Limitations of Transformers on Clinical Text Classification

Bidirectional Encoder Representations from Transformers (BERT) and BERT-based approaches are the current state-of-the-art in many natural language processing (NLP) tasks; however, their application to document classification on long clinical texts is limited. In this work, we introduce four methods to scale BERT, which by default can only handle input sequences up to approximately 400 words long, to perform document classification on clinical texts several thousand words long. We compare these methods against two much simpler architectures - a word-level convolutional neural network and a hierarchical self-attention network - and show that BERT often cannot beat these simpler baselines when classifying MIMIC-III discharge summaries and SEER cancer pathology reports. In our analysis, we show that two key components of BERT - pretraining and WordPiece tokenization - may actually be inhibiting BERT's performance on clinical text classification tasks where the input document is several thousand words long and where correctly identifying labels may depend more on identifying a few key words or phrases rather than understanding the contextual meaning of sequences of text.

Diagnosis-to-surgery interval and survival for different histologies of stage I-IIA lung cancer

Background

Guidelines on timeliness of lung cancer surgery are inconsistent. Lung cancer histologic subtypes have different prognosis and treatment. It is important to understand the consequences of delayed surgery for each lung cancer histologic subtype. This study aimed to examine the association between diagnosis-to-surgery time interval and survival for early stage lung cancer and selected histologic subtypes.

Methods

Patients diagnosed with stage I–IIA lung cancer between 2004 and 2015 receiving definitive surgery and being followed up until Dec. 31, 2018, were identified from Surveillance, Epidemiology, and End Results database. Histologic subtypes included adenocarcinoma, squamous or epidermoid carcinoma, bronchioloalveolar carcinoma, large cell carcinoma, adenosquamous carcinoma, carcinoid carcinoma, and small cell carcinoma. Diagnosis-to-surgery interval was treated as multi-categorical variables (<1, 1–2, 2–3, and ≥3 months) and binary variables (≥1 vs. <1 month, ≥2 vs. <2 months, and ≥3 vs. <3 months). Outcomes included cancer-specific and overall survival. Covariates included age at diagnosis, sex, race, marital status, tumor size, grade, surgery type, chemotherapy, radiotherapy, and study period. Kaplan-Meier survival curves and Cox proportional hazards regression models were applied to examine the survival differences.

Results

With a median follow-up time of 51 months, a total of 40,612 patients were analyzed, including 40.1% adenocarcinoma and 24.5% squamous or epidermoid carcinoma. The proportion of patients receiving surgery <1, 1–2, 2–3, and ≥3 months from diagnosis were 34.2%, 33.9%, 19.8%, and 12.1%, respectively. Delayed surgery was associated with worse cancer-specific and overall survival for all lung cancers, adenocarcinoma, squamous or epidermoid, bronchioloalveolar, and large cell carcinoma (20–40% increased risk). Dose-dependent effects (longer delay, worse survival) were observed in all lung cancers, adenocarcinoma, and squamous and epidermoid carcinoma. No significant association between surgery delay and survival was observed in adenosquamous, carcinoid, and small cell carcinoma.

Conclusions

Our findings support the guidelines of undertaking surgery within 1 month from diagnosis in patients with stage I–IIA lung cancer. The observed dose-dependent effects emphasize the clinical importance of early surgery. Future studies with larger sample size of less frequent histologic subtypes are warranted to provide more evidence for histology-specific lung cancer treatment guidelines.

Privacy-Preserving Deep Learning NLP Models for Cancer Registries

Population cancer registries can benefit from Deep Learning (DL) to automatically extract cancer characteristics from the high volume of unstructured pathology text reports they process annually. The success of DL to tackle this and other real-world problems is proportional to the availability of large labeled datasets for model training. Although collaboration among cancer registries is essential to fully exploit the promise of DL, privacy and confidentiality concerns are main obstacles for data sharing across cancer registries. Moreover, DL for natural language processing (NLP) requires sharing a vocabulary dictionary for the embedding layer which may contain patient identifiers. Thus, even distributing the trained models across cancer registries causes a privacy violation issue. In this paper, we propose DL NLP model distribution via privacy-preserving transfer learning approaches without sharing sensitive data. These approaches are used to distribute a multitask convolutional neural network (MT-CNN) NLP model among cancer registries. The model is trained to extract six key cancer characteristics - tumor site, subsite, laterality, behavior, histology, and grade - from cancer pathology reports. Using 410,064 pathology documents from two cancer registries, we compare our proposed approach to conventional transfer learning without privacy-preserving, single-registry models, and a model trained on centrally hosted data. The results show that transfer learning approaches including data sharing and model distribution outperform significantly the single-registry model. In addition, the best performing privacy-preserving model distribution approach achieves statistically indistinguishable average micro- and macro-F1 scores across all extraction tasks (0.823,0.580) as compared to the centralized model (0.827,0.585).

