Broderick, Brendan T.

ABSTRACT Introduction Community health workers (CHWs) help patients navigate community resources. CHW programs can improve health outcomes and reduce healthcare utilization, but identifying eligible patients is challenging. We developed an electronic health record (EHR)‐based algorithm trained to predict referrals to a CHW program. Our objective was to evaluate whether an algorithm trained on historical referrals identifies a subgroup of patients that differs from the subgroup identified through questionnaire‐based health‐related social needs (HRSN) screening alone. Methods This analysis used data from Mayo Clinic primary care patients in Southeast Minnesota. We developed a gradient boosted time‐to‐event model that incorporated demographics, diagnoses, HRSN questionnaire responses, and HRSNs derived from clinical notes using natural language processing. We compared the characteristics of algorithm‐identified patients with those identified through questionnaires and assessed the variable importance of model features. Results The algorithm had AUC: 0.87 in the training and 0.92 in the hold out set. Patients flagged by the algorithm tended to use interpreter services more commonly (12.6% vs. 1.3%), have a non‐English preferred language (14.1% vs. 2.3%), and have more instances of health literacy documented (5.36 vs. 1.29) compared to patients identified by the HRSN questionnaire. Only 47.0% of patients flagged by the algorithm had responded to the HRSN questionnaire within 2 years. Variable importance was highest for health literacy and financial strain from unstructured notes. Conclusions The algorithm is a chart‐review prioritization tool for referral to the CHW program for patients that may overcome challenges to identifying patients with HRSNs, especially when questionnaire data is incomplete.

INTRODUCTION: Colorectal cancer remains a leading cause of cancer associated death in the United States and colonoscopy the primary screening strategy for prevention. Rates of adenomatous and serrated neoplasia detection are inversely associated with postcolonoscopy colorectal cancer. This crucial quality metric depends on accurate ascertainment of colorectal neoplasia findings from both endoscopy and histopathology records. We aimed to assess the feasibility of a random forest machine learning model to rapidly and accurately categorize colorectal neoplasia from electronic health record data. METHODS: A retrospective cohort study compared neoplasia detection rates among individuals undergoing colonoscopy at a large academic institution to develop a rule-based algorithm to categorize colorectal neoplasia from endoscopy reports and pathology systematized nomenclature of medicine – clinical terms. This cohort provided a large training set to develop a natural language processing system using a random forest approach to automatically classify unstructured pathology findings into adenoma, serrated, or advanced neoplasms. This system was manually validated through an independent holdout set. RESULTS: The training set comprised 35,953 unstructured pathology reports with matched systematized nomenclature of medicine – clinical terms from 95,188 unstructured colonoscopy reports. The final model was assessed on an independent holdout set of 337 manually annotated procedures obtaining an area under the receiver operating characteristic curve of 0.997 (confidence interval [CI] 0.994–1), 0.99 (CI 0.98–1), and 0.99 (CI 0.98–0.99) for prediction of adenoma, serrated, and advanced lesions, respectively. DISCUSSION: The random forest-based hybrid natural language processing system for classification of colonoscopy results was both accurate and explainable. NLP combined with effective machine learning algorithms can provide a scalable strategy for colonoscopy quality monitoring.

