Background: Progression-free survival (PFS) is a crucial endpoint in cancer drug research. Clinician-confirmed cancer progression, namely real-world PFS (rwPFS) in unstructured text (ie, clinical notes), serves as a reasonable surrogate for real-world indicators in ascertaining progression endpoints. Response evaluation criteria in solid tumors (RECIST) is traditionally used in clinical trials using serial imaging evaluations but is impractical when working with real-world data. Manual abstraction of clinical progression from unstructured notes remains the gold standard. However, this process is a resource-intensive, time-consuming process. Natural language processing (NLP), a subdomain of machine learning, has shown promise in accelerating the extraction of tumor progression from real-world data in recent years. Objectives: We aim to configure a pretrained, general-purpose health care NLP framework to transform free-text clinical notes and radiology reports into structured progression events for studying rwPFS on metastatic breast cancer (mBC) cohorts. Methods: This study developed and validated a novel semiautomated workflow to estimate rwPFS in patients with mBC using deidentified electronic health record data from the Nference nSights platform. The developed workflow was validated in a cohort of 316 patients with hormone receptor–positive, human epidermal growth factor receptor-2 (HER-2) 2-negative mBC, who were started on palbociclib and letrozole combination therapy between January 2015 and December 2021. Ground-truth datasets were curated to evaluate the workflow’s performance at both the sentence and patient levels. NLP-captured progression or a change in therapy line were considered outcome events, while death, loss to follow-up, and end of the study period were considered censoring events for rwPFS computation. Peak reduction and cumulative decline in Patient Health Questionnaire-8 (PHQ-8) scores were analyzed in the progressed and nonprogressed patient subgroups. Results: The configured clinical NLP engine achieved a sentence-level progression capture accuracy of 98.2%. At the patient level, initial progression was captured within ±30 days with 88% accuracy. The median rwPFS for the study cohort (N=316) was 20 (95% CI 18-25) months. In a validation subset (n=100), rwPFS determined by manual curation was 25 (95% CI 15-35) months, closely aligning with the computational workflow’s 22 (95% CI 15-35) months. A subanalysis revealed rwPFS estimates of 30 (95% CI 24-39) months from radiology reports and 23 (95% CI 19-28) months from clinical notes, highlighting the importance of integrating multiple note sources. External validation also demonstrated high accuracy (92.5% sentence level; 90.2% patient level). Sensitivity analysis revealed stable rwPFS estimates across varying levels of missing source data and event definitions. Peak reduction in PHQ-8 scores during the study period highlighted significant associations between patient-reported outcomes and disease progression. Conclusions: This workflow enables rapid and reliable determination of rwPFS in patients with mBC receiving combination therapy. Further validation across more diverse external datasets and other cancer types is needed to ensure broader applicability and generalizability.

Background: Alzheimer’s disease (AD) progression is commonly summarized through cognitive and global staging measures, yet routine care captures multidomain deterioration across cognition, function, neuropsychiatric symptoms, caregiver dependence, and systemic physiology. After EVOKE and EVOKE+ showed no population-level slowing of early symptomatic AD progression with oral semaglutide, a key unresolved question is whether disease phase, functional reserve, multimorbidity, and the temporal ordering of decline identify clinically distinct trajectories that are not captured by average trial effects. Methods: We conducted a retrospective observational study using de-identified electronic health record data from a federated network of more than 29 million patients. Large language model (LLM)-enabled extraction was used to recover standardized cognitive, functional, global severity, and neuropsychiatric assessment scores from clinical notes. Functional decline was defined from FAQ, ADCS-ADL, and ADCS-MCI-ADL; cognitive decline was defined from MMSE and MoCA. We reconstructed longitudinal AD trajectories, compared outcome-domain representation in contemporary AD trials, characterized patient-level timing of functional versus cognitive decline, and evaluated clinical trajectories among patients who initiated semaglutide after AD diagnosis. Physician adjudication showed high extraction accuracy for the main reportable instrument set. Findings: Among 131,824 patients with structured AD diagnoses, 42,242 had at least one LLM-derived outcome assessment from clinical notes and 341 initiated semaglutide after AD diagnosis. Normalized trajectories over decades of routine clinical care showed long prediagnostic decline and AD diagnosis-proximal acceleration, with the composite score decreasing from 0.99 at 25.1 years before AD diagnosis to 0.64 at 4.9 years after AD diagnosis. Compared with matched non-AD controls, AD patients showed marked post-index functional separation, including lower ADCS-MCI-ADL scores at year 4 (28.1 versus 47.0 points; P<0.001) and higher FAQ impairment across follow-up. We also evaluated clinical outcome and physiologic trajectories in semaglutide-treated AD patients and matched