Association of the COVID-19 Pandemic with Patterns of Statewide Cancer Services

The coronavirus disease 2019 (COVID-19) pandemic led to delayed medical care in the US. We examined changes in patterns of cancer diagnosis and surgical treatment between January 1 and December 31 in 2020 and 2019 with real-time electronic pathology report data from population-based Surveillance, Epidemiology, and End Results cancer registries from Georgia and Louisiana. During 2020, there were 29,905 fewer pathology reports than in 2019, representing a 10.2% decline. Declines were observed in all age groups, including children and adolescents less than18 years. The nadir was early April 2020, with 42.8% fewer reports than in April 2019. Numbers of reports through December 2020 never consistently exceeded those in 2019 after first declines. Patterns were similar by age group and cancer site. Findings suggest substantial delays in diagnosis and treatment services for cancers during the pandemic. Ongoing evaluation can inform public health efforts to minimize any lasting adverse effects of the pandemic on cancer diagnosis, stage, treatment, and survival.

Sexual function outcomes of radiation and androgen deprivation therapy for localized prostate cancer in men with good baseline function

BACKGROUND

Sexual dysfunction, including erectile dysfunction and loss of libido, are common among men undergoing treatment for localized prostate cancer. Both local treatments and systemic androgen deprivation therapy may contribute to these outcomes and are differentially indicated based on disease characteristics. We sought to compare sexual function through 5 years after radiation treatment with and without androgen deprivation therapy in men with good baseline sexual function to better understand long-term effects in this understudied subset of patients.

METHODS

We retrospectively reviewed a prospectively assembled population-based cohort of men who underwent radiation with and without androgen deprivation therapy for intermediate or high-risk localized prostate cancer. Sexual function was assessed longitudinally over 5 years. Men with erections sufficient for intercourse at baseline were selected for inclusion.

RESULTS

Out of 167 patients included, 73 underwent radiation alone and 94 received androgen deprivation therapy plus radiation (51 with intermediate and 43 with high-risk disease). Androgen deprivation therapy use was associated with worse sexual function through 1 year regardless of disease risk. This difference was no longer statistically significant at 3 years in the intermediate-risk group. Compared to radiation alone, androgen deprivation therapy in high-risk disease was associated with worse sexual function at 3 years (effect: -20.3 points, CI [-31.8, -8.8], p < 0.001) but not at 5 years (effect: -3.4, CI [-17.2, 10.5], p = 0.63).

CONCLUSIONS

Androgen deprivation therapy plus radiation is associated with worse sexual function through 3-years follow-up in men with high-risk prostate cancer compared to radiation alone. The addition of androgen deprivation therapy in the treatment of intermediate-risk disease does not appear to result in worse sexual function at 3 or 5-year follow-up compared to radiation alone.

Racial Disparity and Social Determinants in Receiving Timely Surgery Among Stage I-IIIA Non-small Cell Lung Cancer Patients in a U.S. Southern State

Delayed surgery is associated with worse lung cancer outcomes. Social determinants can influence health disparities. This study aimed to examine the potential racial disparity and the effects from social determinants on receipt of timely surgery among lung cancer patients in Louisiana, a southern state in the U.S. White and black stage I-IIIA non-small cell lung cancer patients diagnosed in Louisiana between 2004 and 2016, receiving surgical lobectomy or a more extensive surgery, were selected. Diagnosis-to-surgery interval >6 weeks were considered as delayed surgery. Social determinants included marital status, insurance, census tract level poverty, and census tract level urbanicity. Multivariable logistic regression and generalized multiple mediation analysis were conducted. A total of 3,616 white (78.9%) and black (21.1%) patients were identified. The median time interval from diagnosis to surgery was 27 days in whites and 42 days in blacks (P < 0.0001). About 28.7% of white and 48.4% of black patients received delayed surgery (P < 0.0001). Black patients had almost two-fold odds of receiving delayed surgery than white patients (adjusted odds ratio: 1.91; 95% confidence interval: 1.59-2.30). Social determinants explained about 26% of the racial disparity in receiving delayed surgery. Having social support, private insurance, and living in census tracts with lower poverty level were associated with improved access to timely surgery. The census tract level poverty level a stronger effect on delayed surgery in black patients than in white patients. Tailored interventions to improve the timely treatment in NSCLC patients, especially black patients, are needed in the future.

Association of HIV infection with survival and cancer-specific death in cancer patients in Louisiana

e18630

Background: Studies have reported elevated mortality in HIV patients (pts), including cancer-specific mortality, compared to non-HIV pts. However, many of these studies did not include data from Louisiana. Louisiana has among the highest HIV and AIDS case rates in the country, but there is no available large-scale data regarding cancer outcomes in HIV pts in the state. We compared our data in Louisiana to the largest known study to date over an expanded time period to evaluate for disparities. Methods: Following linkage of Louisiana Tumor Registry data from 1995-2016 and Louisiana Office of Public Health, STD/HIV Program database, we identified a total of 435,478 cancer cases, of which 2,949 were HIV-positive (0.67%). Analyses of categorical variables were performed using Pearson’s chi-squared test. Cause of death was extracted from death certificate or autopsy report. Survival analyses were performed utilizing SEER*Stat. Results: Despite the higher proportion of HIV in cancer pts in Louisiana, demographics are similar to previously published data in that cancer pts with HIV tended to be male, black, and younger than their HIV-negative counterparts (p<0.001). However, although 5-year survival was substantially decreased in HIV-cancer pts in congruence with other studies, cancer-specific mortality was lower than non-HIV pts in all cohorts, including both AIDS-defining cancers (ADC) and non-AIDS-defining cancers (non-ADC). Conclusions: HIV infection in cancer pts in Louisiana is associated with lower survival but not increased cancer-specific death, independent of race, gender, age, ADC vs non-ADC, and viral vs non-viral etiology. HIV-related deaths still comprise a substantial cause of death in this population.[Table: see text]

