Flaherty, Lawrence E.

Abstract Background S0221 investigated weekly vs every 2 weeks dosing of doxorubicin (A) and cyclophosphamide (C) followed by paclitaxel in patients with high-risk early breast cancer. After an interim analysis, random assignment to the 2 AC arms was stopped for futility, and the trial was modified to study only the paclitaxel schedules. Methods Between December 2003 and November 2010, a total of 2716 patients were randomly assigned in a 2 × 2 factorial design to 15 weeks of weekly A and daily C vs 6 cycles of every 2 weeks AC; and weekly paclitaxel for 12 weeks vs 6 cycles of every 2 weeks paclitaxel. Between January 2011 and January 2012, an additional 578 patients were assigned to 4 cycles of every 2 weeks AC and randomly assigned to weekly vs every 2 weeks paclitaxel. Updated survival was assessed using log-rank tests and Cox regression models. We compared outcomes by breast cancer subtype as well. Results At a median follow-up of 12.1 years, there were no statistically significant differences among the 4 treatment arms in disease-free survival (DFS) (P = .91) or overall survival (P = .34) in the original protocol. Among the 578 patients assigned AC for 4 cycles and randomly assigned to paclitaxel weekly vs every 2 weeks paclitaxel, there were no overall differences in DFS (P = .32) or overall survival (P = .42). Conclusion As there were no statistically significant outcome differences in DFS or overall survival between the studied schedules of AC and paclitaxel with extended follow-up in the original or revised protocol, either paclitaxel schedule may be recommended, with selection based on toxicity, cost, or patient preference.

e13637 Background: Large language models (LLMs) like GPT-4 (OpenAI) and Claude Opus (Anthropic) showed high accuracy in medical multiple-choice exams, but data on their oncology-specific clinical reasoning and performance is limited. This study evaluates their accuracy and clinical reasoning on breast oncology multiple-choice questions (MCQs) from the American Society of Clinical Oncology (ASCO) question bank. Methods: Using OpenAI and Anthropic Application Programming Interface (APIs), questions were tested without additional prompts under consistent settings (GPT-4 and Claude Opus; Temperature = 0, Tokens = Max). Each question was tested three times to assess precision. Then, Chain-of-thought (COT) prompting was applied to promote LLMs stepwise reasoning to increase their accuracy. Accuracy before and after COT prompting was compared. Incorrect responses were reviewed by board-certified medical oncologists speclized in breast cancer, who scored reasoning clarity, bias, and clinical relevance. Qualitative feedback was descriptively analyzed. Results: A total of 273 breast oncology MCQs were evaluated across the two LLMs. GPT-4 achieved an initial accuracy of 81.3% (222/273; 95% CI: 76.3%-85.5%), compared to Claude Opus, which achieved an accuracy of 79.5% (217/273; 95% CI: 74.3%-83.9%). The Chi-squared test for difference between the models before chain-of-thought (COT) prompting yielded a p-value of 0.59, indicating no statistically significant difference in accuracy between GPT-4 and Claude Opus prior to COT prompting. After COT prompting, the performance of the two models diverged significantly. GPT-4 saw a net decline in accuracy, decreasing by a net of 1 correct answer, resulting in an overall accuracy of 80.95% (221/273). In contrast, Claude Opus experienced a notable improvement, with a net of 19 additional correct answers, leading to an accuracy of 86.4% (236/273). A statistical analysis using a Chi-squared test revealed a difference in accuracy between the two models of 5.5% (p = 0.08), demonstrating that the improvement in accuracy for Claude Opus after COT prompting was borderline statistically significant compared to GPT-4. Thematic analysis of oncologists' feedback revealed that the most common reasons for incorrect answers were reliance on outdated guidelines, misinterpretation of clinical trial data, and failure to consider multidisciplinary or patient-specific approaches in clinical decision-making. Conclusions: Although AI models can achieve high scores on multiple-choice exams, they still require human supervision. These models rely on potentially outdated training data and lack the ability to individualize patient care or apply data from clinical trials. especially in unique or unconventional/non textbook scenarios.

CRA1008 The full, final text of this abstract will be available at abstract.asco.org at 7:30 AM (EDT) on Monday, June, 3, 2013, and in the Annual Meeting Proceedings online supplement to the June 20, 2013, issue of Journal of Clinical Oncology. Onsite at the Meeting, this abstract will be printed in the Monday edition of ASCO Daily News.

