Empirical evaluation of language modeling to ascertain cancer outcomes from clinical text reports

BACKGROUND

Longitudinal data on key cancer outcomes for clinical research, such as response to treatment and disease progression, are not captured in standard cancer registry reporting. Manual extraction of such outcomes from unstructured electronic health records is a slow, resource-intensive process. Natural language processing (NLP) methods can accelerate outcome annotation, but they require substantial labeled data. Transfer learning based on language modeling, particularly using the Transformer architecture, has achieved improvements in NLP performance. However, there has been no systematic evaluation of NLP model training strategies on the extraction of cancer outcomes from unstructured text.

RESULTS

We evaluated the performance of nine NLP models at the two tasks of identifying cancer response and cancer progression within imaging reports at a single academic center among patients with non-small cell lung cancer. We trained the classification models under different conditions, including training sample size, classification architecture, and language model pre-training. The training involved a labeled dataset of 14,218 imaging reports for 1112 patients with lung cancer. A subset of models was based on a pre-trained language model, DFCI-ImagingBERT, created by further pre-training a BERT-based model using an unlabeled dataset of 662,579 reports from 27,483 patients with cancer from our center. A classifier based on our DFCI-ImagingBERT, trained on more than 200 patients, achieved the best results in most experiments; however, these results were marginally better than simpler "bag of words" or convolutional neural network models.

CONCLUSION

When developing AI models to extract outcomes from imaging reports for clinical cancer research, if computational resources are plentiful but labeled training data are limited, large language models can be used for zero- or few-shot learning to achieve reasonable performance. When computational resources are more limited but labeled training data are readily available, even simple machine learning architectures can achieve good performance for such tasks.

Genetic variation in the body mass index of adult survivors of childhood acute lymphoblastic leukemia: A report from the Childhood Cancer Survivor Study and the St. Jude Lifetime Cohort

BACKGROUND

Treatment characteristics such as cranial radiation therapy (CRT) do not fully explain adiposity risk in childhood acute lymphoblastic leukemia (ALL) survivors. This study was aimed at characterizing genetic variation related to adult body mass index (BMI) among survivors of childhood ALL.

METHODS

Genetic associations of BMI among 1458 adult survivors of childhood ALL (median time from diagnosis, 20 years) were analyzed by multiple approaches. A 2-stage genome-wide association study in the Childhood Cancer Survivor Study (CCSS) and the St. Jude Lifetime Cohort Study (SJLIFE) was performed. BMI was a highly polygenic trait in the general population. Within the known loci, the BMI percent variance explained was estimated, and additive interactions (chi-square test) with CRT in the CCSS were evaluated. The role of DNA methylation in CRT interaction was further evaluated in a subsample of ALL survivors.

RESULTS

In a meta-analysis of the CCSS and SJLIFE, 2 novel loci associated with adult BMI among survivors of childhood ALL (LINC00856 rs575792008 and EMR1 rs62123082; P_Meta < 5E-8) were identified. It was estimated that the more than 700 known loci explained 6.2% of the variation in adult BMI in childhood ALL survivors. Within the known loci, significant main effects for 23 loci and statistical interactions with CRT at 9 loci (P < 7.0E-5) were further identified. At 2 CRT-interacting loci, DNA methylation patterns may have differed by age.

CONCLUSIONS

Adult survivors of childhood ALL have genetic heritability for BMI similar to that observed in the general population. This study provides evidence that treatment with CRT can modify the effect of genetic variants on adult BMI in childhood ALL survivors.

Abstract 1575: DNA methylation-derived neutrophil-to-lymphocyte ratio and survival among pediatric medulloblastoma patients

Background/Objectives: Medulloblastoma is the most common malignant brain tumor diagnosed in children. Current risk-stratification methods of pediatric medulloblastoma do not fully explain the observed variability in clinical outcomes. Methylation-derived neutrophil-to-lymphocyte ratio (mdNLR) captures immune-specific information and has been identified as a potential prognostic biomarker of outcomes in various cancers, including adult brain tumors. Therefore, we evaluated the association between blood-derived mdNLR and overall survival in a cohort of patients with pediatric medulloblastoma.