Association of the COVID-19 pandemic with patterns of cancer services

1514

Background: The COVID-19 pandemic led to delays in medical care in the United States. We examined changes in patterns of cancer diagnosis and surgical treatment in 2020 using real-time electronic pathology report data from population-based SEER cancer registries in Georgia and Louisiana. Methods: Bi-weekly numbers, distributions, and patterns of pathology reports were compared between January 1st and December 31st in 2020 and the same period in 2019 by age group and cancer site. Results: During 2020, there were 29,905 fewer pathology reports than in 2019, representing a 10.2% decline. Absolute declines were greatest among adults aged ≥50 years (N=23,065); percentage declines were greatest among children and young adults ≤18 years (38.3%). By cancer site, percentage declines were greatest for lung cancer (17.4%), followed by colorectal (12.0%), breast (9.0%) and prostate (5.8%) cancers. Biweekly reports were statistically significantly lower in 2020 than in 2019 from late March through the end of December in most biweekly periods. The nadir was the month of April 2020 – the number of reports was at least 40% lower than in April 2019. The number of reports in 2020 compared with 2019 also declined sharply in early November (26.8%) and late December (32.0%). Numbers of reports in 2020 never consistently exceeded those in 2019 after the first decline. Patterns were similar by cancer site, with variation in magnitude and duration of declines. Conclusions: Significant declines in cancer pathology reports from population-based registries during 2020 suggest substantial delays in screening, evaluation of signs and symptoms, diagnosis, and treatment services for cancers with effective screening tests as well as in cancer sites and age groups without effective screening tests as an indirect result of the COVID-19 pandemic. Ongoing evaluation will be critical for informing public health efforts to minimize any lasting adverse effects of the pandemic on cancer screening, diagnosis, treatment, and survival.[Table: see text]

[Evaluation of intervention effect in the occupational protection of glass fiber workers by occupational health education]

Objective: To understand the mastery of occupational hygienic knowledge and the implementing of occupational health protection measures in the group which were exposed to the procedure of manufacture and use in glass fiber company, and to explore the feasibility of the prevention of the skin injury by occupational health education in glass fiber workers. Methods: We selected 257 on-the-job employees as the research object in a ceramic enterprise in Nanjing from June 2018 to August 2019, with the method of cluster random sampling. According to Solomon's design, the intervention group in which we took measures with health education was divided into RG(1) (O(1)XO(2)) and RG(3) (XO(5)) group, and the control group where we didn't take any intervention was divided into RG(2) (O(3)-O(4)) and RG(4) (-O(6)) group. The intervention effect of health education on the occupational protection of glass fiber workers was evaluated by the results of questionnaire. Results: After training, the average score of occupational health knowledge in the intervention group was 27.34 points higher than that before training, the intervention index was 1.42, 23.62-27.73 points higher than the control glass fiber workers and 33.62-35.52 points higher than the control glass non-glass fiber workers; Compared with the control group, the positive attitude rate of fiber glass workers in the intervention group increased by 13.28%, 13.51%, 11.68% and 11.48%, and the intervention indexes were 1.18, 1.17, 1.14 and 1.15, which was corresponding to using protective cream, wearing gloves, wearing working clothes, washing hands and bathing after work, respectively; Compared with the control group, the implementation rate of occupational protection measures which were represented by wearing gloves、washing hands and bathing for glass fiber workers in the intervention group increased by 29.25% and 7.27% respectively, and the intervention indexes were 1.43 and 1.08 respectively; The skin injury rate of fiberglass workers in the intervention group was reduced by 11.43% comparing to the control group, the intervention index was 1.67. Conclusion: According to the occupational health education of fiberglass workers, it improves the mastery of occupational health knowledge, positive atti-tude rate and the implementation rate of occupational protection measures, meanwhile, it reduces the skin injury rate of the intervention objects to a certain extent.

Understanding the patient journey to diagnosis of lung cancer

Objective

This research describes the clinical pathway and characteristics of two cohorts of patients. The first cohort consists of patients with a confirmed diagnosis of lung cancer while the second consists of patients with a solitary pulmonary nodule (SPN) and no evidence of lung cancer. Linked data from an electronic medical record and the Louisiana Tumor Registry were used in this investigation.

Materials and methods

REACHnet is one of 9 clinical research networks (CRNs) in PCORnet®, the National Patient-Centered Clinical Research Network and includes electronic health records for over 8 million patients from multiple partner health systems. Data from Ochsner Health System and Tulane Medical Center were linked to Louisiana Tumor Registry (LTR), a statewide population-based cancer registry, for analysis of patient’s clinical pathways between July 2013 and 2017. Patient characteristics and health services utilization rates by cancer stage were reported as frequency distributions. The Kaplan-Meier product limit method was used to estimate the time from index date to diagnosis by stage in lung cancer cohort.