non-semaglutide AD controls; in paired score-change analyses, semaglutide-treated patients had more favorable MMSE change than matched controls (+0.4 versus −2.3 points; P=0.022) and more favorable MoCA change (−0.3 versus −2.2 points; P=0.049), whereas functional, global severity, and neuropsychiatric comparisons were not statistically significant and were limited by smaller sample sizes. Among 3,467 patients with at least two repeated functional assessments or two repeated cognitive assessments, 2,883 (83.2%) had confirmed functional decline and 2,535 (73.1%) had confirmed cognitive decline, with both functional and cognitive decline observed in 2,169 (62.6%) patients. Among the 2,169 patients, cognitive decline preceded functional decline (“cognitive decline-first”) in 1,123 patients (51.8%), functional decline preceded cognitive decline (“functional decline-first”) in 741 patients (34.2%), and both declines were confirmed on the same date in 305 patients (14.1%). Cognitive decline-first was more prevalent. Overall, 631 patients (29.1%) had cognitive decline at least 12 months before functional decline and 327 patients (15.1%) had functional decline at least 12 months before cognitive decline. Cognitive decline-first patients had higher medication burden and greater note-derived symptom burden. Functional decline-first patients were enriched for APOE ε3/ε3 (OR 2.82, 95% CI 1.17–6.76; P=0.023), supporting the need for trajectory-aware and genetically stratified approaches in semaglutide trial design and interpretation. In general, AD was also accompanied by systemic physiologic divergence from matched controls, including prediagnostic weight decline and postdiagnostic albumin and hemoglobin decline. In matched paired score-change analyses, semaglutide-treated AD patients showed more favorable MMSE change than matched non-semaglutide AD controls and more favorable MoCA change, whereas functional, global severity, and neuropsychiatric comparisons were not statistically significant and were limited by smaller sample sizes. Interpretation: LLM-enabled clinical-note phenotyping reveals that AD progression in routine care is multidomain and temporally heterogeneous, with functional decline-first and cognitive decline-first subsets that differ in medication and note-derived phenotype burden. These findings do not establish semaglutide efficacy in AD, but they suggest that EVOKE and EVOKE+ leave open a stratification question: whether treatment-associated trajectories vary by disease phase, functional reserve, multimorbidity, and the temporal ordering of functional and cognitive decline. Prospective validation, physician adjudication, and matched comparator analyses are needed before these phenotypes can guide trial enrichment or clinical decision-making.

Real-world benefit from incretin-based therapies depends on actual treatment initiation and sustained exposure, yet structured prescription records may capture prescribing intent rather than confirmed use. Using a large federated, de-identified U.S. EHR platform spanning over 29 million patients, we analyzed patterns of GLP1-RA initiation, persistence, and compounding from clinical notes using an evidence-linked large language model (LLM) pipeline evaluated against an independent physician -adjudicated extracted-event reference standard, achieving 98.4% accuracy for structural chunk triage and 88.2% accuracy for medication-status adjudication. Among 553,073 adults with at least one semaglutide (n=376,697) or tirzepatide (n=176,376) prescription and baseline weight availability, documented initiation within ±3 months of the first structured prescription was identified in 141,189 semaglutide patients and 62,040 tirzepatide patients. Among these patients with documented initiation, frictionless starts accounted for 70.1% of semaglutide and 77.9% of tirzepatide starts, initiation after documented friction accounted for 17.2% and 15.9%, and early interruption after initiation within 3 months accounted for 12.7% and 6.2%, respectively. We analyzed treatment persistence over an 18 month period in a subset of 69,976 patients with “frictionless initiation”. Clinical note-based and structured prescription-derived persistence yielded materially different trajectories within each drug (log-rank p < 0.001). A note-derived ascertainment approach showed that 42.3% and 43.1% of semaglutide and tirzepatide patients, respectively, continued therapy, while structured prescription-based ascertainment yielded higher persistence estimates of 55.4% for semaglutide and 58.3% for tirzepatide. Among the patients with a documented 3-18 month barrier or discontinuation event, insurance, cost barriers, perioperative holds and gastrointestinal adverse events were the top documented reasons. Compounded exposure was detected in 1,696 frictionless initiators, with the first compounding signal occurring on or before the first structured prescription in 50.8% of semaglutide and 57.9% of tirzepatide patients. These findings show that AI curation of clinical notes can distinguish prescription intent from verified exposure. With the potential time-to-market confounding, semaglutide shows higher initiation and lower non-initiation than tirzepatide, with comparable note-confirmed persistence at 18 months. More broadly, these findings support integrative exposure ascertainment to strengthen real-world GLP-1 evidence generation, payer decision-making, safety surveillance, and strategies to improve sustained access and treatment continuity.

