed using ~5% of the total dataset and validated on ~10% of the dataset. Final performance was assessed against a held-out expert annotated test dataset using evaluation metrics (accuracy, F1-score, precision, recall). Additionally, manual annotation was conducted by two independent novice reviewers to compare machine annotated and novice-human annotated risk categorization. Results: A total of 397 patients were included in the evaluation. The median age at diagnosis was 64.8 (IQR: 59.7-68.5); majority of the men were White (n: 366; 92%) and non-Hispanic (n: 371; 94%). The most prevalent risk category was UFint (n: 151; 38%) followed by FInt (n: 80; 20.2%), HR (n: 76; 19.1%), VHR (n: 64; 16.1%), L (21: 5.3%), and VL (n: 5; 1.3%). Novice reviewers achieved an accuracy of 79%, precision of 80%, recall of 82% and a F1-score of 0.81. The hybrid stratification framework using GPT4 achieved an accuracy of 89%, precision of 89%, recall of 88% and a F1-score of 0.88. Among 43 errors by GPT4 risk stratification, majority of the errors were due to miscategorization of clinical T-stage (n: 21; 49%) followed by erroneous number of positive cores (n: 13; 30%). However, additional evaluation showed that GPT4 achieved numerically higher performance (accuracy: 90%; precision: 91%; recall: 90%; F1: 0.90) compared to novice reviewers (accuracy: 77%; precision: 89%; recall: 77%; F1: 0.82) for ascertaining clinical T-stage. Conclusions: Large language models exhibited superior performance than novice clinicians for risk stratification in patients with localized PCa. As the next step, we intend to validate our framework for automated risk classification in a prospective manner.

Abstract Objective Data extraction from the published literature is the most laborious step in conducting living systematic reviews (LSRs). We aim to build a generalizable, automated data extraction workflow leveraging large language models (LLMs) that mimics the real-world 2-reviewer process. Materials and Methods A dataset of 10 trials (22 publications) from a published LSR was used, focusing on 23 variables related to trial, population, and outcomes data. The dataset was split into prompt development (n = 5) and held-out test sets (n = 17). GPT-4-turbo and Claude-3-Opus were used for data extraction. Responses from the 2 LLMs were considered concordant if they were the same for a given variable. The discordant responses from each LLM were provided to the other LLM for cross-critique. Accuracy, ie, the total number of correct responses divided by the total number of responses, was computed to assess performance. Results In the prompt development set, 110 (96%) responses were concordant, achieving an accuracy of 0.99 against the gold standard. In the test set, 342 (87%) responses were concordant. The accuracy of the concordant responses was 0.94. The accuracy of the discordant responses was 0.41 for GPT-4-turbo and 0.50 for Claude-3-Opus. Of the 49 discordant responses, 25 (51%) became concordant after cross-critique, increasing accuracy to 0.76. Discussion Concordant responses by the LLMs are likely to be accurate. In instances of discordant responses, cross-critique can further increase the accuracy. Conclusion Large language models, when simulated in a collaborative, 2-reviewer workflow, can extract data with reasonable performance, enabling truly “living” systematic reviews.

Introduction Prognostication in myelofibrosis (MF) currently relies on manual assessment of clinical, pathological, and laboratory variables, which is inefficient and error-prone. The potential of large language models (LLMs) and rule-based programming to automate data extraction and prognostication in these patients remains unexplored and merits investigation. Methods Patients with MF were identified from a retrospective chart review. Data regarding clinical (age, constitutional symptoms, transfusion requirements), laboratory (hemoglobin, circulating white blood cell, platelet, and blast counts), and pathological (grade of bone marrow fibrosis, chromosomal analyses, cytogenetics) features was extracted from the latest free-text clinical notes and pathology reports using the generative pretrained transformer-4 (GPT-4) model in a zero-shot setting with the prompting parameters set to generate deterministic responses. The patients were randomly sampled into development (15%) and test (85%) sets, with the documents of patients in the development set used for iterative prompt engineering. The finalized prompts were used for data extraction in the test set. Rule-based programming was used for post-processing of the extracted data and ascertaining prognosis via Dynamic International Prognostic Scoring System (DIPSS), DIPSS-plus (DIPSS+), and Mutation-enhanced International Prognostic Scoring System-70 (MIPSS-70). Automatically extracted data and prognostic categories were compared with the manually generated human annotations. Mean accuracy, precision, and recall with 95% confidence intervals (CI) were computed. Detailed error and sensitivity analyses were conducted to assess performance. Results The study included 100 patients (development: 15; test: 85) with 100 clinical notes (development: 15; test: 85) and 90 pathology reports (development: 13; test: 77). The mean age of these patients was 70 (SD: 10.6); 55 (55%) patients were men, and 93 (93%) patients were White. In the development set, the mean accuracy, precision, and recall for data extraction from clinical notes were 94% (95% CI: 90-98%), 92% (84-99%), and 99% (97-100%). For data extraction from pathology reports, the mean accuracy, precision, and recall were 96% (93-100%), 98% (92-100%), and 83% (63-100%). For prognostication, the mean accuracy, precision, and recall were 87% (79-94%), 80% (73-87%), and 100% using DIPSS (accuracy: 80%; precision: 73%; recall: 100%), DIPSS+ (accuracy: 93%; precision: 83%; recall: 100%) and MIPSS-70 (accuracy: 87%, precision: 83%; recall: 100%). In the test set, the mean accuracy, precision, and recall for data extraction from clinical notes were 96% (95% CI: 93-98%), 91% (87-96%), and 98% (97-100%). For data extraction from pathology reports, the mean accuracy, precision, and recall were 95% (92-98%), 93% (87-99%), and 90% (80-100%). For prognostication the mean accuracy, precision, and recall were 95% (94-96%), 89% (88-90%) and 99% (96-100%) using DIPSS (accuracy: 94%; precision: 89%; recall: 100%), DIPSS+ (accuracy: 95%; precision: 88%; recall: 95%) and MIPSS-70 (accuracy: 94%, precision: 90%; recall: 100%). Error analyses identified incorrect extractions by GPT-4 were related to the circulating blasts and constitutional symptoms variables from the clinical notes. Sensitivity analyses excluding the associated variables improved overall accuracy from 94% to 97% and 96% to 97% in the development and test sets, respectively. Similarly, incorrect extractions from the pathology reports were related to the CALR gene mutation variables. Sensitivity analyses excluding the associated variables improved the overall accuracy from 96% to 98% and 95% to 97% in the development and test sets, respectively. Conclusion LLMs and rule-based programming can accurately extract data from the electronic health records and prognosticate MF patients via DIPSS, DIPSS+, and MIPSS-70. The automated approach could significantly streamline clinical and research workflows, providing an efficient and less resource-intensive alternative to manual prognostication.