Results

A total of 30,559 potentially eligible patients were identified and 2929 (9.58%) had primary lung cancer. Of these, 1496 (51.1%) were documented in LTR and their clinical pathway to diagnosis was further studied. Time to diagnosis varied significantly by cancer stage. A total of 24,140 patients with an SPN were identified in REACHnet and 15,978 (66.6%) had documented follow up care for 1 year. 1612 (10%) had no evidence of any work up for their SPN. The remaining 14,366 had some evidence of follow up, primarily office visits and additional chest imaging.

Conclusion

In both cohorts multiple biopsies were evident in the clinical pathway. Despite clinical workup, 70% of patients in the lung cancer cohort had stage III or IV disease. In the SPN cohort, only 66% were identified as receiving a diagnostic work-up.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-021-08067-1.

Deep active learning for classifying cancer pathology reports

Background

Automated text classification has many important applications in the clinical setting; however, obtaining labelled data for training machine learning and deep learning models is often difficult and expensive. Active learning techniques may mitigate this challenge by reducing the amount of labelled data required to effectively train a model. In this study, we analyze the effectiveness of 11 active learning algorithms on classifying subsite and histology from cancer pathology reports using a Convolutional Neural Network as the text classification model.

Results

We compare the performance of each active learning strategy using two differently sized datasets and two different classification tasks. Our results show that on all tasks and dataset sizes, all active learning strategies except diversity-sampling strategies outperformed random sampling, i.e., no active learning. On our large dataset (15K initial labelled samples, adding 15K additional labelled samples each iteration of active learning), there was no clear winner between the different active learning strategies. On our small dataset (1K initial labelled samples, adding 1K additional labelled samples each iteration of active learning), marginal and ratio uncertainty sampling performed better than all other active learning techniques. We found that compared to random sampling, active learning strongly helps performance on rare classes by focusing on underrepresented classes.

Conclusions

Active learning can save annotation cost by helping human annotators efficiently and intelligently select which samples to label. Our results show that a dataset constructed using effective active learning techniques requires less than half the amount of labelled data to achieve the same performance as a dataset constructed using random sampling.

Supplementary Information

The online version supplementary material available at 10.1186/s12859-021-04047-1.

Automatic extraction of cancer registry reportable information from free-text pathology reports using multitask convolutional neural networks

OBJECTIVE

We implement 2 different multitask learning (MTL) techniques, hard parameter sharing and cross-stitch, to train a word-level convolutional neural network (CNN) specifically designed for automatic extraction of cancer data from unstructured text in pathology reports. We show the importance of learning related information extraction (IE) tasks leveraging shared representations across the tasks to achieve state-of-the-art performance in classification accuracy and computational efficiency.

MATERIALS AND METHODS

Multitask CNN (MTCNN) attempts to tackle document information extraction by learning to extract multiple key cancer characteristics simultaneously. We trained our MTCNN to perform 5 information extraction tasks: (1) primary cancer site (65 classes), (2) laterality (4 classes), (3) behavior (3 classes), (4) histological type (63 classes), and (5) histological grade (5 classes). We evaluated the performance on a corpus of 95 231 pathology documents (71 223 unique tumors) obtained from the Louisiana Tumor Registry. We compared the performance of the MTCNN models against single-task CNN models and 2 traditional machine learning approaches, namely support vector machine (SVM) and random forest classifier (RFC).

RESULTS

MTCNNs offered superior performance across all 5 tasks in terms of classification accuracy as compared with the other machine learning models. Based on retrospective evaluation, the hard parameter sharing and cross-stitch MTCNN models correctly classified 59.04% and 57.93% of the pathology reports respectively across all 5 tasks. The baseline models achieved 53.68% (CNN), 46.37% (RFC), and 36.75% (SVM). Based on prospective evaluation, the percentages of correctly classified cases across the 5 tasks were 60.11% (hard parameter sharing), 58.13% (cross-stitch), 51.30% (single-task CNN), 42.07% (RFC), and 35.16% (SVM). Moreover, hard parameter sharing MTCNNs outperformed the other models in computational efficiency by using about the same number of trainable parameters as a single-task CNN.

CONCLUSIONS

The hard parameter sharing MTCNN offers superior classification accuracy for automated coding support of pathology documents across a wide range of cancers and multiple information extraction tasks while maintaining similar training and inference time as those of a single task-specific model.

Five-year outcomes from a prospective comparative effectiveness study evaluating external-beam radiotherapy with or without low-dose-rate brachytherapy boost for localized prostate cancer

BACKGROUND

To inform patients who are in the process of selecting prostate cancer treatment, the authors compared disease-specific function after external-beam radiotherapy (EBRT) alone versus EBRT plus a low-dose-rate (LDR) brachytherapy boost (EBRT-LDR).

METHODS

For this prospective study, men who had localized prostate cancer in 2011 and 2012 were enrolled. Assessments at baseline, 0.5, 1, 3, and 5 years included the patient-reported Expanded Prostate Index Composite, the 36-item Medical Outcomes Study Short-Form Health Survey, and treatment-related regret. Regression models were adjusted for baseline function and for patient and treatment characteristics. The minimum clinically important difference in scores on the Expanded Prostate Index Composite 26-item instrument was from 5 to 7 for urinary irritation and from 4 to 6 for bowel function.