5526 Background: HR-HPV DNA testing, with or without cervical cytology, provides excellent sensitivity for detection of cervical cancer (CC) and its precursors; negative test results indicate that risk of disease is extremely low and enable women to undergo reduced screening with safety. However, management of women who screen positive remains challenging as many will prove to have self-limited HR-HPV infections. DNA methylation is an early event in carcinogenesis that could enhance CC screening specificity. Methods: For discovery, DNA from 70 FFPE CC (36 squamous, 34 adenocarcinoma) tissues that were reviewed microscopically, 18 fresh frozen benign cervicovaginal (BCV) tissues collected at the time of benign hysterectomy, and 18 buffy coats from cancer-free women underwent reduced representation bisulfite sequencing (RRBS) to identify MDMs associated with CC. Candidate MDM selection was based on area under the receiver operating characteristic curve (AUC) discrimination, methylation fold change, and low background methylation among benign controls. Candidate MDMs were re-tested using methylation-specific PCR (MSP) to confirm performance. Blinded biological validation was performed using MSP on DNA extracted from independent FFPE CC (38 squamous, 43 adenocarcinoma) and BCV (40) tissues. The performance of CC MDMs was also tested in DNA extracted from cervical dysplasia (36 adenocarcinoma in situ (AIS), 32 cervical intraepithelial neoplasia (CIN) 2/3, 11 CIN 1) FFPE tissues. Results: From RRBS discovery and technical validation via MSP, 30 candidate MDMs showed marked methylation fold changes (10 to >1000) across both CC histologies compared to BCV tissue from cancer-free women. Each of the 30 MDMs highly discriminated CC from BCV tissue with 9 MDMs having an AUC >0.90 (Table). CC MDMs also highly discriminated AIS from BCV but did not perform well in CIN 2/3 and CIN 1 (Table). Conclusions: Whole methylome sequencing, stringent filtering criteria, and biological validation have yielded outstanding candidate MDMs for CC that highly discriminate CC from BCV, notably with high specificity. Performance in cervical dysplasias varied with higher positivity rates in AIS than in CIN 2/3 and CIN 1. Translation to testing these novel MDMs in lower genital tract biospecimens and the addition of HR-HPV to the CC panel are warranted.[Table: see text]

Abstract Background: Biopsy diagnosis of benign breast disease (BBD) based on the most severe lesion in a sample predicts future breast cancer risk and has implications for screening and management. Lobules are the functional unit of the breast and the structures from which BBD arises. We developed and preliminarily validated an automated computational pathology algorithm to discriminate normal from BBD breast lobules as a step toward automated comprehensive characterization of benign biopsies. Methods: 152 BBD biopsies (27 training, 125 validation) from the Mayo Clinic were examined using scanned digital images of H&E stained sections. For each image, a pathologist annotated up to 10 representative lobules as normal or BBD using standard pathology criteria. A deep learning algorithm to quantify lobular features was developed using 129 lobules from 27 subjects. Nine features were identified that discriminate normal vs BBD lobules, expressing lobule size, acini size and number, acinar lumen size, proportion of lobular stroma, and capillaries. Here, we validate their performance to discriminate normal vs BBD lobules in a set of 1250 lobules from 125 subjects using area under the ROC curve (AUC) analysis. Random forest analysis was used for multivariable modeling; model performance was assessed with a tenfold cross-validation approach. Results: Median subject age was 52 years. Among the 125 validation subjects, BBD findings were nonproliferative in 39%, proliferative in 45%, and atypical hyperplasia in 16%. Sections included 552 (44%) normal lobules and 698 BBD (56%) lobules, with representation of both lobule types on each section. In univariate analyses, four individual features showed good discrimination between normal vs. BBD lobules, yielding the following AUCs: lobule size= 0.74, mean acini size= 0.75, epithelial area= 0.75, and number of acini with large lumens= 0.76. Lobule size and epithelial area were highly correlated (Spearman rank correlation r = 0.95), but both of these features showed lower correlation with mean acini size and number of acini with large lumens (each r < 0.5). With random forest modeling, number of acini with large lumens was the strongest discriminating feature, followed by mean acini size, lobule size, and epithelial area. With ten-fold cross validation of the multivariable random forest model, the overall AUC was 0.82 (95% CI: 0.79-0.85). Conclusion: Our validation showed that automated quantitative computational pathology assessment of breast lobules can discriminate normal versus BBD on a per lobule basis. This finding supports the feasibility of developing automated algorithms to classify every lesion in breast biopsies, expanding beyond visual assessment. Further studies using deep learning may reveal novel pathology features for classifying BBD biopsies with the potential to strengthen estimation of breast cancer risk. Citation Format: Amy C. Degnim, Thomas de Bel, Mark E. Sherman, Derek C. Radisky, Stacey J. Winham, Tanya L. Hoskin, Melody L. Stallings Mann, Marlene Frost, Robert A. Vierkant, Brendan T. Broderick, Ethan P. Heinzen, Rushin Brahmbhatt, Muhammad Arshad, Celine M. Vachon, Jodi M. Carter, Lori A. Denison, Daniel W. Visscher, Jeroen van der Laak. Discrimination of benign breast disease from normal lobules using an automated computational pathology algorithm [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2113.

