AUTHORS
Janamejey Gaur, MD, Vaishali Gautam, BS2, Chunyi Zhang3, James P Long, PhD4, Benjamin Mescher5, Sattva S Neelapu Sr, MD6, Sairah Ahmed, MD7, Anath Christopher Lionel, MD, PhD8, Swaminathan P. Iyer, MD9
AFFILIATIONS
1 UT MD Anderson Cancer Center, Houston, Texas, United States
2 UT MD Anderson Cancer Center, Houston, Texas, United States
3 UT MD Anderson Cancer Center, Houston, Texas, United States
4 UT MD Anderson Cancer Center, Houston, Texas, United States
5 UT MD Anderson Cancer Center, Houston, Texas, United States
6 UT MD Anderson Cancer Center, Houston, Texas, United States
7 UT MD Anderson Cancer Center, Houston, Texas, United States
8 UT MD Anderson Cancer Center, Houston, Texas, United States
9 UT MD Anderson Cancer Center, Houston, Texas, United States
ABSTRACT
Background: PET/CT using the Deauville 5-point scale (DS) is the standard for response assessment after CAR T-cell therapy for B-cell lymphomas. However, extracting serial scores and reconciling radiology reports with clinical interpretations require labor-intensive chart review. We evaluated a large language model (LLM)-driven pipeline for automated DS extraction and interpretation, testing whether the resulting response classifications reproduced the expected association between early metabolic response and overall survival (OS).
Methods: This IRB-approved retrospective study used de-identified data under a consent waiver. We analyzed 744 consecutive patients with aggressive B-cell lymphomas who received standard-of-care CAR T-cell therapy at a tertiary center, encompassing 1,841 PET/CT reports at baseline (n=713) and on days 30 (n=687), 60 (n=48), and 90 (n=393). Clinical and imaging data were retrieved using Context Engine, preserving data provenance from the electronic health record. A multi-model pipeline (GPT-4o, Gemini 2.0, and Claude 3.5 Haiku) processed PET reports by extracting an explicitly stated DS or, when absent, by interpreting FDG-uptake descriptions. Clinical oncology notes were independently processed to capture physician assessments, addenda, supplemental imaging, and revised interpretations. Concordance between PET-only LLM-derived scores (L-scores) and clinical-contextualized scores (C-scores) was the primary outcome. Post-infusion scores were operationally classified as complete metabolic response (CMR; DS 1–3), partial response (PR; DS 4), or progressive disease (PD; DS 5). OS was analyzed using Kaplan-Meier, log-rank, and multivariable Cox regression from day-30 and day-90 scan-date landmarks.
Results: Among 1,495 reports with both L- and C-scores, exact agreement was 93.5% (1,398/1,495; κ=0.914), and 97.9% agreed within +/- 1 DS point. Among 158 paired reports without an explicitly stated DS, for which the L-score was inferred from FDG uptake descriptions, 133 (84.2%) matched exactly and 92.4% agreed within +/- 1 point. Of the 97 discordant reports, 38 (39.2%) involved the adjacent DS 4/5 boundary - the largest discordance category. Contributing factors included physician reinterpretation, non-lymphoma FDG uptake, and addenda issued after the original report. CMR rates were 65.8% at day 30 and 63.6% at day 90. Among 379 patients with baseline DS5 and an evaluable day-30 assessment, 220 (58.0%) achieved CMR. In the day-30 landmark analysis (n=637; 267 deaths), median OS was 62.9 months for CMR, 19.5 months for PR, and 4.9 months for PD (log-rank p<0.0001). At the day-90 landmark (n=355; 130 deaths), median OS was 78.7, 23.9, and 9.7 months, respectively (log-rank p<0.0001). In a multivariable day-30 model adjusted for baseline DS, PD (HR 3.81, 95% CI 2.80–5.18; p<0.001) and PR (HR 1.84, 95% CI 1.31–2.58; p<0.001) independently predicted mortality, while baseline DS5 retained modest prognostic significance (HR 1.60, 95% CI 1.05–2.46; p=0.029). Among 320 patients with both baseline and day-90 scores, day-90 PD (HR 4.34, 95% CI 2.86–6.60; p<0.001) and PR (HR 2.57, 95% CI 1.40–4.71; p=0.002) remained prognostic, whereas baseline DS5 did not (HR 1.35, 95% CI 0.71–2.58; p=0.36).
Conclusion: LLM-derived early metabolic response strongly stratified OS after CAR T-cell therapy, with day-30 PD conferring a nearly four-fold adjusted mortality hazard relative to CMR. By day 90, metabolic response remained a strong prognostic factor and largely superseded the prognostic contribution of baseline DS. The pipeline also demonstrated high concordance between PET-only L-scores and clinical-contextualized C-scores. However, clinical note context remained important for resolving reports without an explicit DS and for addressing disagreements at the DS 4/5 boundary. These findings support automated Deauville score extraction as a scalable approach for response assessment and survival stratification. Limitations include the absence of an image-adjudicated gold standard, reliance on the C-score as the comparator, sparse day-60 sampling, and an untested proportional hazards assumption. Adjudicated validation and incorporation of treatment history are ongoing and will be presented at the ASH Annual Meeting.
DISCLOSURES
Janamejey Gaur, MD: No relevant conflicts of interest to declare.