RESULTS

Six-hundred ninety-five men met inclusion criteria and received either EBRT (n = 583) or EBRT-LDR (n = 112). Patients in the EBRT-LDR group were younger (median age, 66 years [interquartile range [IQR], 60-71 years] vs 69 years [IQR, 64-74 years]; P < .001), were less likely to receive pelvic radiotherapy (10% vs 18%; P = .040), and had higher baseline 36-item Medical Outcomes Study Short-Form Health Survey physical function scores (median score, 95 [IQR, 86-100] vs 90 [IQR, 70-100]; P < .001). Over a 3-year period, compared with EBRT, EBRT-LDR was associated with worse urinary irritative scores (adjusted mean difference at 3 years, -5.4; 95% CI, -9.3, -1.6) and bowel function scores (-4.1; 95% CI, -7.6, -0.5). The differences were no longer clinically meaningful at 5 years (difference in urinary irritative scores: -4.5; 95% CI, -8.4, -0.5; difference in bowel function scores: -2.1; 95% CI, -5.7, -1.4). However, men who received EBRT-LDR were more likely to report moderate or big problems with urinary function bother (adjusted odds ratio, 3.5; 95% CI, 1.5-8.2) and frequent urination (adjusted odds ratio, 2.6; 95% CI, 1.2-5.6) through 5 years. There were no differences in survival or treatment-related regret between treatment groups.

CONCLUSIONS

Compared with EBRT alone, EBRT-LDR was associated with clinically meaningful worse urinary irritative and bowel function over 3 years after treatment and more urinary bother at 5 years.

LAY SUMMARY

In men with prostate cancer who received external-beam radiation therapy (EBRT) with or without a brachytherapy boost (EBRT-LDR), EBRT-LDR was associated with clinically worse urinary irritation and bowel function through 3 years but resolved after 5 years. Men who received EBRT-LDR continued to report moderate-to-big problems with urinary function bother and frequent urination through 5 years. There was no difference in treatment-related regret or survival between patients who received EBRT and those who received EBRT-LDR. These intermediate-term estimates of function may facilitate counseling for men who are selecting treatment.

Abstract PS7-40: Triple negative breast cancer in New Orleans, part deux

Version 7/7/2020 with edit Triple-Negative Breast Cancer in New Orleans – part deaux Objectives: To update our previously reported findings with more current data from SEER 18 registries to compare race-specific incidence rates of triple-negative breast cancer (TNBC) in New Orleans (NO) and Louisiana (LA) with other SEER metro areas and identify the predictors of TNBC. Background: We previously reported significantly increased incidence rates of TNBC in black women (BW) in NO and LA than other SEER metro areas. Methods: We analyzed tumor characteristics of invasive female breast cancers diagnosed from 2011-2017 from SEER 18 for BW and white women (WW). We compared LA data with SEER 18 and metropolitan areas (Atlanta, Detroit, Los Angeles and San Francisco) as we did with our previous data set. Predictors of TNBC in NO and LA were identified in multivariate logistic regression adjusting for age, BMI, poverty, insurance, tumor size, Bloom-Richardson grade and AJCC stage. Results: The overall incidence rate of TNBC in BW was again significantly higher in NO (29.8 per 100,000) than any metro area and the SEER metro areas combined (23.9 per 100,000). Detroit again had the second highest rate (27.2) followed by San Francisco/Oakland (23.8) Los Angeles (22.3) and Atlanta (22.1). The multivariable analysis showed that younger age, large tumor size, high grade were associated with increased risk of TNBC in BW. Conclusions: The incidence rates of TNBC in BW in NO continue to be consistently higher compared to LA and other SEER 18 metro areas. This data reiterates the importance of ongoing translational research to broaden the understanding and optimize the treatment of this aggressive disease, especially in this vulnerable population.

Citation Format: Michelle M Loch, Mei-Chin Hsieh, Thomas Reske, Vivien Chen, Xiao-Cheng Wu. Triple negative breast cancer in New Orleans, part deux [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS7-40.

Contributions of Adenocarcinoma and Carcinoid Tumors to Early-Onset Colorectal Cancer Incidence Rates in the United States

BACKGROUND

Early-onset colorectal cancer (EOCRC) incidence rates (IRs) are rising, according to previous cancer registry analyses. However, analysis of histologic subtypes, including adenocarcinoma (the focus of CRC screening and diagnostic testing) and carcinoid tumors (which are classified as "colorectal cancer" in SEER [Surveillance, Epidemiology, and End Results] databases but have a distinct pathogenesis and are managed differently from adenocarcinoma), has not been reported.

OBJECTIVE

To assess EOCRC IRs and changes in IRs over time, stratified by histology.

DESIGN

Retrospective analysis.

SETTING

Yearly IRs according to SEER 18 data from 2000 to 2016 on age-specific colon-only, rectal-only, and combined-site CRC cases, stratified by histology ("overall" CRC [all histologic subtypes], adenocarcinoma, and carcinoid tumors) and age.

PATIENTS

119 624 patients with CRC.

MEASUREMENTS

IRs per 100 000 population, changes in 3-year average annual IRs (pooled IRs from 2000 to 2002 vs. those from 2014 to 2016), and annual percentage change (APC) in persons aged 20 to 29, 30 to 39, 40 to 49, and 50 to 54 years.