521 Background: S0221 investigated weekly vs q 2 week dosing of doxorubicin/cyclophosphamide (AC) and paclitaxel (P) in patients (pts) with high risk early breast cancer as previously reported (JCO 33:58-64, 2015). After enrollment of 2716 pts randomization to the two AC arms was stopped for futility and an additional 578 pts received 4 cycles of q 2 week AC and were randomized to P weekly (Pw) or P q 2 weeks (P2). We report updated results of the original trial design and the first report of the 578 pts treated with AC x 4 and Pw x 12 or P2 x 6. Methods: Between December 2003 and November 2010, 2716 pts were randomized in a 2x2 factorial design to 1) 15 weeks of weekly AC (A 24 mg/m2/week and C 60 mg/m2/day po) vs 6 cycles of q 2 week AC (A 60 mg/m2 and C 600 mg/m2) and 2) Pw (paclitaxel 80 mg/m2/week x 12) vs P2 (paclitaxel 175 mg/m2 q 2 weeks x 6), with growth factor support as previously described. After study amendment 578 patients received 4 cycles of q 2 week AC followed by Pw or P2. Updated survival was assessed using log-rank tests and Cox regression models. Results: At a median follow-up of 8.5 years, among the pts treated in the original protocol, there were no significant differences among the four treatments for DFS (p=0.21) or OS (p=0.08). The triple-negative subset had worse DFS (P<0.001) than the HER2-positive or ER/PR+/HER2- subsets, with 5 year DFS of 75% vs 83% and 84%, respectively. While we previously found in the triple negative subset that the arm using q 2 weeks for both AC and paclitaxel was marginally superior, the differences among the arms are no longer significant for DFS (p=0.12) or OS (p=0.11). Among the 578 pts assigned ACx4 and randomized to Pw v P2 there were no overall differences in DFS (p=0.70) or OS (p=0.63) after 4.4 years median follow-up. Conclusions: There were no significant differences in DFS or OS between any of the schedules with extended follow-up in the original cohort and no difference in outcome by paclitaxel schedule for the 578 additional patients in the revised protocol. Either paclitaxel schedule may be recommended, with selection based on toxicity, cost, or patient preference rather than efficacy. Support: NCI grants CA32102, CA38926, CA21115, CA21076, CA77597, CA25224, CA77202, CCSRI15469, and Amgen, Inc. Clinical trial information: NCT00070564.

CRA1008 Background: S0221 is a SWOG-coordinated phase III adjuvant chemotherapy intergroup trial in node-positive and high-risk node-negative operable breast cancer which hypothesized that 1) the weekly AC+G regimen is superior to ddAC x 6 and 2) 12 weeks of weekly paclitaxel (wP) is superior to q 2 week paclitaxel x 6 (ddP). Methods: Between December 2003 and November 2010, 2,716 patients were randomized in a 2 x 2 factorial design to 1) AC+G vs ddAC and 2) P 80 mg/m2/week x 12 vs P 175 mg/m2 q 2 weeks x 6. If there was no significant interaction between the factors, the trial was powered to find a disease-free survival hazard ratio (HR) ≤ 0.82 for weekly vs q 2 week for each factor. At the first interim analysis, the AC randomization was halted for futility, and S0221 was closed to accrual 10 November 2010. S0221 reopened 15 December 2010, after which time all patients received 4 cycles of ddAC and randomization to P weekly x 12 and ddP x 6 continued. Accrual halted at a total of 3,294 in January 2012. Results: By September 7, 2012, 487 events and 340 deaths had occurred, prompting the third planned interim analysis. The Data Safety and Monitoring Committee recommended reporting the results since the futility boundary was crossed. A Cox model adjusting for the AC arms had a HR = 1.08 (95% CI 0.90-1.28; p=0.42), with the 99.5% CI excluding the original alternative hypothesis that the HR=0.82. There was no significant interaction of the two factors. Estimated 5-year progression-free survivals were 82% for weekly P and 81% for ddP. Toxicity data were available for 1,385 patients treated with ddP and 1,367 treated with weekly P. Grade 5 toxicity occurred in 4 patients on ddP and 2 on weekly P. Percent grade 3-4 toxicity per arm are shown in the Table. Conclusions: Either ddPx6 or weekly P x 12 are acceptable schedules of P administration. The differences in leukopenia likely reflect ascertainment bias against weekly P. If this is accepted, weekly P x 12 produces less overall toxicity than 6 cycles of ddP. Support: NCI grants CA32102, CA38926, CA21115, CA21076, CA77597, CA25224, CA77202, CCSRI15469, and Amgen, Inc. Clinical trial information: NCT00070564. [Table: see text]

Purpose To determine the optimal dose and schedule of anthracycline and taxane administration as adjuvant therapy for early-stage breast cancer. Patients and Methods A 2 × 2 factorial design was used to test two hypotheses: (1) that a novel continuous schedule of doxorubicin-cyclophosphamide was superior to six cycles of doxorubicin-cyclophosphamide once every 2 weeks and (2) that paclitaxel once per week was superior to six cycles of paclitaxel once every 2 weeks in patients with node-positive or high-risk node-negative early-stage breast cancer. With 3,250 patients, a disease-free survival (DFS) hazard ratio of 0.82 for each randomization could be detected with 90% power with two-sided α = .05. Overall survival (OS) was a secondary outcome. Results Interim analyses crossed the futility boundaries for demonstrating superiority of both once-per-week regimens and once-every-2-weeks regimens. After a median follow-up of 6 years, a significant interaction developed between the two randomization factors (DFS P = .024; OS P = .010) in the 2,716 patients randomly assigned in the original design, which precluded interpretation of the two factors separately. Comparing all four arms showed a significant difference in OS (P = .040) but not in DFS (P = .11), with all treatments given once every 2 weeks associated with the highest OS. This difference in OS seemed confined to patients with hormone receptor–negative/human epidermal growth factor receptor 2 (HER2) –negative tumors (P = .067), with no differences seen with hormone receptor–positive/HER2-negative (P = .90) or HER2-positive tumors (P = .40). Conclusion Patients achieved a similar DFS with any of these regimens. Subset analysis suggests the hypothesis that once-every-2-weeks dosing may be best for patients with hormone receptor–negative/HER2-negative tumors.