AUTHORS
Ali Kamali, PhD, Kuo-Kai Chin, MD2, Baher Krayem3, Alexander Chan, MD4, Khawaja Bilal5, Chad Vanderbilt6, Justin Jee7, James Brogan, MD8, Lia Jueng, MD9, Ying Liu, MD10, Wenbin Xiao, MD, PhD11, Ahmet Dogan, MD12, Eytan M Stein I, MD13, Gregory Goldgof, MD, PHD, MS14, Xin Wang, MD, MSCE15
AFFILIATIONS
1 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
2 Memorial Sloan Kettering Cancer Center, New York, New York, United States
3 Memorial Sloan Kettering Cancer Center, New York, New York, United States
4 Memorial Sloan Kettering Cancer Center, New York, New York, United States
5 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
6 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
7 Memorial Sloan Kettering Cancer Center, New York, New York, United States
8 Memorial Sloan Kettering Cancer Center, New York, New York, United States
9 Weill Cornell Medical College, New York, New York, United States
10 Memorial Sloan Kettering Cancer Center, New York, New York, United States
11 Memorial Sloan Kettering Cancer Center, New York, New York, United States
12 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
13 Memorial Sloan Kettering Cancer Center, New York, New York, United States
14 Memorial Sloan Kettering Cancer Center, New York, New York, United States, Memorial Sloan Kettering Cancer Center, New York, New York, United States
15 Memorial Sloan Kettering Cancer Center, New York, New York, United States
ABSTRACT
Introduction: Real-world evidence and prognostic modeling in acute myeloid leukemia (AML) are limited by challenges in curating longitudinal records. AML clinicogenomic data are scattered across heterogeneous unstructured documents spanning years. Many critical concepts, such as risk stratification and response, are not readily extracted from single reports and require expert longitudinal synthesis, making cohort construction slow to build and maintain.
Methods: We developed the Hematologic Multimodal Analytics Platform (HeMAP), an LLM-enabled system that reconstructs AML trajectories from clinical records. The pipeline performs disease-status extraction, prior history extraction, cytomolecular extraction and classification, treatment-timeline reconstruction, and response classification. Each stage uses task-specific prompts that were iteratively refined with oncologist and hematopathologist oversight. Data from pathology reports, oncology notes, treatment records, and transplant records are organized by timepoint and propagated across stages to provide temporal and clinical context for downstream inference. HeMAP produces 71 directly extracted and 68 deterministically derived variables spanning disease status, genomic risk, treatment, response, and outcomes.
We applied HeMAP to 2,850 patients treated at Memorial Sloan Kettering Cancer Center from 2015 to 2026 with at least one molecular testing record assigned an AML-associated diagnostic code at report review. The disease-status extraction step identified 2,499 confirmed AML cases (median 92 source documents per patient [IQR 42–173]; 317,462 total documents; 16,792 timepoints), which were used for further analysis. Output was validated against two independent, non-model-seeded references within the cohort (a 295-pt reference set curated by hematopathologist; an 18-patient independent, blinded review comparison set by oncology trainees). Utility was assessed by generating a frontline venetoclax/azacitidine (VEN/AZA) cohort and analyzing outcomes.
Results: Against the 295-patient set, molecular extraction was accurate for key genomic features relevant to ELN risk stratification (range, 96–100%; mean, 98%); diagnosis date matched within 14 days in 88%; induction regimen backbone and induction start date (±14 days) were concordant in 88% and 89%, respectively; best pre-transplant response, categorized as composite CR versus no composite CR, and flow-MRD status at best response were concordant in 92% and 84%, respectively; and cytogenetic risk and ELN 2022 classification were concordant in 86% and 84%, respectively.
Comparing HeMAP and trainees in the 18-patient subset, diagnosis date was identified within 14 days in 100% of patients. Direct extraction of molecular abnormalities was 100% accurate for key ELN genes, HeMAP-derived risk classification matched trainee annotation for cytogenetic risk and ELN2024 (both 100%), ELN2022 (94%), Mayo (78%), and BEAT-AML (72%). However, attending hematologist adjudication attributed all risk classification discordances to be trainee errors, yielding 100% accuracy across all 5 frameworks for HeMAP. For ELN 2022 response classification, trainee-HeMAP agreement was 82%, whereas attending hematologist adjudication found model accuracy to be 91%. Extrapolating the mean annotation time documented by trainees (31.3 minutes/patient), manual abstraction of the full cohort (n=2,499) would require approximately 1,304 hours (163 workdays).
We then assembled a frontline VEN/AZA cohort (n=448) using automatically reconstructed trajectories. Median overall survival (OS) was 18.6 months; composite complete remission (CR/CRi/CRh) was 74%. All evaluated frameworks significantly stratified OS (log-rank p<0.001), but discrimination differed: ELN2024 (Harrell C 0.60), BEAT-AML (0.60), and Mayo (0.58) outperformed ELN2022 (0.54), consistent with prior reports in venetoclax-treated AML. These results provide orthogonal validation that HeMAP preserves prognostic relationships.
Conclusions: HeMAP reconstructs longitudinal AML disease trajectories with high concordance from routine records at scale and generates analysis-ready cohorts that reproduce established prognostic associations without manual chart abstraction. It supports fully automated hypothesis-generating studies or targeted review when greater precision is required, providing scalable infrastructure for real-world evidence, prognostic model evaluation, and multimodal AML risk stratification.
DISCLOSURES
Ali Kamali, PhD: No relevant conflicts of interest to declare.