RESULTS

The steepest changes in adenocarcinoma 3-year average annual IRs were for rectal-only cases in persons aged 20 to 29 years (+39% [0.33 to 0.46 per 100 000]; P < 0.050) and 30 to 39 years (+39% [1.92 to 2.66 per 100 000]; P < 0.050) and colon-only cases in those aged 30 to 39 years (+20% [3.30 to 3.97 per 100 000]; P < 0.050). Corresponding APCs were 1.6% (P < 0.050), 2.2% (P < 0.050), and 1.2% (P < 0.050), respectively. In persons aged 40 to 49 years, 3-year average annual IRs increased in both colon-only (+13% [12.21 to 13.85 per 100 000]; P < 0.050) and rectal-only (+16% [7.50 to 8.72 per 100 000]; P < 0.050) subsites. Carcinoid tumors were common, representing approximately 4% to 20% of all colorectal and 8% to 34% of all rectal cancer cases, depending on age group and calendar year. Colon-only carcinoid tumors were rare. Colorectal carcinoid tumor IRs increased more steeply than adenocarcinoma in all age groups, thus affecting the contribution of carcinoid tumors to overall cancer cases over time. These changes were driven by rectal subsites and were most pronounced in persons aged 50 to 54 years, in whom rectal carcinoid tumors increased by 159% (2.36 to 6.10 per 100 000) between 2000 to 2002 and 2014 to 2016, compared with 10% for adenocarcinoma (18.07 to 19.84 per 100 000), ultimately accounting for 22.6% of all rectal cancer cases.

LIMITATION

Population-based data.

CONCLUSION

These findings underscore the importance of assessing histologic CRC subtypes independently. Doing so may lead to a better understanding of the drivers of temporal changes in overall CRC incidence and a more accurate measurement of outcomes from efforts to reduce adenocarcinoma risk, and can guide future research.

PRIMARY FUNDING SOURCE

None.

Pathology Laboratory Policies and Procedures for Releasing Diagnostic Tissue for Cancer Research

CONTEXT.—

The Surveillance, Epidemiology, and End Results (SEER) cancer registry program is currently evaluating the use of archival, diagnostic, formalin-fixed, paraffin-embedded (FFPE) tissue obtained through SEER cancer registries, functioning as honest brokers for deidentified tissue and associated data. To determine the feasibility of this potential program, laboratory policies for sharing tissue for research needed to be assessed.

OBJECTIVE.—

To understand the willingness of pathology laboratories to share archival diagnostic tissue for cancer research and related policies.

DESIGN.—

Seven SEER registries administered a 27-item questionnaire to pathology laboratories within their respective registry catchment areas. Only laboratories that processed diagnostic FFPE specimens and completed the questionnaire were included in the analysis.

RESULTS.—

Of the 153 responding laboratories, 127 (83%) responded that they process FFPE specimens. Most (n = 88; 69%) were willing to share tissue specimens for research, which was not associated with the number of blocks processed per year by the laboratories. Most laboratories retained the specimens for at least 10 years. Institutional regulatory policies on sharing deidentified tissue varied considerably, ranging from requiring a full Institutional Review Board review to considering such use exempt from Institutional Review Board review, and 43% (55 of 127) of the laboratories did not know their terms for sharing tissue for research.

CONCLUSIONS.—

This project indicated a general willingness of pathology laboratories to participate in research by sharing FFPE tissue. Given the variability of research policies across laboratories, it is critical for each SEER registry to work with laboratories in their catchment area to understand such policies and state legislation regulating tissue retention and guardianship.

10-Year Survival after Breast-Conserving Surgery Compared with Mastectomy in Louisiana Women with Early-Stage Breast Cancer: A Population-Based Study

BACKGROUND

Louisiana has the third highest breast cancer mortality in the US, despite ranking 30th in incidence. Whether surgical approach contributes to such a poor outcome is unknown. We compared outcomes of breast-conserving surgery plus radiation (BCT) vs mastectomy (MST) for Louisiana women with early-stage breast cancer.

STUDY DESIGN

Data on women diagnosed with Stage I-II breast cancer from 2004 to 2016 were analyzed from the Louisiana Tumor Registry. Overall survival (OS) and breast cancer-specific survival (CSS) were compared between BCT and MST. Kaplan-Meier method and log-rank test were used to compare survival curves, and logistic regression was used to examine the association of sociodemographic and clinical factors with the type of breast cancer surgery. Values of p < 0.05 were considered significant.

RESULTS

Of the 18,260 patients, 9,968 patients (54.6%) had BCT and 8,292 patients (45.4%) had MST. Compared with BCT, the MST group tended to be underinsured/Medicare/Medicaid, more impoverished, had higher stage 2 disease, were more likely to reside in rural regions, travel ≥25 miles to radiation treatment facility, be treated at low volume centers, and have T3, node positive, and poorly differentiated tumors. Ten-year OS and CSS were significantly better among those who had BCT (OS: 80.0%; 95% CI: 79.0%-81.1%; CSS: 92.7%; 95% CI: 92.1%-93.4%) than those having MST (OS: 69.3%; 95% CI: 68.0%-70.5%; CSS: 88.8%:95% CI: 87.9%-89.7%) (p < 0.05). Even after controlling for sociodemographic and clinical variables, MST was associated with a 28.6% increased risk of death from all causes (hazard ratio [HR]:1.286; 95% CI:1.197-1.380) and a 29.8% increased risk of breast-cancer specific death (HR:1.298; 95% CI: 1.150-1.465).