Design/Methods: We identified pediatric patients diagnosed and treated for medulloblastoma at Texas Children’s Cancer Center between 1995 and 2015. Peripheral blood DNA methylation was measured using the Infinium Human Methylation 450K Beadchip. Methylation data underwent quality control, including beta-mixture quantile normalization and batch correction. Immune cell proportions (CD4+ T-Cells, CD8+ T-Cells, B-cells, natural killer cells, monocytes, and granulocytes) were estimated using cellular deconvolution methods and used to estimate mdNLR, which was then log-transformed to improve normality. Cox regression models were estimated to evaluate the association between mdNLR and overall survival.

Results: Of the 78 eligible patients include in this analysis, 83% (n=65) were alive at last follow-up (median follow-up=7.8 years). Deceased patients (n=13; median follow-up=2.6 years) had a higher mean mdNLR than patients who were alive at last contact (12.3 vs. 4.0, P = 0.046). Elevated log-transformed mdNLR was associated with an increased risk of death in both unadjusted models (HR=1.68, 95%CI: 1.11-2.55) and models accounting for age, sex, race, and clinical risk group (HR=1.97, 95%CI: 1.12-3.45).

Conclusion: We identified a significant association between peripheral blood mdNLR and survival in pediatric medulloblastoma. As a promising prognostic biomarker of outcomes, mdNLR captures immune-specific information and is a potential avenue of research in settings where cytologic determination of NLR may not be possible. Future work should investigate the relationship between elevated mdNLR and specific pediatric medulloblastoma molecular subtypes.

Citation Format: Vidal M. Arroyo, Philip J. Lupo, Michael E. Scheurer, Surya P. Rednam, Jeffrey C. Murray, M F. Okcu, Murali M. Chintagumpala, Austin L. Brown. DNA methylation-derived neutrophil-to-lymphocyte ratio and survival among pediatric medulloblastoma patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1575.

Abstract 3287: Survival disparities for second primary malignancies diagnosed among childhood cancer survivors: A population-based assessment

Background: Curative therapy places childhood cancer survivors at increased risk of second primary malignancies (SPMs). However, there have been few population-based attempts to characterize differences in outcomes between SPMs in childhood cancer survivors and comparable de novo first primary malignancies (FPMs).

Methods: We extracted clinical and demographic information from childhood cancer survivors who developed SPMs and from individuals with comparable de novo FPMs using the Surveillance, Epidemiology, and End Results (SEER) 1973-2015 database. Hazard ratios (HR) and 95% confidence intervals (CI) were estimated with Cox proportional hazards models comparing overall survival (OS) between individuals with and without a history of childhood cancer. OS was evaluated overall and within specific cancers diagnosed in ≥50 childhood cancer survivors. Models accounted for potential confounders, including sex, race, age, decade of diagnosis, histology, and disease stage.

Results: Compared to individuals with FPMs (n=1,332,203), childhood cancer survivor with a similar SPM as the FPMs (n=1,409) experienced poorer OS (HR=1.86, 95%CI: 1.72-2.02) after accounting for age, sex, race, and decade of diagnosis. Estimated five-year overall survival for SPMs diagnosed in survivors of childhood cancer was 61.7% (95% CI: 58.9-64.4), compared to 77.4% (95% CI: 75.0-79.5) among de novo FPMs propensity score-matched on confounding factors. A history of childhood cancer remained a poor prognostic factor for all specific cancer evaluated, including: breast (HR=2.07, 95%CI: 1.63-2.62), thyroid (HR=3.59, 95%CI: 2.08-6.19), acute myeloid leukemia (HR=2.38, 95%CI:1.87-3.05), brain (HR=2.09, 95%CI:1.72-2.55), melanoma (HR=2.57, 95%CI: 1.55-4.27), bone (HR=1.88, 95%CI:1.37-2.57), and soft tissue sarcoma (HR=2.44, 95%CI: 1.78-3.33). Survival disparities were most pronounced in survivors diagnosed with a childhood cancer before age 10 (HR=2.83, 95%CI: 2.46-3.25), diagnosed with a SPM within 10 years of the childhood cancer (HR=2.61, 95%CI: 2.29-2.99), and exposed to radiotherapy during childhood (HR=2.13, 95%CI: 1.92-2.37).

