Use of Natural Language Processing to Infer Sites of Metastatic Disease From Radiology Reports at Scale.

PURPOSE: To evaluate natural language processing (NLP) methods to infer metastatic sites from radiology reports. METHODS: A set of 4,522 computed tomography (CT) reports of 550 patients with 14 types of cancer was used to fine-tune four clinical large language models (LLMs) for multilabel classification of metastatic sites. We also developed an NLP information extraction (IE) system (on the basis of named entity recognition, assertion status detection, and relation extraction) for comparison. Model performances were measured by F1 scores on test and three external validation sets. The best model was used to facilitate analysis of metastatic frequencies in a cohort study of 6,555 patients with 53,838 CT reports. RESULTS: The RadBERT, BioBERT, GatorTron-base, and GatorTron-medium LLMs achieved F1 scores of 0.84, 0.87, 0.89, and 0.91, respectively, on the test set. The IE system performed best, achieving an F1 score of 0.93. F1 scores of the IE system by individual cancer type ranged from 0.89 to 0.96. The IE system attained F1 scores of 0.89, 0.83, and 0.81, respectively, on external validation sets including additional cancer types, positron emission tomography-CT ,and magnetic resonance imaging scans, respectively. In our cohort study, we found that for colorectal cancer, liver-only metastases were higher in de novo stage IV versus recurrent patients (29.7% v 12.2%; P < .001). Conversely, lung-only metastases were more frequent in recurrent versus de novo stage IV patients (17.2% v 7.3%; P < .001). CONCLUSION: We developed an IE system that accurately infers metastatic sites in multiple primary cancers from radiology reports. It has explainable methods and performs better than some clinical LLMs. The inferred metastatic phenotypes could enhance cancer research databases and clinical trial matching, and identify potential patients for oligometastatic interventions.

Alterations to DNA methylation patterns induced by chemotherapy treatment are associated with negative impacts on the olfactory pathway.

BACKGROUND: Exposure to cytotoxic chemotherapy treatment may alter DNA methylation (DNAm) in breast cancer patients. METHODS: We performed DNAm analysis in 125 breast cancer patients with blood drawn before and after chemotherapy, using the Illumina MethylationEPIC array. DNAm changes of 588,798 individual CpGs (including 41,207 promoter regions) were evaluated using linear regression models adjusted for monocyte proportion. Gene set enrichment analyses (GSEA) were conducted to identify key Gene Ontology (GO) biological processes or Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways associated with chemotherapy. Results were validated in a separate cohort of breast cancer patients who were treated (n = 1273) and not treated (n = 872) by chemotherapy (1808 blood, 337 saliva). RESULTS: A total of 141 differentially methylated CpGs and 11 promoters were significantly associated with chemotherapy after multiple testing corrections in both the paired sample and single time point analyses. GSEA of promoter regions (pre-ranked by test statistics) identified six suppressed biological processes (p < 4.67e-8) related to sensory perception and detection of chemical stimuli, including smell perception (GO:0007606, GO:0007608, GO:0009593, GO:0050906, GO:0050907, and GO:0050911). The same six biological processes were significantly suppressed in the validation dataset (p < 9.02e-14). The KEGG pathway olfactory transduction (hsa04740) was also found to be significantly suppressed (ppaired-samples = 1.72e-9, psingle-timepoint-blood = 2.03e-15 and psingle-timepoint-saliva = 7.52e-56). CONCLUSION: The enrichment of imprinted genes within biological processes and pathways suggests a biological mechanism by which chemotherapy could affect the perception of smell.

Inferring cancer disease response from radiology reports using large language models with data augmentation and prompting.

OBJECTIVE: To assess large language models on their ability to accurately infer cancer disease response from free-text radiology reports. MATERIALS AND METHODS: We assembled 10 602 computed tomography reports from cancer patients seen at a single institution. All reports were classified into: no evidence of disease, partial response, stable disease, or progressive disease. We applied transformer models, a bidirectional long short-term memory model, a convolutional neural network model, and conventional machine learning methods to this task. Data augmentation using sentence permutation with consistency loss as well as prompt-based fine-tuning were used on the best-performing models. Models were validated on a hold-out test set and an external validation set based on Response Evaluation Criteria in Solid Tumors (RECIST) classifications. RESULTS: The best-performing model was the GatorTron transformer which achieved an accuracy of 0.8916 on the test set and 0.8919 on the RECIST validation set. Data augmentation further improved the accuracy to 0.8976. Prompt-based fine-tuning did not further improve accuracy but was able to reduce the number of training reports to 500 while still achieving good performance. DISCUSSION: These models could be used by researchers to derive progression-free survival in large datasets. It may also serve as a decision support tool by providing clinicians an automated second opinion of disease response. CONCLUSIONS: Large clinical language models demonstrate potential to infer cancer disease response from radiology reports at scale. Data augmentation techniques are useful to further improve performance. Prompt-based fine-tuning can significantly reduce the size of the training dataset.

HER2 expression, copy number variation and survival outcomes in HER2-low non-metastatic breast cancer: an international multicentre cohort study and TCGA-METABRIC analysis.

BACKGROUND: HER2-low breast cancer (BC) is currently an area of active interest. This study evaluated the impact of low expression of HER2 on survival outcomes in HER2-negative non-metastatic breast cancer (BC). METHODS: Patients with HER2-negative non-metastatic BC from 6 centres within the Asian Breast Cancer Cooperative Group (ABCCG) (n = 28,280) were analysed. HER2-low was defined as immunohistochemistry (IHC) 1+ or 2+ and in situ hybridization non-amplified (ISH-) and HER2-zero as IHC 0. Relapse-free survival (RFS) and overall survival (OS) by hormone receptor status and HER2 IHC 0, 1+ and 2+ ISH- status were the main outcomes. A combined TCGA-BRCA and METABRIC cohort (n = 1967) was also analysed to explore the association between HER2 expression, ERBB2 copy number variation (CNV) status and RFS. RESULTS: ABCCG cohort median follow-up was 6.6 years; there were 12,260 (43.4%) HER2-low BC and 16,020 (56.6%) HER2-zero BC. The outcomes were better in HER2-low BC than in HER2-zero BC (RFS: centre-adjusted hazard ratio (HR) 0.88, 95% CI 0.82-0.93, P < 0.001; OS: centre-adjusted HR 0.82, 95% CI 0.76-0.89, P < 0.001). On multivariable analysis, HER2-low status was prognostic (RFS: HR 0.90, 95% CI 0.85-0.96, P = 0.002; OS: HR 0.86, 95% CI 0.79-0.93, P < 0.001). These differences remained significant in hormone receptor-positive tumours and for OS in hormone receptor-negative tumours. Superior outcomes were observed for HER2 IHC1+ BC versus HER2-zero BC (RFS: HR 0.89, 95% CI 0.83-0.96, P = 0.001; OS: HR 0.85, 95% CI 0.78-0.93, P = 0.001). No significant differences were seen between HER2 IHC2+ ISH- and HER2-zero BCs. In the TCGA-BRCA and METABRIC cohorts, ERBB2 CNV status was an independent RFS prognostic factor (neutral versus non-neutral HR 0.71, 95% CI 0.59-0.86, P < 0.001); no differences in RFS by ERBB2 mRNA expression levels were found. CONCLUSIONS: HER2-low BC had a superior prognosis compared to HER2-zero BC in the non-metastatic setting, though absolute differences were modest and driven by HER2 IHC 1+ BC. ERBB2 CNV merits further investigation in HER2-negative BC.