CONCLUSIONS

Surgical approach, a factor that is within the control of the surgeon, has an impact on mortality for Louisiana women with early-stage breast cancer.

Abstract PO-155: Social determinants of receptor positive breast cancer patients’ survival in Louisiana

Breast carcinoma is the most commonly diagnosed cancer in Louisiana women. Four molecular subtypes of breast cancer are classified by the presence or lack of certain hormone receptors (HR), specifically estrogen and progesterone receptors, and the Human epidermal growth factor 2 (HER2) including: Luminal A (HR+/HER2-), Luminal B (HR+/HER2+), HER2-enriched (HR-/HER2+), and Triple-Negative Breast Cancer, TNBC (HR-/HER2-). Although it is now known that genetics has a role in the risk of breast cancer incidence, other factors such as age, race, and neighborhood also have an effect on prognosis and survival. Previously we have reported that in Louisiana neighborhood concentrated disadvantage index (CDI) explained the racial disparities in TNBC stage at diagnosis and survival, but not the disparities present in incidence. The purpose of this study is to assess the association of social determinants in Luminal A, Luminal B, and HER2-enriched breast cancers patients’ survival. This population-based study used 2010-2016 breast cancer data from the Louisiana Tumor Registry (LTR) and the data of the population at risk was obtained from the US Census. To assess neighborhood environment, CDI scores were calculated based on the PhenX Toolkit Protocol. We performed regression analysis of survival data based on the cox proportional hazard models using PHREG. A total of 21,005 breast cancer patient’s data were initially screened for the study with Luminal A (14,437), Luminal B (2,345), HER2-enriched (1,089) and TNBC (3134) diagnosis. Caucasian women (White) represented 68.5%, African women (Black) represented 30.42% and “other” race represented 1.09% of the breast cancer patients. Using Wald Chi-square tests, we compared stage of diagnosis (distant, regional) and CDI among Black and White patients of all three receptor positive breast cancers. For Luminal A, stage of diagnosis, [Distant: (Pr&gt;ChiSq &lt;0.0001), Hazard Ratio, HR:32.4] and Regional: (Pr&gt;ChiSq &lt;0.0001), HR:3.56], and race (Pr&gt;ChiSq &lt;0.0001, HR:1.58]) were significant, but not the CDI [(Pr&gt;ChiSq: &lt; 0.812), HR:1.0]. For Luminal B, stage of diagnosis, [Distant: (Pr&gt;ChiSq &lt;0.0001), HR:30.17] and Regional: (Pr&gt;ChiSq &lt;0.0001), HR:4.49], and race (Pr&gt;ChiSq &lt;0.0065, HR:1.62]) were significant, but not the CDI [(Pr&gt;ChiSq: &lt; 0.288), HR:0.91]. For HER2, stage of diagnosis, [Distant: (Pr&gt;ChiSq &lt;0.0001), HR:14.798] and Regional: (Pr&gt;ChiSq &lt;0.0001), HR:2.765], was significant, but race (Pr&gt;ChiSq &lt;0.254, HR:1.26]) and CDI [(Pr&gt;ChiSq: &lt; 0.75), HR:1.0] were not. We found significant racial difference in survival in Luminal A and Luminal B breast cancer subtypes but not in HER2-enriched subtype. However, we didn’t find any significant contribution of CDI in receptor positive breast cancer patient survival.

Future studies should focus on assessing the processes by which social determinants affect receptor positive breast cancer disease progression.

Citation Format: Alejandra M. Rivera-Irizarry, Denise Danos, Xinnan Wang, Qiufan Fu, Xiao-Cheng Wu, Richard Scribner, Lucio Miele, Fokhrul Hossain. Social determinants of receptor positive breast cancer patients’ survival in Louisiana [abstract]. In: Proceedings of the AACR Virtual Conference: Thirteenth AACR Conference on the Science of Cancer Health Disparities in Racial/Ethnic Minorities and the Medically Underserved; 2020 Oct 2-4. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2020;29(12 Suppl):Abstract nr PO-155.

Positive impact of the Patient Protection and Affordable Care Act Medicaid expansion on Louisiana women with breast cancer

BACKGROUND

Louisiana is one of the few Southern states that enacted the Medicaid expansion of the Patient Protection and Affordable Care Act (ACA). To the authors' knowledge, the issue of how this has affected the breast cancer landscape in Louisiana is unknown. The authors have postulated that ACA expansion had a positive impact for Louisiana women diagnosed with breast cancer.

METHODS

Data from the Louisiana Tumor Registry regarding 14,640 women aged 20 to 64 years who resided in Louisiana and were diagnosed with American Joint Committee on Cancer stage 0 to stage IV breast cancer between 2012 and 2018 were analyzed. The study period was divided into 2 groups: 1) before ACA expansion (January 1, 2012-May 31, 2016); and 2) after ACA expansion (June 1, 2016-December 31, 2018). The chi-square test and multivariable logistic regression models were used to assess the impact of ACA expansion. A P value <.05 was considered statistically significant.