Conclusion: Compared to individuals without a prior cancer diagnosis, survivors of childhood cancer with a SPM experience inferior outcomes. These survival disparities persist across cancer types, therapeutic exposures, and clinical factors.

Citation Format: Austin L. Brown, Vidal M. Arroyo, Jennifer Agrusa, Michael E. Scheurer, Maria M. Gramatges, Philip J. Lupo. Survival disparities for second primary malignancies diagnosed among childhood cancer survivors: A population-based assessment [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 3287.

Survival disparities for second primary malignancies diagnosed among childhood cancer survivors: A population-based assessment

BACKGROUND

Curative therapy places childhood cancer survivors at increased risk for second primary malignancies (SPMs). However, there have been few population-based attempts to characterize differences between outcomes of SPMs in childhood cancer survivors and outcomes of first primary malignancies (FPMs).

METHODS

Clinical and demographic information about childhood cancer survivors who developed SPMs and individuals with comparable FPMs was extracted from the Surveillance, Epidemiology, and End Results program. Hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated with Cox proportional hazards models comparing the overall survival (OS) of individuals with and without a history of childhood cancer. OS was evaluated both overall and for specific cancers diagnosed in 50 or more childhood cancer survivors. Models accounted for potential confounders, including sex, race, age, treatment decade, histology, and disease stage.

RESULTS

Compared with individuals with FPMs (n = 1,332,203), childhood cancer survivors (n = 1409) with an SPM experienced poorer OS (HR, 1.86; 95% CI, 1.72-2.02) after the study had accounted for cancer type, age, sex, race, and decade of diagnosis. A history of childhood cancer remained a poor prognostic factor for all specific cancers evaluated, including breast cancer (HR, 2.07; 95% CI, 1.63-2.62), thyroid cancer (HR, 3.59; 95% CI, 2.08-6.19), acute myeloid leukemia (HR, 2.38; 95% CI, 1.87-3.05), brain cancer (HR, 2.09; 95% CI, 1.72-2.55), melanoma (HR, 2.57; 95% CI, 1.55-4.27), bone cancer (HR, 1.88; 95% CI, 1.37-2.57), and soft-tissue sarcoma (HR, 2.44; 95% CI, 1.78-3.33).

CONCLUSIONS

Compared with individuals without a prior cancer diagnosis, survivors of childhood cancer with an SPM experienced inferior outcomes. Survival disparities were observed for the most frequent SPMs diagnosed in childhood cancer survivors.

Pilot study of DNA methylation-derived neutrophil-to-lymphocyte ratio and survival in pediatric medulloblastoma

INTRODUCTION

Methylation-derived neutrophil-to-lymphocyte ratio (mdNLR) has been identified as a potential prognostic biomarker of outcomes in various cancers. We evaluated the prognostic value of blood-derived mdNLR within a retrospective cohort of pediatric medulloblastoma patients.

MATERIALS AND METHODS

DNA methylation was measured in archival peripheral blood samples collected on 56 pediatric medulloblastoma patients. Hazard ratios (HR) and 95% confidence intervals (CI) for the association between mdNLR and survival were evaluated using Cox proportional hazard models.

RESULTS

Compared to patients who were alive at last follow-up (n = 43), the mean mdNLR value was slightly higher in deceased patients (n = 13) (12.3 vs. 5.2,P =  0.163). Elevated log-transformed mdNLR was suggestively associated with an increased likelihood of death in unadjusted models (HR=1.43, 95%CI: 0.92-2.22) and significantly associated with mortality in adjusted models (HR=1.61, 95%CI: 1.01-2.58).

DISCUSSION

Future work is warranted to investigate the relationship between mdNLR outcomes in specific pediatric medulloblastoma molecular subgroups.

