Citations

PURPOSE To evaluate open-source large language models (LLMs) for extracting cancer-specific phenotypic data, benchmark their performance against GPT4 models, and assess the impact of fine-tuning with training data sizes. METHODS Open-source LLMs (Mistral, LLaMa, MAMBA, BioMistral) were evaluated in zero-/one-shot and fine-tuned setups against GPT4-turbo/GPT4o to extract the cancer presence, progression, response, and metastatic sites from radiology impressions of patients with solid tumors treated at Dana-Farber Cancer Institute. Performance metrics (accuracy, precision, recall, F1-score) were computed. McNemar's odds ratio (OR), measuring which model is more likely to be correct when they disagree, was computed with 95% CI. Statistical significance was assessed using the alpha of .000139. RESULTS This study included 2,623 patients (25,273 radiology impressions). In zero-/one-shot settings, GPT4-turbo/GPT4o outperformed open-source LLMs. However, fine-tuned open-source LLMs achieved higher F1-scores than GPT4 models. Compared with the best-performing GPT4 model, fine-tuned Mistral0.2-7.3B (OR, 0.27 [95% CI, 0.20 to 0.36]; P < .00001), Mistral0.3-7.3B (OR, 0.26 [95% CI, 0.19 to 0.36]; P < .00001), LLaMa2-6.7B (OR, 0.30 [95% CI, 0.22 to 0.40]; P < .00001), LLaMa3.1-8B (OR, 0.37 [95% CI, 0.28 to 0.48]; P < .00001), and MAMBA-2.8B (OR, 0.32 [95% CI, 0.24 to 0.42]; P < .00001) showed significantly better performance in ascertaining disease progression. Performance was consistently better for inferring overall response, any evidence of cancer, and sites of metastases, with no significant differences among fine-tuned open-source LLMs. Fine-tuning gains plateaued at 25% of training data (5,718 impressions) and remained comparable at 5% (1,144 impressions). CONCLUSION Open-source LLMs, when fine-tuned using labeled data, can effectively automate the ascertainment of key radiophenotypic variables using only the impression section of radiology reports, without the full report text. Their consistent performance in small training sets suggests that these models may provide a scalable approach for phenotypic characterization of patients with cancer in real-world clinical settings.

Abstract Background The ability of large language models (LLMs) to work collaboratively and screen studies in a systematic review (SR) is under-explored. Hence, we aimed to evaluate the effectiveness of LLMs in automating the process of screening in systematic reviews. Methods This is an observational study which included labeled data (title and abstracts) for five SRs. Originally, two reviewers screened the citations independently for eligibility. A third reviewer cross-checked each citation for quality assurance. GPT-4, Claude-3-Sonnet, and Gemini-Pro-1.0 were used using zero-shot chain-of-thought prompting. Collaborative approaches included (i): conflict resolution using benefit of the doubt, (ii) majority voting using an independent third LLM and (iii) conflict resolution using an informed third LLM. Performance was assessed using accuracy, precision for exclusion, and recall for inclusion. Work saved over samples (WSS) was computed to estimate the reduction in manual human effort. Results A total of 11300 articles were included in this study. The individual models, GPT-4, Claude-3-Sonnet, and Gemini-Pro-1.0 exhibited a high precision for exclusion, achieving 99.7%, 99.7%, and 99.2% and high recall for inclusion achieving 95.5%, 96.6% and 85.7%, respectively. However, the collaborative approach utilizing the two best-performing models (GPT-4 and Claude-3S) achieved an average precision of 99.9% and a recall of 98.5% (across all collaborative approaches). Furthermore, the proposed collaborative approach resulted in an average WSS of 63.5%, compared to the average WSS of 45.2% for individual models. Conversational LLM interactions showed a consistent pattern of results. Limitations This study was limited due to reliance on proprietary models, and evaluation on oncology datasets. Conclusion Evidence shows that collaborative LLMs enable efficient, high-performing screening in systematic reviews, supporting continuous evidence updates. Primary funding source NIH (U24CA265879-01-1) and Carolyn-Ann-Kennedy-Bacon Fund.

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.

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.

2054 Background: Systematic reviews that summarize the toxicity of Immune checkpoint inhibitors (ICIs) become outdated very soon after publication. Therefore, we reported results of a toxicity meta-analysis at 2019 ASCO meeting and informed the intent to create a living systematic review (LSR). LSRs combine human and machine effort and support rapid evidence synthesis and living clinical practice guidelines. Now, we report our experience maintaining a LSR on toxicity of ICIs. Methods: Steps include quarterly literature searches to identify new clinical trials reporting ICI-associated adverse events (AEs), AI-enabled screening of new citations which meet the inclusion criteria, automated cumulative meta-analysis and an online reporting platform. Standard data formats and protocols were designed for inputting text, tables and graphics. Software was written to interpret these data and output the information in the appropriate format, such as a forest plot and summary tables. Finally, a dynamic interface that enables user inputs and displays the associated output was designed. Results: The LSR is continuously updated incorporating toxicity data from new clinical trials as it becomes available. We have screened 8000 relevant citations and summarized the odds of Grade 3 or higher AEs and AEs of special interest in patient receiving ICIs. The results are updated on quarterly basis and are available online. The results are updated on quarterly basis and will be available on a website at the time of publication. Prototype with dummy data is available at this link . This interface can also be manipulated via user input to organize and sort data tables and forest plots by type of cancer, name or mechanism (PD-1 or PD-L1) of ICI agent, single agent or combination, type of control arm, line of treatment and several other clinically relevant filters. For example, a user can instantaneously generate a meta-analysis summarizing the risk of colitis or pneumonitis in metastatic lung cancer trials with pembrolizmuab. Conclusions: This LSR engine can prospectively synthesize toxicity data from ICI trials in an efficient manner providing accurate and timely information for advanced clinical decision support at point-of-care. Efforts are ongoing to improve efficiency of screening, improve AI-enabled processes for automated screening and data abstraction, and test across multiple clinical questions.

80 Background: Several previous systematic reviews and meta-analyses have attempted to summarize toxicity of Immune Checkpoint inhibitors (ICIs). However, very soon after each one of these reviews has been published, it became outdated. ICIs are currently used in 14 different cancers and data is rapidly evolving from new clinical trials. A living Systematic review, which is defined as a systematic review that is continually updated to incorporate relevant new evidence as it becomes available, is necessary in this situations. Methods: In order to dynamically display content -- including textual data, tables, charts, and graphs – standard data formats and protocols were designed for various components of the systematic reviews carried out by researchers. Given these standardized formats, software was written to interpret this data and output the information in the appropriate format, such as a forest plot or data table. Finally, a dynamic interface that enables user inputs (such as which aspect of the meta-analysis to view) and displays the associated output was designed. The overall living Systematic Review infrastructure includes monthly literature searches, cumulative meta-analysis, and an online reporting platform. Results: We have developed a Living Systematic Review engine that can take in standardized data and reflect this data in an interactive, intuitive manner in the form of text, tables, and plots. This interface can also be manipulated via user input to organize and sort data tables or omit/highlight specific pieces of data. Conclusions: We have initiated the first living systematic review in immune-oncology that will be continuously updated, incorporating relevant new evidence as it becomes available, and will provide accurate and up to date toxicity estimates to support clinical decision making. The Living Systematic Review interface allows users to easily and intuitively access the most recent available clinical evidence on toxicity of ICIs.