INTRODUCTION: Nonendoscopic Barrett's esophagus (BE) screening may help improve esophageal adenocarcinoma outcomes. We previously demonstrated promising accuracy of methylated DNA markers (MDMs) for the nonendoscopic diagnosis of BE using samples obtained from a capsule sponge-on-string (SOS) device. We aimed to assess the accuracy of these MDMs in an independent cohort using a commercial grade assay. METHODS: BE cases had ≥ 1 cm of circumferential BE with intestinal metaplasia; controls had no endoscopic evidence of BE. The SOS device was withdrawn 8 minutes after swallowing, followed by endoscopy (the criterion standard). Highest performing MDMs from a previous study were blindly assessed on extracted bisulfite-converted DNA by target enrichment long-probe quantitative amplified signal (TELQAS) assays. Optimal MDM combinations were selected and analyzed using random forest modeling with in silico cross-validation. RESULTS: Of 295 patients consented, 268 (91%) swallowed the SOS device; 112 cases and 89 controls met the pre-established inclusion criteria. The median BE length was 6 cm (interquartile range 4–9), and 50% had no dysplasia. The cross-validated sensitivity and specificity of a 5 MDM random forest model were 92% (95% confidence interval 85%–96%) and 94% (95% confidence interval 87%–98%), respectively. Model performance was not affected by age, gender, or smoking history but was influenced by the BE segment length. SOS administration was well tolerated (median [interquartile range] tolerability 2 [0, 4] on 10 scale grading), and 95% preferred SOS over endoscopy. DISCUSSION: Using a minimally invasive molecular approach, MDMs assayed from SOS samples show promise as a safe and accurate nonendoscopic test for BE prediction.

Abstract Background: Pancreatic cancer (PC) ranks among the top causes of cancer mortality. Risk stratification is an important step toward prevention and early detection. Prior work has shown that ABO blood type (non-O) and first-degree family history of PC are associated with increased PC risk. Individuals with first-degree family history of PC have a 2- to 3-fold increased risk, but it is unclear if ABO blood type plays a role in the familial aggregation of PC. We investigated PC risk among first-degree relatives (FDRs) of PC probands based on the probands' serologic blood type (O vs non-O). Methods: A total of 702 probands sequentially enrolled in the Mayo Clinic Pancreas Research Registry from 2000-2016 with serologic ABO blood type provided information on family history of cancer for 5,178 FDRs. FDRs were grouped based on their related probands' blood type as type O (n=242; 1,768 FDRs) vs non-O (n=460; 3,410 FDRs). Standardized incidence ratios (SIRs) and 95% confidence intervals (CIs) were calculated by comparing the number of PC cases observed among FDRs to those expected using data from the Surveillance, Epidemiology and End Results (SEER) Program as the reference population. SIRs for PC risk were calculated separately for FDRs of probands with type O vs FDRs of probands with non-O blood, and stratified by sex of FDRs and relationship to proband (parent, sibling, or offspring). Results: Compared to the SEER reference, PC risk was nearly 2-fold higher among FDRs of PC probands (SIR=1.76; 95%CI: 1.35-2.26). When stratified by probands' ABO type, risk for PC was higher among FDRs of probands with non-O blood (SIR=1.94; 95%CI: 1.41-2.60), compared to FDRs of probands with O blood (SIR=1.42; 95%CI: 0.83-2.28). Parents of probands with non-O blood had increased PC risk (SIR=3.28; 95%CI: 2.18-4.74, n=920), but not parents of probands with O blood (SIR=1.75; 95%CI: 0.75-3.45, n=484), p-value for between-group difference=0.10. Among FDRs of probands with O blood, female FDRs had increased PC risk (SIR=2.12; 95%CI: 1.06-3.79), but male FDRs did not (SIR=0.86; 95%CI: 0.31-1.88). Higher PC risk was observed among female FDRs of probands with non-O blood (SIR=2.28; 95%CI: 1.44-3.41) than male FDRs of probands with non-O blood (SIR=1.62; 95%CI: 1.00-2.48). Mothers of probands with O blood had increased PC risk (SIR=2.89; 95%CI: 1.05-6.28) compared to fathers (SIR=0.79; 95%CI: 0.09-2.84). Both mothers (SIR=3.60; 95%CI: 1.96-6.03) and fathers (SIR=2.95; 95%CI: 1.61-4.95) of probands with non-O blood had increased PC risk. No major findings were observed among probands' siblings or offspring. Conclusions: PC risk is greater among FDRs of PC probands with non-O blood than among FDRs of probands with type O. PC risk is higher among parents of probands with non-O blood, and female FDRs and mothers of probands with non-O blood. Possible clustering of non-O blood type in families may contribute to increased PC risk. Citation Format: Aarti Kolluri, Samuel O. Antwi, Sarah E. Fagan, Kari G. Chaffee, Brendan T. Broderick, William R. Bamlet, Ann L. Oberg, Robert R. McWilliams, Gloria M. Petersen. Risk of pancreatic cancer is increased among first-degree relatives of pancreatic cancer probands who have non-O ABO blood type [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 1239.

