Can machine learning be used to reduce overtreatment of the axilla in breast cancer? Results from a retrospective cohort study

Jozsa, F, Ahmed, M, Baker, RD ORCID: and Douek, M 2020, Can machine learning be used to reduce overtreatment of the axilla in breast cancer? Results from a retrospective cohort study , in: The European Society of Surgical Oncology (ESSO) 2020 Virtual, 23rd-24th October 2020, Online.

Full text not available from this repository.
Access Information: This is an abstract of a conference session presentation.


Background: Patients with early breast cancer undergoing primary surgery who have low axillary nodal burden can safely forego axillary node clearance (ANC). However, routine use of axillary ultrasound (AUS) leads to 43% of patients in this group having ANC unnecessarily following a positive AUS. The intersection of machine learning with medicine can provide innovative ways to understand specific risk within large patient data sets, but this has not yet been trialled in the arena of axillary node management in breast cancer. Materials and Methods: A single-centre retrospective analysis was performed on patients with breast cancer who had a preoperative axillary ultrasound, and the specificity and sensitivity of AUS were calculated. Machine learning and standard statistical methods were then applied to the data to see if, when used preoperatively, they could have improved the accuracy of AUS to better discern between high and low axillary burden. The python programming language was used, including the pandas, matplotlib, scikitlearn, and tensorflow modules. Data was divided into a 70% and 30% train-test-split and fed into a dense, feed-forward artificial neural network with 3 layers of 11, 6, and 1 neurons, respectively. Backpropagation was optimized using the Adam method. Results: The study included 459 patients; 31% (n=142) had a positive AUS, and, among this group, 62% (n=88) had two or fewer macrometastatic nodes at ANC. When applied to the dataset, logistic regression outperformed AUS and machine learning methods with a specificity of 0.950, correctly identifying 66 patients in this group who had been incorrectly classed as having high axillary burden by AUS alone. Of all the methods, the artificial neural network had the highest accuracy (0.919). Interestingly, AUS had the highest sensitivity of all methods (0.777), underlining its utility in this setting. Conclusions: In our dataset, standard logistic regression outperformed machine learning when identifying false positive axillary ultrasound results in preoperative breast cancer patients. Future applications with larger and higher-dimensional datasets may lend themselves better to machine learning. The finding that pre-operative ultrasound has better sensitivity than the machine learning and statistical methods underlines its use as a tool to potentially allow omission of any surgical staging in patients with early breast cancer and a radiologically and clinically negative axilla.

Item Type: Conference or Workshop Item (Speech)
Additional Information: ** From Elsevier via Jisc Publications Router
Schools: Schools > Salford Business School
Journal or Publication Title: European Journal of Surgical Oncology
Publisher: Elsevier
ISSN: 0748-7983
Related URLs:
SWORD Depositor: Publications Router
Depositing User: Publications Router
Date Deposited: 01 Feb 2021 09:28
Last Modified: 27 Aug 2021 21:49

Actions (login required)

Edit record (repository staff only) Edit record (repository staff only)