EVOKE and EVOKE+ trials showed no population-level slowing of early symptomatic Alzheimer’s disease (AD) with semaglutide, motivating an assessment of whether clinical-note phenotyping over decades of routine care could identify distinct AD trajectories. Here we analyzed de-identified EHRs from a 29 million-patient federated network using LLMs to extract standardized cognitive, functional, global severity, and neuropsychiatric note-derived scores with high physician-adjudicated accuracy. Among 131,824 AD patients, 42,242 had ≥1 extracted assessment and 341 initiated semaglutide after diagnosis. Composite scores declined from 0.99 at 25.1 years before diagnosis to 0.64 at 4.9 years after diagnosis. Compared with matched non-AD controls, AD patients had lower ADCS-MCI-ADL scores at year 4 (28.1 versus 47.0 points; P<0.001) and higher FAQ impairment across follow-up. In paired score-change analyses, semaglutide-treated patients had more favorable MMSE change than matched controls (+0.4 versus −2.3 points; P=0.022) and more favorable MoCA change (−0.3 versus −2.2 points; P=0.049). Among 2,169 patients with both decline domains, cognitive decline preceded functional decline (“cognitive decline-first”) in 1,123 patients (51.8%), functional decline preceded cognitive decline (“functional decline-first”) in 741 patients (34.2%). Cognitive decline-first trajectories were more prevalent, with 631 patients (29.1%) showing cognitive decline at least 12 months before functional decline, compared with 327 patients (15.1%) showing functional decline at least 12 months before cognitive decline. Cognitive decline-first patients also had higher medication burden and greater note-derived symptom burden. Functional decline-first patients were enriched for APOE ε3/ε3 (OR 2.82; P=0.023) motivating the need for trajectory-aware and genetically stratified approaches to Alzheimer’s disease trial design.

Background: Real-world evidence (RWE) is increasingly recognized as a valuable type of oncology research but extracting fit-for-purpose real-world data (RWD) from electronic health records (EHRs) remains challenging. Manual abstraction from free-text clinical documents, although the gold standard for information extraction, is resource-intensive. RWD generation using natural language processing (NLP) has been limited by performance ceilings and annotation requirements, which recent LLMs improve on. We evaluate new NLP workflows in abstracting multiple myeloma (MM) information from de-identified EHRs. Methods: NLP workflows (BERT and Llama-based using various prompt types) were developed for 12 MM-specific data fields and evaluated with manually curated data from 125 clinical notes. The best Llama-based workflow for three data fields was applied to all recent notes in selected charts to generate patient journey timelines. Results: Average F ₁ for the best Llama and BERT workflows was 0.82 and 0.65 respectively. Best workflow performance ranged across the data fields (F ₁ = 0.59–0.99). Statistical analysis of the results showed model size, inter-rater reliability (IRR), variable type, and prompt design significantly predicted workflow performance, in descending order of significance ( p < 0.05). Conclusion: The overall performance improvements seen with larger LLMs and chain-of-thought prompting was greater in ambiguous data fields. IRR can be used to prioritize NLP resources and increase efficiency of RWD generation without sacrificing data quality.