A multicenter study validated an integrated deep learning model for precision malignancy risk assessment and reducing unnecessary biopsies in BI-RADS 4 cases

Stephen T. Wong, Chika F. Ezeana, Tiancheng He, Tejal Patel, Virginia G Kaklamani, Maryam Elmi, Erica Ibarra, Pamela M. Otto, Kenneth A. Kist, Heather Speck, Lin Wang, Joe Edward Ensor, Jr., Ya-Chen T. Shih, Kim Bumyang, I-Wen Pan, David Spak, Wei T. Yang, Jenny C. Chang

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Introduction: BI-RADS category 4 is associated with a wide variability in probability of malignancy, ranging from 2 to 95% while biopsy-derived positive predictive value (PPV3) for this category’s lesions remains low at 21.1% in the US. A major fallout of these facts is that we have way very high false positive rate leading to too many unnecessary biopsies and their associated costs and emotional burden. We improved our in-house intelligent-augmented Breast cancer RISK calculator (iBRISK), an integrated deep learning (DL) based decision support app and assessed its performance in a multicenter IRB-approved study.

Methods: We improved iBRISK by retraining the DL model with an expanded dataset of 9,700 patient records of clinical risk-factors and mammographic descriptors from Houston Methodist Hospital (HMH) and validated using another 1,078 patient records. These patients were all seen between March 2006 and December 2016. We assessed the model using blinded, independent retrospective BI-RADS 4 patients who had biopsies subsequently after mammography and seen January 2015 - June 2019 at three major healthcare institutions in Texas, USA: MD Anderson Cancer Center, the University of Texas Health Sciences Center at San Antonio, and HMH. We dichotomized and trichotomized the data to evaluate precision of risk stratification and probability of malignancy (POM) estimation translated into biopsy decision augmentation. The iBRISK score as a continuous predictor of malignancy and possible cost savings was also analyzed.

Results: The multicenter validation dataset had 4,209 women, median age (interquartile) was 56 (45, 65) years. The use of iBRISK score as a continuous predictor of malignancy yielded an AUC of 0.97. Among “low” and “moderate” POM patients, only two out of 1,228 patients (0.16%) and 118 out of 1788 (6.6%) were malignant respectively. This translates to an even better precision when compared to newly introduced BI-RADS 4 subcategories 4A and 4B, with associated PPV3s of 7.6% and 22%, respectively. The “high” POM group had a malignancy rate of 85.9% (1,025/1,193). Estimated potential cost savings in the US was over $260 million.

Conclusion: The iBRISK app demonstrated high sensitivity in malignancy prediction and can potentially be used to safely obviate biopsies in up to 50% of patients in low/moderate POM-groups. This would result in significant healthcare quality improvement, cost savings, and help reduce patient anxiety.
Original languageEnglish (US)
Title of host publicationProceedings of the American Association for Cancer Research Annual Meeting 2023
Subtitle of host publicationCancer Res 2023;83(7_Suppl):Abstract nr 5698
DOIs
StatePublished - Apr 4 2023
EventAACR Annual Meeting 2023 - Orlando, United States
Duration: Apr 14 2023Apr 19 2023

Conference

ConferenceAACR Annual Meeting 2023
Country/TerritoryUnited States
CityOrlando
Period4/14/234/19/23

Fingerprint

Dive into the research topics of 'A multicenter study validated an integrated deep learning model for precision malignancy risk assessment and reducing unnecessary biopsies in BI-RADS 4 cases'. Together they form a unique fingerprint.

Cite this