Purpose Women with atypical hyperplasia (AH) on breast biopsy have an aggregate increased risk of breast cancer (BC), but existing risk prediction models do not provide accurate individualized estimates of risk in this subset of high-risk women. Here, we used the Mayo benign breast disease cohort to develop and validate a model of BC risk prediction that is specifically for women with AH, which we have designated as AH-BC. Patients and Methods Retrospective cohorts of women age 18 to 85 years with pathologically confirmed benign AH from Rochester, MN, and Nashville, TN, were used for model development and external validation, respectively. Clinical risk factors and histologic features of the tissue biopsy were selected using L1-penalized Cox proportional hazards regression. Identified features were included in a Fine and Gray regression model to estimate BC risk, with death as a competing risk. Model discrimination and calibration were assessed in the model-building set and an external validation set. Results The model-building set consisted of 699 women with AH, 142 of whom developed BC (median follow-up, 8.1 years), and the external validation set consisted of 461 women with 114 later BC events (median follow-up, 11.4 years). The final AH-BC model included three covariates: age at biopsy, age at biopsy squared, and number of foci of AH. At 10 years, the AH-BC model demonstrated good discrimination (0.63 [95% CI, 0.57 to 0.70]) and calibration (0.87 [95% CI, 0.66 to 1.24]). In the external validation set, the model showed acceptable discrimination (0.59 [95% CI, 0.51 to 0.67]) and calibration (0.91 [95% CI, 0.65 to 1.42]). Conclusion We have created a new model with which to refine BC risk prediction for women with AH. The AH-BC model demonstrates good discrimination and calibration, and it validates in an external data set.

Abstract Background: Epidemiologic associations of leukocyte telomere length (LTL) and pancreatic ductal adenocarcinoma (PDAC) have been inconsistent owing, in part, to variation in telomere length (TL) assessment across studies. To overcome this limitation and address concerns of potential reverse causation, we used carriage of telomere-related alleles to genetically predict TL and examined its association with PDAC. Methods: A case–control study of 1,500 PDAC cases and 1,500 controls, frequency-matched on age and sex was performed. Eight of nine polymorphisms previously associated with variation in LTL were analyzed. Genetic risk scores (GRS) consisting of the TL-related polymorphisms were computed as the number of long TL alleles carried by an individual scaled to published kilobase pairs of TL associated with each allele. Participants were further categorized on the basis of the number of short TL alleles they carry across all eight SNPs. Associations were examined in additive and dominant models using logistic regression to calculate ORs and 95% confidence intervals (CI). Results: In age- and sex-adjusted models, one short TL allele (rs10936599, T) was associated with reduced risk, whereas another short TL allele (rs2736100, A) was associated with increased risk, with per-allele ORs of 0.89 (95% CI, 0.79–0.99) and 1.13 (95% CI, 1.01–1.24), respectively. No association was observed with GRS or short TL allele counts, and no associations were observed in the dominant models. Conclusions: Findings suggest that genetically predicted short TL is not associated with PDAC risk. Impact: Common genetic determinants of short TL do not appear to influence PDAC risk. Cancer Epidemiol Biomarkers Prev; 26(6); 971–4. ©2017 AACR.