DNA methylation and obesity in survivors of pediatric acute lymphoblastic leukemia: A report from the Childhood Cancer Survivor Study

Because survivors of pediatric acute lymphoblastic leukemia (ALL) are more likely to be obese than unaffected contemporaries, we compared DNA methylation profiles between normal-weight and obese survivors at adiposity-associated CpG sites previously-reported by epigenome-wide association studies (EWAS) of body mass index (BMI) in the general population. We selected 96 ALL survivors from the Childhood Cancer Survivor Study: 48 obese and 48 normal weight. The Illumina HumanMethylation450 BeadChip was used to compare DNA methylation at 211 loci identified in EWAS of BMI in the general population. Thirty-nine loci were associated (false discovery rate <0.05) with obesity among survivors who only received chemotherapy (n = 49). No loci were significantly associated with obesity among CRT-exposed survivors (n = 47). Our results suggest that previously identified BMI-DNA methylation loci are associated with obesity in ALL survivors who were spared CRT, while no loci were significantly associated with obesity in survivors who received CRT.

Abstract 1197: Soft tissue sarcoma clinical presentation, treatment, and survival in adolescents and young adults compared to older adults: A report from the Scandinavian Sarcoma Group

Background: Five-year survival rates for those diagnosed with soft tissue sarcoma (STS) have improved significantly among children and older adults (OAs), but these same trends have not been observed for adolescents and young adults (AYAs). While these disparities could be due to differences in biology or treatment, few studies have evaluated STS occurrence and outcome in AYAs. Therefore, the purpose of this study was to evaluate differences between adolescents and young adults (AYAs) and older adults (OAs) diagnosed with STS by stratifying analysis by: (1) clinical presentation; (2) treatment; and (3) survival.

Methods: Data were obtained from the Scandinavian Sarcoma Group (SSG) Central Register, which includes information on 5,747 patients from Sweden and Norway, diagnosed with a STS during 1986-2011. Variables included: age at diagnosis, metastasis at diagnosis, tumor size, histology, adjuvant treatment, date of death or last follow-up. AYAs were defined as those diagnosed ages 15-39 years. Categorical variables were analyzed using chi-square tests, and continuous variables were analyzed using t-tests. Overall survival (OS) and recurrence-free survival (RFS) were compared between AYAs and OAs using Kaplan-Meier estimates and log-rank tests. All analyses were conducted overall and by common STS subtypes.

Results: Overall and by STS subtype, there were significant differences between AYAs and OAs on presentation, treatment, and survival. The distribution of STS subtypes was different between OAs and AYAs. For example, OAs were more likely to be diagnosed with leiomyosarcoma compared to AYAs (18% vs. 10%, p&lt;0.001), whereas AYAs were more likely to be diagnosed with malignant peripheral nerve sheath tumor (MPNST, 9% vs. 4%, p&lt;0.001). OAs were also more likely to have larger tumors (&gt;5 cm, 67% vs. 52%, p&lt;0.001) and higher malignancy grade (grade IV, 45% vs. 31%, p&lt;0.001). Interestingly, AYAs were more likely to be treated with radiotherapy and chemotherapy compared to OAs (12% vs. 5%, p&lt;0.001). There were also differences within STS subtypes. For example, OAs were more likely to have metastasis compared to AYAs if diagnosed with leiomyosarcoma (18% vs. 10%, p=0.04). In most scenarios AYAs had significantly better OS and RFS compared to OAs, other than for MPNST (OS: p=0.19, RFS: p=0.28).

Conclusions: There were several differences between AYAs and OAs on STS presentation, treatment, and outcome. AYAs not only had differences in terms of STS subtypes but also tumor size and malignancy grade within subtypes. Additional work is needed to characterize the biology underlying these differences, which will inform future treatment strategies for both AYAs and OAs with STS.

Citation Format: Vidal M. Arroyo, Philip J. Lupo, Beatrice Melin, Emelie Styring, Olga Zaikova, Karin Papworth. Soft tissue sarcoma clinical presentation, treatment, and survival in adolescents and young adults compared to older adults: A report from the Scandinavian Sarcoma Group [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 1197.