RESULTS

After ACA expansion, the rate of uninsured patients decreased from 5.4% to 3.0% (P < .0001), and the rate of Medicaid recipients increased from 11.6% to 17.7% (P < .0001). The diagnosis of stage I breast cancer increased from 36.8% to 44.7% (P < .0001), whereas the diagnosis of stage III breast cancer decreased from 10.7% to 8.5% (P < .0001). The receipt of radiotherapy after breast-conserving surgery increased from 81.2% to 84.0% (P = .0035), and the receipt of radiotherapy within 90 days increased from 57.2% to 61.7% (P = .0012). After adjustment for sociodemographic and clinical variables, the models demonstrated that ACA expansion decreased the uninsured rate by 48% (odds ratio [OR], 0.52; 95% CI, 0.43-0.63), increased the diagnosis of early-stage disease (stage0 to stage II) by 27% (OR, 1.27; 95% CI, 1.15-1.41), increased receipt of radiotherapy after breast-conserving surgery by 19% (OR, 1.19; 95% CI, 1.03-1.37), and reduced the delay of receipt of radiotherapy by 16% (OR, 0.84; 95% CI, 0.74-0.95).

CONCLUSIONS

ACA expansion in Louisiana reduced the uninsured rate, increased the diagnosis of early-stage disease, and increased access to treatment.

Geographic and intra-racial disparities in early-onset colorectal cancer in the SEER 18 registries of the United States

Background

Although early‐onset colorectal cancer (EOCRC) incidence rates (IRs) are increasing, geographic and intra‐racial IR disparities are not well defined.

Methods

2000‐2015 Surveillance, Epidemiology, and End Results (SEER) program CRC IR Analysis (170,434 cases) was performed from ages 30 to 60 in four US regions, 18 individual registries, metropolitan and nonmetropolitan locations and stratified by race. Analyses were conducted in 1‐year and 5‐year age increments.

Results

Wide US regional EOCRC IR variations exist: For example, age 45 IRs in the south are 26.8/100,000, 36.0% higher than the West, 19.7/100,000 (p < 0.0001). Disparities magnify between individual registries: EOCRC IRs in highest risk registries were 177‐348% (Alaska Natives), 75‐200% (Hawaii), 76‐128% (Louisiana), and 61‐125% (Kentucky) higher than lowest risk registries depending on age. EOCRC IRs are 18.2%‐25.6% higher in nonmetropolitan versus metropolitan settings. Wide geographic intra‐racial disparities exist. Within the White population, the greatest IR difference (78.8%) was between Kentucky (5.9/100,000) and Los Angeles (3.3/100,000) in 30‐ to 34‐year‐olds (p < .0001). Within the Black population, the greatest difference (136.2%) was between rural Georgia (30.7/100,000) and California excluding San Francisco‐Oakland/San Jose‐Monterey/Los Angeles (13/100,000) in 40‐ to 44‐year‐olds (p = 0003).

Conclusion

Marked geographic EOCRC disparities exist with disproportionately high IRs in Alaska Natives, Hawaii, and southern registries. Geographic intra‐racial disparities are present within White and Black populations. In Blacks, there are disproportionately high EOCRC IRs in rural Georgia. Although vigilance is required in all populations, attention must be paid to these higher risk populations. Potential interventions include assuring early investigation of symptoms, targeting modifiable risk factors and utilizing earlier age 45 screening options supported by some guidelines.

Accelerated training of bootstrap aggregation-based deep information extraction systems from cancer pathology reports

OBJECTIVE

In machine learning, it is evident that the classification of the task performance increases if bootstrap aggregation (bagging) is applied. However, the bagging of deep neural networks takes tremendous amounts of computational resources and training time. The research question that we aimed to answer in this research is whether we could achieve higher task performance scores and accelerate the training by dividing a problem into sub-problems.

MATERIALS AND METHODS

The data used in this study consist of free text from electronic cancer pathology reports. We applied bagging and partitioned data training using Multi-Task Convolutional Neural Network (MT-CNN) and Multi-Task Hierarchical Convolutional Attention Network (MT-HCAN) classifiers. We split a big problem into 20 sub-problems, resampled the training cases 2,000 times, and trained the deep learning model for each bootstrap sample and each sub-problem-thus, generating up to 40,000 models. We performed the training of many models concurrently in a high-performance computing environment at Oak Ridge National Laboratory (ORNL).

RESULTS

We demonstrated that aggregation of the models improves task performance compared with the single-model approach, which is consistent with other research studies; and we demonstrated that the two proposed partitioned bagging methods achieved higher classification accuracy scores on four tasks. Notably, the improvements were significant for the extraction of cancer histology data, which had more than 500 class labels in the task; these results show that data partition may alleviate the complexity of the task. On the contrary, the methods did not achieve superior scores for the tasks of site and subsite classification. Intrinsically, since data partitioning was based on the primary cancer site, the accuracy depended on the determination of the partitions, which needs further investigation and improvement.

CONCLUSION

Results in this research demonstrate that 1. The data partitioning and bagging strategy achieved higher performance scores. 2. We achieved faster training leveraged by the high-performance Summit supercomputer at ORNL.